%======================== \section{Construction of...
बनाया गया: 24 फ़रवरी 2026
का उपयोग करके उत्तर दिया गया GPT-5.2 Thinking द्वारा Chat01
बनाया गया: 24 फ़रवरी 2026
का उपयोग करके उत्तर दिया गया GPT-5.2 Thinking द्वारा Chat01
%========================
\section{Construction of Dataset}
%========================
Consider a five-stand tandem cold rolling mill indexed by .
Let denote the -th sampling instant and the corresponding sampling interval length.
Define the local within-interval time variable .
Let denote the exit thickness of stand (),
and let denote the inter-stand strip tension between stand and ().
The system is strongly coupled because the inter-stand tensions propagate along the mill line and are affected by neighboring stands' actions.
Let and be the desired references (setpoints) given by process requirements. We define deviation variables
\begin{equation}
\Delta h_i(t)\triangleq h_i(t)-h_i^{\mathrm{ref}}(t),\qquad
\Delta T_i(t)\triangleq T_i(t)-T_i^{\mathrm{ref}}(t).
\label{eq:dev_def}
\end{equation}
Throughout this paper, the symbol ``'' attached to states always means deviation from reference.
For each stand , we choose the local deviation state as
\begin{equation}
\Delta x_i(t)\triangleq
\begin{bmatrix}
\Delta h_i(t)\
\Delta T_{i-1}(t)\
\Delta T_i(t)
\end{bmatrix}\in\mathbb{R}^{d},\qquad d=3,
\label{eq:xi_def_clean}
\end{equation}
with the boundary convention and to keep a unified dimension for all stands.
For a five-stand tandem mill, the dominant coupling is between adjacent stands, hence we define
\begin{equation}
Z_1={2},\quad
Z_i={i-1,i+1}\ (i=2,3,4),\quad
Z_5={4}.
\label{eq:Zi_clean}
\end{equation}
Define the neighbor-state stack
\begin{equation}
\Delta x_{Z_i}(t_n)=\mathrm{col}{\Delta x_k(t_n),|,k\in Z_i}.
\label{eq:xZi_clean}
\end{equation}
\begin{bmatrix}
\Delta s_i(t_n)\
\Delta v_i(t_n)
\end{bmatrix}.
\label{eq:du_discrete_clean}
\end{equation}
Throughout this paper, the symbol ``'' attached to inputs means sample-to-sample increment.
Thus, and are conceptually different, and this is fixed by definition.
Let denote exogenous disturbances.
We denote the interval-level equivalent disturbance by .
denotes the identity matrix; denotes the zero matrix.
M_d,\Delta x_i(t_n)
+
N_d,\Delta u_i(t_n)
+
F_d,\Delta d_i(t_n),
\label{eq:linear_form_concept}
\end{equation}
where represent equivalent discrete-time matrices around operating conditions.
In a practical five-stand cold rolling mill, accurately deriving and identifying these matrices and disturbance models from first principles is difficult,
due to strong coupling, unmodeled nonlinearities, and time-varying operating regimes.
Therefore, this paper aims to learn a high-fidelity approximation of the interval evolution from data and then embed it into distributed MPC.
\begin{remark}
In fact, due to the existence of complex coupling relationships, it is difficult to directly and accurately establish \eqref{eq:linear_form_concept}
based on first principles. Therefore, in this paper, we learn an approximate mapping of \eqref{eq:true_mapping_clean} from data.
\end{remark}
Although decisions are updated at discrete instants , the hydraulic gap and drive systems evolve continuously inside each interval,
and abrupt within-interval changes may excite tension oscillations and deteriorate thickness stability.
Thus, parameterizing the within-interval increment trajectory by a low-order polynomial:
(i) yields a compact finite-dimensional decision representation;
(ii) enforces smooth profiles inside the interval;
(iii) enables enforcing increment constraints for all .
This is appropriate when is not excessively large relative to actuator bandwidth and the within-interval evolution is well approximated by a low-order basis.
\Gamma_{i,n0}+\Gamma_{i,n1}\tau+\Gamma_{i,n2}\tau^2,
\qquad \tau\in[0,\delta_n],
\label{eq:du_poly_vec_clean}
\end{equation}
where are coefficient vectors ().
Component-wise, \eqref{eq:du_poly_vec_clean} corresponds to
\begin{equation}
\begin{aligned}
\Delta s_{i,n}(\tau) &= \gamma^{(s)}{i,n0}+\gamma^{(s)}{i,n1}\tau+\gamma^{(s)}{i,n2}\tau^2,\
\Delta v{i,n}(\tau) &= \gamma^{(v)}{i,n0}+\gamma^{(v)}{i,n1}\tau+\gamma^{(v)}{i,n2}\tau^2.
\end{aligned}
\label{eq:du_components_clean}
\end{equation}
Define the stacked parameter vector
\begin{equation}
\Gamma{i,n}\triangleq
\big[
(\Gamma_{i,n0})^\top,,
(\Gamma_{i,n1})^\top,,
(\Gamma_{i,n2})^\top
\big]^\top
\in\mathbb{R}^{p},
\qquad
p=3n_u=6.
\label{eq:Gamma_clean}
\end{equation}
Here, is the baseline increment at , while and describe the linear and quadratic variation rates.
Define the interval-averaged equivalent increments as
\begin{equation}
\begin{aligned}
\Delta u_i(t_n) &\triangleq \frac{1}{\delta_n}\int_0^{\delta_n}\Delta u_{i,n}(\tau),d\tau,\
\Delta d_i(t_n) &\triangleq \frac{1}{\delta_n}\int_0^{\delta_n}\Delta d_i(\tau),d\tau.
\end{aligned}
\label{eq:avg_def_clean}
\end{equation}
With \eqref{eq:du_poly_vec_clean}, the input average has a closed form:
\begin{equation}
\Delta u_i(t_n)=
\Gamma_{i,n0}
+\Gamma_{i,n1}\frac{\delta_n}{2}
+\Gamma_{i,n2}\frac{\delta_n^2}{3}.
\label{eq:avg_closed_clean}
\end{equation}
Let denote the sampling domain of deviation states and neighbor stacks ,
and let denote the sampling domain of polynomial parameters .
These domains specify the operating envelope used to generate supervised training data.
Given the above parameterization, one training sample is generated on each interval .
In addition to the local deviation state, the neighbor deviation states are included to represent inter-stand coupling.
The process is summarized in Table~\ref{tab:interval_sample_generation_en}.
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.15}
\caption{Procedure for generating one interval-level sample on (five-stand coupled mill).}
\label{tab:interval_sample_generation_en}
\begin{tabularx}{\linewidth}{>{\centering\arraybackslash}p{0.09\linewidth} X}
\toprule
\textbf{Step} & \textbf{Operation} \
\midrule
1 & \textbf{State sampling:} sample and from . \
2 & \textbf{Parameter sampling:} draw (coefficients for both and ). \
3 & \textbf{Control construction:} compute via \eqref{eq:du_poly_vec_clean}. \
4 & \textbf{State propagation:} integrate the \emph{five-stand coupled} mill model on (e.g., RK4) using the within-interval control trajectory, and record . \
\bottomrule
\end{tabularx}
\end{table}
Accordingly, an interval sample for subsystem can be represented as
\begin{equation}
\mathcal{D}{i,n}=\big{\Delta x_i(t_n),\ \Delta x{Z_i}(t_n),\ \Delta u_{i,n}(\tau),\ \Delta x_i(t_{n+1})\big}.
\label{eq:interval_sample_clean}
\end{equation}
Note that is fully determined by via \eqref{eq:du_poly_vec_clean},
therefore it is sufficient to store as the learning input.
For each subsystem , by repeating the above procedure across multiple intervals and randomized draws,
the local one-step training dataset is formed as
\begin{equation}
\begin{split}
S_i=\Big{&
\big(\Delta x_i^{(j)}(t_n),,\Delta x_{Z_i}^{(j)}(t_n),,\Delta x_i^{(j)}(t_{n+1});,
\Gamma_{i,n}^{(j)},,\delta_n^{(j)}\big)
\ \Big|\ j=1,\ldots,J
\Big}.
\end{split}
\label{eq:S_i_clean}
\end{equation}
Here is the number of one-step samples for subsystem .
The overall dataset for the five-stand mill is denoted by .
The point-cloud visualization of the training dataset is shown in Figure~\ref{2}.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=0.5]{picture/Fig2.pdf}
\caption{Point cloud map of the training dataset.}\label{2}
\end{figure*}
The one-step set is sufficient for one-step regression, but it is not sufficient for training with multi-step rollout loss
and reciprocal-consistency regularization, because these objectives require ground-truth deviation-state trajectories over a horizon of consecutive intervals.
Therefore, without changing the single-interval sampling mechanism above, we additionally organize the offline-simulated samples
into -step trajectory segments.
Specifically, for each starting time we generate a segment of length by consecutively sampling
(and the corresponding inputs/disturbances),
and integrating the five-stand coupled mill model over for .
Hence, we obtain the deviation-state sequence as well as the neighbor stacks
.
Define a -step segment sample for subsystem as
\begin{equation}
\begin{aligned}
\mathcal{W}{i,n}=
\Big{&
\big(\Delta x_i(t{n+s}),,\Delta x_{Z_i}(t_{n+s}),,\Gamma_{i,n+s},,\delta_{n+s}\big){s=0}^{K-1}; \
&\big(\Delta x_i(t{n+s+1})\big){s=0}^{K-1}
\Big}.
\end{aligned}
\label{eq:segment_clean}
\end{equation}
By repeating the above segment generation, we form the multi-step training set
\begin{equation}
S_i^{(K)}=\Big{\mathcal{W}{i,n}^{(j)}\ \Big|\ j=1,\ldots,J_K\Big},
\label{eq:S_i_K_clean}
\end{equation}
where is the number of -step segment samples.
Note that can be viewed as the marginal one-step projection of (keeping only ),
thus the original dataset design is preserved, and only an additional \emph{segment organization} is introduced for multi-step training.
%========================
\section{Construction of Residual Neural Network}
%========================
\subsection{Residual Neural Network Structure Construction and Training Method}
Given the dataset, the neural network model is trained to learn a stand-wise, control-dependent one-step evolution law of deviation states:
\begin{equation}
\Delta x_i(t_{n+1})
\approx
\Delta x_i(t_n)+
\mathcal{N}i!\Big(\Delta x_i(t_n),,\Delta x{Z_i}(t_n),,\Gamma_{i,n},,\delta_n;,\Theta_i\Big),
\label{eq:learned_dyn_clean}
\end{equation}
where outputs the one-step deviation-state change and are trainable parameters.
\begin{remark}
If does not take control information as input (here and ),
the predictor becomes an autoregressive model that only reproduces trajectories under the training input patterns
and cannot answer the counterfactual question: ``what will happen if we choose a different roll gap and speed trajectory?''
Since MPC optimizes over candidate decisions, a control-dependent predictor \eqref{eq:learned_dyn_clean} is necessary
to evaluate the predicted thickness and tension behavior under different candidate actuator trajectories.
\end{remark}
Let (state dimension), be the number of neighbors of stand in \eqref{eq:Zi_clean}, and in \eqref{eq:Gamma_clean}.
Define the input vector
\begin{equation}
X_{i,\text{in}} \triangleq
\big[
\Delta x_i(t_n)^\top,,
\Delta x_{Z_i}(t_n)^\top,,
\Gamma_{i,n}^\top,,
\delta_n
\big]^\top
\in \mathbb{R}^{d(1+|Z_i|)+p+1}.
\label{eq:X_in_clean}
\end{equation}
The network mapping is
\begin{equation}
\mathcal{N}_i:\mathbb{R}^{d(1+|Z_i|)+p+1}\rightarrow\mathbb{R}^{d}.
\end{equation}
To improve training stability and long-horizon rollout robustness, we use a residual form.
Let be a selection matrix extracting the local state block:
\begin{equation}
\hat{I}i = [I_d,, 0{d\times(d|Z_i|+p+1)}].
\label{eq:Ihat_clean}
\end{equation}
Then the one-step predictor is written as
\begin{equation}
X_{i,\text{out}} = \hat{I}i X{i,\text{in}} + \mathcal{N}i(X{i,\text{in}}; \Theta_i),
\label{eq:res_predict_clean}
\end{equation}
where represents the predicted .
This structure implements a baseline-plus-correction interpretation:
the shortcut propagates the current deviation state , while the network learns the correction capturing
unmodeled nonlinearities and inter-stand coupling (via ) under varying operating conditions.
To improve robustness when varies, we introduce an auxiliary branch inside :
\begin{equation}
\mathcal{N}i(X{i,\text{in}};\Theta_i)\triangleq
\psi_i(X_{i,\text{in}};\Theta_{\psi_i}) + \rho_i(X_{i,\text{in}};\theta_i),
\label{eq:aux_clean}
\end{equation}
where is a lightweight feedforward branch that captures low-frequency/scale effects strongly related to ,
and captures the remaining nonlinear coupling corrections.
When , the model reduces to a standard residual network.
For the -th sample in \eqref{eq:S_i_clean}, define
\begin{equation}
X_{i,\text{in}}^{(j)} =
\big[
\Delta x_i^{(j)}(t_n),\ \Delta x_{Z_i}^{(j)}(t_n),\
\Gamma_{i,n}^{(j)},\ \delta_n^{(j)}
\big]^{\top},
\end{equation}
and the supervised residual target
\begin{equation}
\Delta r_i^{(j)}=\Delta x_i^{(j)}(t_{n+1})-\Delta x_i^{(j)}(t_n).
\label{eq:target_clean}
\end{equation}
To suppress accumulation drift induced by long-horizon recursion and to improve long-term predictive stability,
we train the forward predictor jointly with an auxiliary backward residual model
and impose a multi-step reciprocal-consistency regularization over a -step segment from .
Construct a backward residual network
\begin{equation}
\mathcal{B}i:\mathbb{R}^{d(1+|Z_i|)+p+1}\rightarrow\mathbb{R}^{d},
\end{equation}
parameterized by . For the backward step associated with interval , define
\begin{equation}
\begin{aligned}
X{i,\mathrm{in}}^{b}
&=
\big[
\Delta x_i(t_{n+1}),\ \Delta x_{Z_i}(t_{n+1}),
\Gamma_{i,n},\ \delta_n
\big]^{\top},\
X_{i,\mathrm{out}}^{b}
&=
\hat{I}i X{i,\mathrm{in}}^{b} + \mathcal{B}i(X{i,\mathrm{in}}^{b};\bar{\Theta}i),
\end{aligned}
\label{eq:back_clean}
\end{equation}
where represents the backward estimate of .
The supervised backward residual target is
\begin{equation}
\Delta r_i^{b}=\Delta x_i(t_n)-\Delta x_i(t{n+1}).
\end{equation}
Given a segment sample , initialize
\begin{equation}
\Delta \hat{x}i(t_n)=\Delta x_i(t_n),
\end{equation}
and recursively apply the forward predictor for steps:
\begin{equation}
\begin{aligned}
\Delta \hat{x}i(t{n+s+1})
&=
\Delta \hat{x}i(t{n+s})
+
\mathcal{N}i!\Big(
\Delta \hat{x}i(t{n+s}),,\Delta \hat{x}{Z_i}(t{n+s}),,
\Gamma_{i,n+s},,\delta_{n+s};,\Theta_i
\Big),\
&\qquad s=0,\ldots,K-1.
\end{aligned}
\label{eq:fwd_roll_clean}
\end{equation}
Set the terminal condition
\begin{equation}
\Delta \bar{x}i(t{n+K})=\Delta \hat{x}i(t{n+K}),
\end{equation}
and roll back using :
\begin{equation}
\begin{aligned}
\Delta \bar{x}i(t{n+s})
&=
\hat{I}i X{i,\mathrm{in}}^{b}(t_{n+s})
+
\mathcal{B}i!\Big(X{i,\mathrm{in}}^{b}(t_{n+s});,\bar{\Theta}i\Big),
\quad s=K-1,\ldots,0,
\end{aligned}
\label{eq:bwd_roll_clean}
\end{equation}
where
\begin{equation}
X{i,\mathrm{in}}^{b}(t_{n+s})=
\big[
\Delta \bar{x}i(t{n+s+1}),\ \Delta \hat{x}{Z_i}(t{n+s+1}),
\Gamma_{i,n+s},\ \delta_{n+s}
\big]^{\top}.
\end{equation}
\sum_{s=0}^{K}
\left|
\Delta \hat{x}i(t{n+s})-\Delta \bar{x}i(t{n+s})
\right|^2.
\end{equation}
We jointly minimize:
\begin{equation}
\begin{aligned}
L_{\mathrm{1step}}(\Theta_i)
&= \frac{1}{J_K}\sum_{j=1}^{J_K}\frac{1}{K}\sum_{s=0}^{K-1}
\Big|
\big(\Delta x_i^{(j)}(t_{n+s+1})-\Delta x_i^{(j)}(t_{n+s})\big)
-\mathcal{N}i!\left(
X{i,\mathrm{in}}^{(j)}(t_{n+s});\Theta_i
\right)
\Big|^2,\[2mm]
L_{\mathrm{bwd}}(\bar{\Theta}i)
&= \frac{1}{J_K}\sum{j=1}^{J_K}\frac{1}{K}\sum_{s=0}^{K-1}
\Big|
\big(\Delta x_i^{(j)}(t_{n+s})-\Delta x_i^{(j)}(t_{n+s+1})\big)
-\mathcal{B}i!\left(
X{i,\mathrm{in}}^{b,(j)}(t_{n+s});\bar{\Theta}i
\right)
\Big|^2,\[2mm]
L{\mathrm{msrp}}(\Theta_i,\bar{\Theta}i)
&= \frac{1}{J_K}\sum{j=1}^{J_K} E_i^{(j)}(t_n),\[2mm]
L_{\mathrm{roll}}(\Theta_i)
&= \frac{1}{J_K}\sum_{j=1}^{J_K}\sum_{s=1}^{K}
\Big|
\Delta x_i^{(j)}(t_{n+s})-\Delta \hat{x}i^{(j)}(t{n+s})
\Big|^2.
\end{aligned}
\label{eq:loss_clean}
\end{equation}
Here, enforces one-step accuracy; explicitly suppresses long-horizon drift under recursion;
regularizes the learned dynamics by enforcing reciprocal consistency between forward and backward rollouts;
and trains the backward model for the consistency regularization.
In implementation, these terms are combined as
\begin{equation}
L_{\mathrm{total}}=\lambda_1 L_{\mathrm{1step}}+\lambda_2 L_{\mathrm{roll}}+\lambda_3 L_{\mathrm{msrp}}+\lambda_4 L_{\mathrm{bwd}},
\end{equation}
where are tuned on a validation set.
\Delta x_i(t_n)
+
\mathcal{N}i!\Big(
\Delta x_i(t_n),,\Delta x{Z_i}(t_n),,
\Gamma_{i,n},,\delta_n;,\Theta_i^*
\Big),
\label{eq:pred_clean}
\end{equation}
and multi-step prediction is obtained by recursive rollout of \eqref{eq:pred_clean}.
This learned predictor is the internal model used by the MPC optimizer in the next section.
Finally, network parameters are optimized using Adam:
\begin{equation}
\Theta_{i,t+1} = \Theta_{i,t} - \alpha \frac{\hat{m}{i,t}}{\sqrt{\hat{v}{i,t}} + \varepsilon},
\end{equation}
where is the learning rate (we use to avoid conflict with other symbols),
and are bias-corrected moment estimates, and is a small constant for numerical stability.
Figure~\ref{fig:rnn_logic} illustrates the overall structure.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.85]{picture/x6.pdf}
\caption{Logic diagram of the residual neural network.}
\label{fig:rnn_logic}
\end{figure}
\subsection{Explainability of the residual network}
\Delta x_i(t_n)
+
\underbrace{\int_{0}^{\delta_n}
f_i!\Big(\Delta x_i(t_n+\tau),,\Delta x_{Z_i}(t_n+\tau),,u_i(t_n+\tau),,d_i(t_n+\tau)\Big),d\tau}{\triangleq\ \varphi{i,n}},
\label{eq:increment_integral}
\end{equation}
where is the one-interval state increment generated by the mechanistic dynamics.
Our learned model \eqref{eq:learned_dyn_clean} adopts the same increment form as \eqref{eq:increment_integral}:
\begin{equation}
\Delta x_i(t_{n+1})
\approx
\Delta x_i(t_n)
+
\mathcal{N}i!\Big(\Delta x_i(t_n),,\Delta x{Z_i}(t_n),,\Gamma_{i,n},,\delta_n;\Theta_i\Big).
\end{equation}
Here, plays the role of a data-driven approximation of the integral increment ,
i.e., it approximates the accumulated effect of the mechanistic dynamics over .
This is consistent with the well-known interpretation that a residual network behaves like a one-step time integrator:
the identity path propagates the current state, while the residual branch represents the increment over the time lag.
Inside each interval, we do not optimize point-wise but parameterize the increment trajectory by
.
Hence, the mechanistic increment in \eqref{eq:increment_integral} depends on the \emph{whole} within-interval trajectory.
Feeding into is therefore a compact way to represent how different candidate
gap/speed trajectories change the integral effect and thus the next thickness--tension state.
When varies and is not very small, directly learning may be harder.
Motivated by the generalized residual idea,
we decompose the increment predictor into two parts in \eqref{eq:aux_clean}:
\begin{equation}
\mathcal{N}_i(\cdot)=\psi_i(\cdot)+\rho_i(\cdot).
\end{equation}
Conceptually, captures low-frequency and scale effects strongly related to ,
while captures the remaining nonlinear coupling corrections.
This provides a mechanism-consistent interpretation: a baseline increment plus a residual correction
that compensates unmodeled nonlinearities and inter-stand coupling.
%========================
\section{Nash Equilibrium-Based RNE-DMPC}
%========================
The five-stand tandem cold rolling system is strongly coupled through inter-stand tension propagation.
As a result, changes in control actions (roll gap and stand speed) at one stand can affect both upstream and downstream stands,
making centralized online optimization over all stands' decision variables computationally demanding.
To mitigate this issue, we decompose the global predictive-control problem into local subproblems associated with individual stands.
Each local controller optimizes its own decision variables while accounting for coupling via limited information exchange with neighboring controllers.
Motivated by game-theoretic coordination \citep{rawlings2008coordinating}, we formulate distributed coordination as a Nash-equilibrium-seeking iteration.
Based on the trained residual neural network surrogate model, we construct a Nash-equilibrium-based distributed MPC method (RNE-DMPC)
for coordinated thickness--tension regulation and tracking. The overall control structure is shown in Figure~\ref{4}.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/x2.pdf}
\caption{Schematic diagram of the control architecture for a tandem cold rolling mill.}\label{4}
\end{figure*}
At sampling time , stand chooses the polynomial-parameter sequence
, where .
Let
denote the joint strategy profile, and let denote the collection of all strategies except stand .
Given the current measured/estimated deviation state and the strategies
,
the multi-step prediction used by stand is written explicitly as
\begin{equation}
\begin{aligned}
\Delta \hat{x}i(t{n+s+1};\mathbf{\Gamma}i,\mathbf{\Gamma}{Z_i})
&=
\Delta \hat{x}i(t{n+s};\mathbf{\Gamma}i,\mathbf{\Gamma}{Z_i})
+
\mathcal{N}i!\Big(
\Delta \hat{x}i(t{n+s};\cdot),,
\Delta \hat{x}{Z_i}(t_{n+s};\mathbf{\Gamma}{Z_i}),\
&\qquad
\Gamma{i,n+s},,
\delta_{n+s};\Theta_i^*
\Big),
\end{aligned}
\label{eq:rollout_mpc_game}
\end{equation}
for , with initialization .
Here the neighbor stack is generated from neighbors' strategies via the same learned predictors.
\Gamma_{i,n+s,0}
+\Gamma_{i,n+s,1}\frac{\delta_{n+s}}{2}
+\Gamma_{i,n+s,2}\frac{\delta_{n+s}^2}{3}.
\label{eq:du_avg_clean}
\end{equation}
\begin{remark}
Because inter-stand tension is jointly affected by the adjacent stands and ,
the predicted evolution of depends on neighbors' future actions,
hence the MPC problems are not independent but form a coupled dynamic game.
\end{remark}
\mathrm{col}{\Gamma_{i,n},\Gamma_{i,n+1},\ldots,\Gamma_{i,n+N_c-1}}
\in \mathbb{R}^{pN_c}.
\end{equation}
In deviation coordinates, the regulation/tracking objective is , i.e.
\begin{equation}
\Delta x_{i,\mathrm{ref}}(t_{n+s})\equiv 0\in\mathbb{R}^{d},\qquad d=3.
\end{equation}
Recall .
Define the row selectors
\begin{equation}
C^- \triangleq [0\ \ 1\ \ 0]\in\mathbb{R}^{1\times 3},\qquad
C^+ \triangleq [0\ \ 0\ \ 1]\in\mathbb{R}^{1\times 3},
\end{equation}
so that (upstream interface) and (downstream interface).
C^-\Delta \hat{x}{i+1}(t{n+s};\mathbf{\Gamma}),\qquad i=1,\ldots,4.
\label{eq:shared_tension_mismatch}
\end{equation}
\sum_{s=1}^{N_p}
\left|
\Delta \hat{x}i(t{n+s};\mathbf{\Gamma}i,\mathbf{\Gamma}{Z_i})
\right|{Q_i}^{2}
+
\sum{s=0}^{N_c-1}
\left|\Gamma_{i,n+s}\right|_{R_i}^{2}
+
J_i^{\mathrm{cpl}}(\mathbf{\Gamma}i;\mathbf{\Gamma}{-i})
\label{eq:Ji_game}
\end{equation}
where weights thickness and tension deviations, and penalizes actuation magnitudes.
C^-\Delta \hat{x}{i+1}(t{n+s};\mathbf{\Gamma})
\big|^2
\Big),
\end{aligned}
\label{eq:coupling_cost}
\end{equation}
with and indicator .
This term makes the coupling conflict explicit: unilateral actions that locally reduce thickness error may worsen shared-tension
compatibility and thus increase , and also affect neighbors' objectives.
We enforce the absolute-input bounds and within-interval increment bounds.
Absolute input bounds (roll gap and speed):
\begin{equation}
u_{i,\min}\le u_i(t_{n+s})\le u_{i,\max},
\qquad s=0,\ldots,N_p-1,
\label{eq:u_abs_game}
\end{equation}
where .
Within-interval increment-trajectory bounds:
\begin{equation}
\Delta u_{i,\min}\le \Delta u_{i,n+s}(\tau;\Gamma_{i,n+s})\le \Delta u_{i,\max},
\qquad \forall\tau\in[0,\delta_{n+s}],\ s=0,\ldots,N_p-1,
\label{eq:du_traj_game}
\end{equation}
where is given by the polynomial parameterization.
\Gamma_{i,n+s,0}
+\Gamma_{i,n+s,1}\frac{\delta_{n+s}}{2}
+\Gamma_{i,n+s,2}\frac{\delta_{n+s}^2}{3},
\label{eq:du_avg_game}
\end{equation}
and propagate the absolute input along the horizon by
\begin{equation}
u_i(t_n)=u_i(t_{n-1})+\Delta u_i(t_n),\qquad
u_i(t_{n+s})=u_i(t_{n+s-1})+\Delta u_i(t_{n+s}),\ s=1,\ldots,N_p-1,
\label{eq:u_prop_game}
\end{equation}
where is the applied (measured) input from the previous sampling instant.
Compact feasible set:
\begin{equation}
\Omega_i \triangleq
\Big{\mathbf{\Gamma}_i\ \Big|\
\eqref{eq:rollout_mpc_game}\ \text{holds and}
\eqref{eq:u_abs_game},\eqref{eq:du_traj_game},\eqref{eq:u_prop_game}\ \text{are satisfied}
\Big}.
\label{eq:Omega_i_game}
\end{equation}
\arg\min_{\mathbf{\Gamma}_i\in\Omega_i}\
J_i(\mathbf{\Gamma}i;\mathbf{\Gamma}{-i}).
\label{eq:local_BR}
\end{equation}
Because the learned surrogate is differentiable, \eqref{eq:local_BR} can be solved by standard gradient-based NLP solvers.
At each sampling time , the distributed MPC coordination induces a finite-horizon dynamic game:
players are stands ; strategy sets are ; and payoff (cost) functions are
defined in \eqref{eq:Ji_game}--\eqref{eq:coupling_cost}.
A joint strategy profile
is a Nash equilibrium if
\begin{equation}
\forall i\in{1,\ldots,5},\qquad
\mathbf{\Gamma}i^*\in
\arg\min{\mathbf{\Gamma}_i\in\Omega_i}
J_i(\mathbf{\Gamma}i;\mathbf{\Gamma}{-i}^*).
\label{eq:NE_def}
\end{equation}
This definition explicitly characterizes the strategic coupling:
each player's optimal decision depends on neighbors' decisions through the shared-tension dynamics and the coupling term.
(1-\omega)\mathbf{\Gamma}_i^{(l-1)}
+
\omega,\mathbf{\Gamma}_i^{\mathrm{BR},(l)},
\qquad \omega\in(0,1].
\label{eq:relaxed_BR}
\end{equation}
The relaxation factor mitigates oscillations caused by strong coupling and improves practical convergence.
The Nash equilibrium is computed through distributed best-response iterations, summarized in Table ~\ref{tab:nash_iter_en}.
\max_i
\frac{\left|
\mathbf{\Gamma}_i^{(l)}-\mathbf{\Gamma}_i^{(l-1)}
\right|_2}{
\left|
\mathbf{\Gamma}_i^{(l-1)}
\right|_2+\epsilon},
\end{equation}
with small.
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.12}
\setlength{\tabcolsep}{3.5pt}
\caption{Relaxed distributed Nash best-response iteration for RNE-DMPC (five-stand).}
\label{tab:nash_iter_en}
\begin{tabularx}{\linewidth}{>{\centering\arraybackslash}p{0.11\linewidth} X}
\toprule
\textbf{Step} & \textbf{Description} \
\midrule
A &
Initialize and warm-start (e.g., from the previous sampling time). \
B &
Communicate (or the induced predicted trajectories) among neighbors; form . \
C &
Given , solve the best-response NLP \eqref{eq:local_BR} to obtain . \
D &
Update the relaxed strategy using \eqref{eq:relaxed_BR} and compute the induced predictions
via \eqref{eq:rollout_mpc_game}. \
E &
Broadcast and the predicted interface quantities needed in \eqref{eq:coupling_cost}
(e.g., ) to neighbors. \
F &
Compute ; if , stop and set ; otherwise and repeat. \
\bottomrule
\end{tabularx}
\end{table}
\Gamma_{i,n,0}^*
+
\Gamma_{i,n,1}^\frac{\delta_n}{2}
+
\Gamma_{i,n,2}^\frac{\delta_n^2}{3},
\end{equation}
followed by .
The overall control flow chart is shown in Fig~\ref{liu}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{picture/x5.pdf}
\caption{The overall system control flow chart}\label{liu}
\end{figure}
\begin{remark}
Equations \eqref{eq:Ji_game}--\eqref{eq:coupling_cost} and the NE definition \eqref{eq:NE_def} explicitly show that each player's objective
depends on others' strategies through the coupled prediction dynamics and the shared-tension interface mismatch penalty,
thus making the multi-stand strategic interaction and coordination mechanism transparent.
\end{remark}这是我最新的第二章,第三章,第四章。在多项式参数化的部分, 这个地方一般的情况用n阶进行方案设计,仿真可以选为二阶的,所有的时变信号参数化都这样改。给出改后的完整的latex代码,上下文要符合逻辑,在符合逻辑的前提下其他地方都不要改
下面给出按你要求修改后的完整 LaTeX 代码:仅将“多项式参数化”相关内容统一改为一般 n 阶形式(并在文中说明“仿真可取二阶 n=2”),同时把所有依赖该参数化的地方(如 、平均增量闭式、MPC 中的参数化与平均式、收敛后施加式等)做了对应一致更新;其余内容保持不变、上下文逻辑自洽。
latex%======================== \section{Construction of Dataset} %======================== Consider a five-stand tandem cold rolling mill indexed by $i\in\{1,2,3,4,5\}$. Let $t_n$ denote the $n$-th sampling instant and $\delta_n=t_{n+1}-t_n$ the corresponding sampling interval length. Define the local within-interval time variable $\tau=t-t_n\in[0,\delta_n]$. Let $h_i(t)$ denote the exit thickness of stand $i$ ($i=1,\dots,5$), and let $T_i(t)$ denote the inter-stand strip tension between stand $i$ and $i+1$ ($i=1,\dots,4$). The system is strongly coupled because the inter-stand tensions propagate along the mill line and are affected by neighboring stands' actions. Let $h_i^{\mathrm{ref}}(t)$ and $T_i^{\mathrm{ref}}(t)$ be the desired references (setpoints) given by process requirements. We define deviation variables \begin{equation} \Delta h_i(t)\triangleq h_i(t)-h_i^{\mathrm{ref}}(t),\qquad \Delta T_i(t)\triangleq T_i(t)-T_i^{\mathrm{ref}}(t). \label{eq:dev_def} \end{equation} Throughout this paper, the symbol ``$\Delta$'' attached to states always means deviation from reference. For each stand $i$, we choose the local deviation state as \begin{equation} \Delta x_i(t)\triangleq \begin{bmatrix} \Delta h_i(t)\\ \Delta T_{i-1}(t)\\ \Delta T_i(t) \end{bmatrix}\in\mathbb{R}^{d},\qquad d=3, \label{eq:xi_def_clean} \end{equation} with the boundary convention $\Delta T_0(t)\equiv 0$ and $\Delta T_5(t)\equiv 0$ to keep a unified dimension $d=3$ for all stands. For a five-stand tandem mill, the dominant coupling is between adjacent stands, hence we define \begin{equation} Z_1=\{2\},\quad Z_i=\{i-1,i+1\}\ (i=2,3,4),\quad Z_5=\{4\}. \label{eq:Zi_clean} \end{equation} Define the neighbor-state stack \begin{equation} \Delta x_{Z_i}(t_n)=\mathrm{col}\{\Delta x_k(t_n)\,|\,k\in Z_i\}. \label{eq:xZi_clean} \end{equation} Each stand $i$ is manipulated by roll gap $s_i(t)$ and stand speed $v_i(t)$: \begin{equation} u_i(t)= \begin{bmatrix} s_i(t)\\ v_i(t) \end{bmatrix}\in\mathbb{R}^{n_u},\qquad n_u=2. \label{eq:ui_clean} \end{equation} To ensure smooth actuation and match industrial practice, we optimize \emph{discrete input increments}: \begin{equation} \Delta u_i(t_n)\triangleq u_i(t_n)-u_i(t_{n-1}) = \begin{bmatrix} \Delta s_i(t_n)\\ \Delta v_i(t_n) \end{bmatrix}. \label{eq:du_discrete_clean} \end{equation} Throughout this paper, the symbol ``$\Delta$'' attached to inputs $\Delta u_i(t_n)$ means sample-to-sample increment. Thus, $\Delta x$ and $\Delta u$ are conceptually different, and this is fixed by definition. Let $d_i(t)$ denote exogenous disturbances. We denote the interval-level equivalent disturbance by $\Delta d_i(t_n)$. $I_d$ denotes the $d\times d$ identity matrix; $0_{a\times b}$ denotes the $a\times b$ zero matrix. The stand-wise deviation-state evolution over $[t_n,t_{n+1}]$ can be expressed by a discrete-time mapping \begin{equation} \Delta x_i(t_{n+1}) = \Phi_i\Big(\Delta x_i(t_n),\,\Delta x_{Z_i}(t_n),\,\Delta u_i([t_n,t_{n+1}]),\,\Delta d_i([t_n,t_{n+1}])\Big), \label{eq:true_mapping_clean} \end{equation} where $\Phi_i(\cdot)$ is generally nonlinear and coupled due to rolling deformation and tension propagation. A commonly used conceptual equivalent discrete linear form is \begin{equation} \Delta x_i(t_{n+1}) = M_d\,\Delta x_i(t_n) + N_d\,\Delta u_i(t_n) + F_d\,\Delta d_i(t_n), \label{eq:linear_form_concept} \end{equation} where $M_d,N_d,F_d$ represent equivalent discrete-time matrices around operating conditions. In a practical five-stand cold rolling mill, accurately deriving and identifying these matrices and disturbance models from first principles is difficult, due to strong coupling, unmodeled nonlinearities, and time-varying operating regimes. Therefore, this paper aims to learn a high-fidelity approximation of the interval evolution from data and then embed it into distributed MPC. \begin{remark} In fact, due to the existence of complex coupling relationships, it is difficult to directly and accurately establish \eqref{eq:linear_form_concept} based on first principles. Therefore, in this paper, we learn an approximate mapping of \eqref{eq:true_mapping_clean} from data. \end{remark} Although decisions are updated at discrete instants $t_n$, the hydraulic gap and drive systems evolve continuously inside each interval, and abrupt within-interval changes may excite tension oscillations and deteriorate thickness stability. Thus, parameterizing the within-interval increment trajectory by a low-order polynomial: (i) yields a compact finite-dimensional decision representation; (ii) enforces smooth profiles inside the interval; (iii) enables enforcing increment constraints for all $\tau\in[0,\delta_n]$. This is appropriate when $\delta_n$ is not excessively large relative to actuator bandwidth and the within-interval evolution is well approximated by a low-order basis. On the interval $[t_n,t_{n+1}]$, parameterize the control increment trajectory as an $n$-th order polynomial: \begin{equation} \Delta u_{i,n}(\tau;\Gamma_{i,n}) = \sum_{m=0}^{n}\Gamma_{i,nm}\tau^{m}, \qquad \tau\in[0,\delta_n], \label{eq:du_poly_vec_clean} \end{equation} where $\Gamma_{i,nm}\in\mathbb{R}^{n_u}$ are coefficient vectors ($n_u=2$), and $n\ge 1$ is the chosen polynomial order. In simulations, one may select $n=2$ (quadratic) for a good accuracy--complexity trade-off. Component-wise, \eqref{eq:du_poly_vec_clean} corresponds to \begin{equation} \begin{aligned} \Delta s_{i,n}(\tau) &= \sum_{m=0}^{n}\gamma^{(s)}_{i,nm}\tau^{m},\\ \Delta v_{i,n}(\tau) &= \sum_{m=0}^{n}\gamma^{(v)}_{i,nm}\tau^{m}. \end{aligned} \label{eq:du_components_clean} \end{equation} Define the stacked parameter vector \begin{equation} \Gamma_{i,n}\triangleq \big[ (\Gamma_{i,n0})^\top,\, (\Gamma_{i,n1})^\top,\, \ldots,\, (\Gamma_{i,nn})^\top \big]^\top \in\mathbb{R}^{p}, \qquad p=(n+1)n_u. \label{eq:Gamma_clean} \end{equation} Here, $\Gamma_{i,n0}$ is the baseline increment at $\tau=0$, while $\Gamma_{i,nm}$ ($m\ge 1$) describe the higher-order variation rates. Define the interval-averaged equivalent increments as \begin{equation} \begin{aligned} \Delta u_i(t_n) &\triangleq \frac{1}{\delta_n}\int_0^{\delta_n}\Delta u_{i,n}(\tau)\,d\tau,\\ \Delta d_i(t_n) &\triangleq \frac{1}{\delta_n}\int_0^{\delta_n}\Delta d_i(\tau)\,d\tau. \end{aligned} \label{eq:avg_def_clean} \end{equation} With \eqref{eq:du_poly_vec_clean}, the input average has a closed form: \begin{equation} \Delta u_i(t_n)= \sum_{m=0}^{n}\Gamma_{i,nm}\frac{\delta_n^{m}}{m+1}. \label{eq:avg_closed_clean} \end{equation} Let $\mathcal{I}_x$ denote the sampling domain of deviation states $\Delta x_i(t_n)$ and neighbor stacks $\Delta x_{Z_i}(t_n)$, and let $\mathcal{I}_\Gamma$ denote the sampling domain of polynomial parameters $\Gamma_{i,n}$. These domains specify the operating envelope used to generate supervised training data. Given the above parameterization, one training sample is generated on each interval $[t_n,t_{n+1}]$. In addition to the local deviation state, the neighbor deviation states are included to represent inter-stand coupling. The process is summarized in Table~\ref{tab:interval_sample_generation_en}. \begin{table}[t] \centering \small \renewcommand{\arraystretch}{1.15} \caption{Procedure for generating one interval-level sample on $[t_n,t_{n+1}]$ (five-stand coupled mill).} \label{tab:interval_sample_generation_en} \begin{tabularx}{\linewidth}{>{\centering\arraybackslash}p{0.09\linewidth} X} \toprule \textbf{Step} & \textbf{Operation} \\ \midrule 1 & \textbf{State sampling:} sample $\Delta x_i(t_n)$ and $\Delta x_{Z_i}(t_n)$ from $\mathcal{I}_x$. \\ 2 & \textbf{Parameter sampling:} draw $\Gamma_{i,n}\sim\mathcal{I}_\Gamma$ (coefficients for both $\Delta s_{i,n}(\tau)$ and $\Delta v_{i,n}(\tau)$). \\ 3 & \textbf{Control construction:} compute $\Delta u_{i,n}(\tau)$ via \eqref{eq:du_poly_vec_clean}. \\ 4 & \textbf{State propagation:} integrate the \emph{five-stand coupled} mill model on $[t_n,t_{n+1}]$ (e.g., RK4) using the within-interval control trajectory, and record $\Delta x_i(t_{n+1})$. \\ \bottomrule \end{tabularx} \end{table} Accordingly, an interval sample for subsystem $i$ can be represented as \begin{equation} \mathcal{D}_{i,n}=\big\{\Delta x_i(t_n),\ \Delta x_{Z_i}(t_n),\ \Delta u_{i,n}(\tau),\ \Delta x_i(t_{n+1})\big\}. \label{eq:interval_sample_clean} \end{equation} Note that $\Delta u_{i,n}(\tau)$ is fully determined by $(\Gamma_{i,n},\delta_n)$ via \eqref{eq:du_poly_vec_clean}, therefore it is sufficient to store $(\Gamma_{i,n},\delta_n)$ as the learning input. For each subsystem $i$, by repeating the above procedure across multiple intervals and randomized draws, the local one-step training dataset is formed as \begin{equation} \begin{split} S_i=\Big\{& \big(\Delta x_i^{(j)}(t_n),\,\Delta x_{Z_i}^{(j)}(t_n),\,\Delta x_i^{(j)}(t_{n+1});\, \Gamma_{i,n}^{(j)},\,\delta_n^{(j)}\big) \ \Big|\ j=1,\ldots,J \Big\}. \end{split} \label{eq:S_i_clean} \end{equation} Here $J$ is the number of one-step samples for subsystem $i$. The overall dataset for the five-stand mill is denoted by $\{S_i\}_{i=1}^{5}$. The point-cloud visualization of the training dataset is shown in Figure~\ref{2}. \begin{figure*}[htbp] \centering \includegraphics[scale=0.5]{picture/Fig2.pdf} \caption{Point cloud map of the training dataset.}\label{2} \end{figure*} The one-step set $S_i$ is sufficient for one-step regression, but it is not sufficient for training with multi-step rollout loss and reciprocal-consistency regularization, because these objectives require ground-truth deviation-state trajectories over a horizon of $K$ consecutive intervals. Therefore, without changing the single-interval sampling mechanism above, we additionally organize the offline-simulated samples into $K$-step trajectory segments. Specifically, for each starting time $t_n$ we generate a segment of length $K$ by consecutively sampling $\{\Gamma_{i,n+s},\delta_{n+s}\}_{s=0}^{K-1}$ (and the corresponding inputs/disturbances), and integrating the five-stand coupled mill model over $[t_{n+s},t_{n+s+1}]$ for $s=0,\ldots,K-1$. Hence, we obtain the deviation-state sequence $\{\Delta x_i(t_{n+s})\}_{s=0}^{K}$ as well as the neighbor stacks $\{\Delta x_{Z_i}(t_{n+s})\}_{s=0}^{K}$. Define a $K$-step segment sample for subsystem $i$ as \begin{equation} \begin{aligned} \mathcal{W}_{i,n}= \Big\{& \big(\Delta x_i(t_{n+s}),\,\Delta x_{Z_i}(t_{n+s}),\,\Gamma_{i,n+s},\,\delta_{n+s}\big)_{s=0}^{K-1}; \\ &\big(\Delta x_i(t_{n+s+1})\big)_{s=0}^{K-1} \Big\}. \end{aligned} \label{eq:segment_clean} \end{equation} By repeating the above segment generation, we form the multi-step training set \begin{equation} S_i^{(K)}=\Big\{\mathcal{W}_{i,n}^{(j)}\ \Big|\ j=1,\ldots,J_K\Big\}, \label{eq:S_i_K_clean} \end{equation} where $J_K$ is the number of $K$-step segment samples. Note that $S_i$ can be viewed as the marginal one-step projection of $S_i^{(K)}$ (keeping only $s=0$), thus the original dataset design is preserved, and only an additional \emph{segment organization} is introduced for multi-step training. %======================== \section{Construction of Residual Neural Network} %======================== \subsection{Residual Neural Network Structure Construction and Training Method} Given the dataset, the neural network model is trained to learn a stand-wise, control-dependent one-step evolution law of deviation states: \begin{equation} \Delta x_i(t_{n+1}) \approx \Delta x_i(t_n)+ \mathcal{N}_i\!\Big(\Delta x_i(t_n),\,\Delta x_{Z_i}(t_n),\,\Gamma_{i,n},\,\delta_n;\,\Theta_i\Big), \label{eq:learned_dyn_clean} \end{equation} where $\mathcal{N}_i(\cdot)$ outputs the one-step deviation-state change and $\Theta_i$ are trainable parameters. \begin{remark} If $\mathcal{N}_i$ does not take control information as input (here $\Gamma_{i,n}$ and $\delta_n$), the predictor becomes an autoregressive model that only reproduces trajectories under the training input patterns and cannot answer the counterfactual question: ``what will happen if we choose a different roll gap and speed trajectory?'' Since MPC optimizes over candidate decisions, a control-dependent predictor \eqref{eq:learned_dyn_clean} is necessary to evaluate the predicted thickness and tension behavior under different candidate actuator trajectories. \end{remark} Let $d=3$ (state dimension), $|Z_i|$ be the number of neighbors of stand $i$ in \eqref{eq:Zi_clean}, and $p=(n+1)n_u$ in \eqref{eq:Gamma_clean}. Define the input vector \begin{equation} X_{i,\text{in}} \triangleq \big[ \Delta x_i(t_n)^\top,\, \Delta x_{Z_i}(t_n)^\top,\, \Gamma_{i,n}^\top,\, \delta_n \big]^\top \in \mathbb{R}^{d(1+|Z_i|)+p+1}. \label{eq:X_in_clean} \end{equation} The network mapping is \begin{equation} \mathcal{N}_i:\mathbb{R}^{d(1+|Z_i|)+p+1}\rightarrow\mathbb{R}^{d}. \end{equation} To improve training stability and long-horizon rollout robustness, we use a residual form. Let $\hat{I}_i\in\mathbb{R}^{d\times(d(1+|Z_i|)+p+1)}$ be a selection matrix extracting the local state block: \begin{equation} \hat{I}_i = [I_d,\, 0_{d\times(d|Z_i|+p+1)}]. \label{eq:Ihat_clean} \end{equation} Then the one-step predictor is written as \begin{equation} X_{i,\text{out}} = \hat{I}_i X_{i,\text{in}} + \mathcal{N}_i(X_{i,\text{in}}; \Theta_i), \label{eq:res_predict_clean} \end{equation} where $X_{i,\text{out}}$ represents the predicted $\Delta x_i(t_{n+1})$. This structure implements a baseline-plus-correction interpretation: the shortcut propagates the current deviation state $\Delta x_i(t_n)$, while the network learns the correction capturing unmodeled nonlinearities and inter-stand coupling (via $\Delta x_{Z_i}$) under varying operating conditions. To improve robustness when $\delta_n$ varies, we introduce an auxiliary branch inside $\mathcal{N}_i$: \begin{equation} \mathcal{N}_i(X_{i,\text{in}};\Theta_i)\triangleq \psi_i(X_{i,\text{in}};\Theta_{\psi_i}) + \rho_i(X_{i,\text{in}};\theta_i), \label{eq:aux_clean} \end{equation} where $\psi_i(\cdot)$ is a lightweight feedforward branch that captures low-frequency/scale effects strongly related to $\delta_n$, and $\rho_i(\cdot)$ captures the remaining nonlinear coupling corrections. When $\psi_i(\cdot)\equiv 0$, the model reduces to a standard residual network. For the $j$-th sample in \eqref{eq:S_i_clean}, define \begin{equation} X_{i,\text{in}}^{(j)} = \big[ \Delta x_i^{(j)}(t_n),\ \Delta x_{Z_i}^{(j)}(t_n),\ \Gamma_{i,n}^{(j)},\ \delta_n^{(j)} \big]^{\top}, \end{equation} and the supervised residual target \begin{equation} \Delta r_i^{(j)}=\Delta x_i^{(j)}(t_{n+1})-\Delta x_i^{(j)}(t_n). \label{eq:target_clean} \end{equation} To suppress accumulation drift induced by long-horizon recursion and to improve long-term predictive stability, we train the forward predictor jointly with an auxiliary backward residual model and impose a multi-step reciprocal-consistency regularization over a $K$-step segment from $S_i^{(K)}$. Construct a backward residual network \begin{equation} \mathcal{B}_i:\mathbb{R}^{d(1+|Z_i|)+p+1}\rightarrow\mathbb{R}^{d}, \end{equation} parameterized by $\bar{\Theta}_i$. For the backward step associated with interval $[t_n,t_{n+1}]$, define \begin{equation} \begin{aligned} X_{i,\mathrm{in}}^{b} &= \big[ \Delta x_i(t_{n+1}),\ \Delta x_{Z_i}(t_{n+1}),\ \Gamma_{i,n},\ \delta_n \big]^{\top},\\ X_{i,\mathrm{out}}^{b} &= \hat{I}_i X_{i,\mathrm{in}}^{b} + \mathcal{B}_i(X_{i,\mathrm{in}}^{b};\bar{\Theta}_i), \end{aligned} \label{eq:back_clean} \end{equation} where $X_{i,\mathrm{out}}^{b}$ represents the backward estimate of $\Delta x_i(t_n)$. The supervised backward residual target is \begin{equation} \Delta r_i^{b}=\Delta x_i(t_n)-\Delta x_i(t_{n+1}). \end{equation} Given a segment sample $\mathcal{W}_{i,n}\in S_i^{(K)}$, initialize \begin{equation} \Delta \hat{x}_i(t_n)=\Delta x_i(t_n), \end{equation} and recursively apply the forward predictor for $K$ steps: \begin{equation} \begin{aligned} \Delta \hat{x}_i(t_{n+s+1}) &= \Delta \hat{x}_i(t_{n+s}) + \mathcal{N}_i\!\Big( \Delta \hat{x}_i(t_{n+s}),\,\Delta \hat{x}_{Z_i}(t_{n+s}),\, \Gamma_{i,n+s},\,\delta_{n+s};\,\Theta_i \Big),\\ &\qquad s=0,\ldots,K-1. \end{aligned} \label{eq:fwd_roll_clean} \end{equation} Set the terminal condition \begin{equation} \Delta \bar{x}_i(t_{n+K})=\Delta \hat{x}_i(t_{n+K}), \end{equation} and roll back using $\mathcal{B}_i$: \begin{equation} \begin{aligned} \Delta \bar{x}_i(t_{n+s}) &= \hat{I}_i X_{i,\mathrm{in}}^{b}(t_{n+s}) + \mathcal{B}_i\!\Big(X_{i,\mathrm{in}}^{b}(t_{n+s});\,\bar{\Theta}_i\Big), \quad s=K-1,\ldots,0, \end{aligned} \label{eq:bwd_roll_clean} \end{equation} where \begin{equation} X_{i,\mathrm{in}}^{b}(t_{n+s})= \big[ \Delta \bar{x}_i(t_{n+s+1}),\ \Delta \hat{x}_{Z_i}(t_{n+s+1}),\ \Gamma_{i,n+s},\ \delta_{n+s} \big]^{\top}. \end{equation} Define the multi-step reciprocal prediction error \begin{equation} E_i(t_n) = \sum_{s=0}^{K} \left\| \Delta \hat{x}_i(t_{n+s})-\Delta \bar{x}_i(t_{n+s}) \right\|^2. \end{equation} We jointly minimize: \begin{equation} \begin{aligned} L_{\mathrm{1step}}(\Theta_i) &= \frac{1}{J_K}\sum_{j=1}^{J_K}\frac{1}{K}\sum_{s=0}^{K-1} \Big\| \big(\Delta x_i^{(j)}(t_{n+s+1})-\Delta x_i^{(j)}(t_{n+s})\big) -\mathcal{N}_i\!\left( X_{i,\mathrm{in}}^{(j)}(t_{n+s});\Theta_i \right) \Big\|^2,\\[2mm] L_{\mathrm{bwd}}(\bar{\Theta}_i) &= \frac{1}{J_K}\sum_{j=1}^{J_K}\frac{1}{K}\sum_{s=0}^{K-1} \Big\| \big(\Delta x_i^{(j)}(t_{n+s})-\Delta x_i^{(j)}(t_{n+s+1})\big) -\mathcal{B}_i\!\left( X_{i,\mathrm{in}}^{b\,(j)}(t_{n+s});\bar{\Theta}_i \right) \Big\|^2,\\[2mm] L_{\mathrm{msrp}}(\Theta_i,\bar{\Theta}_i) &= \frac{1}{J_K}\sum_{j=1}^{J_K} E_i^{(j)}(t_n),\\[2mm] L_{\mathrm{roll}}(\Theta_i) &= \frac{1}{J_K}\sum_{j=1}^{J_K}\sum_{s=1}^{K} \Big\| \Delta x_i^{(j)}(t_{n+s})-\Delta \hat{x}_i^{(j)}(t_{n+s}) \Big\|^2. \end{aligned} \label{eq:loss_clean} \end{equation} Here, $L_{\mathrm{1step}}$ enforces one-step accuracy; $L_{\mathrm{roll}}$ explicitly suppresses long-horizon drift under recursion; $L_{\mathrm{msrp}}$ regularizes the learned dynamics by enforcing reciprocal consistency between forward and backward rollouts; and $L_{\mathrm{bwd}}$ trains the backward model for the consistency regularization. In implementation, these terms are combined as \begin{equation} L_{\mathrm{total}}=\lambda_1 L_{\mathrm{1step}}+\lambda_2 L_{\mathrm{roll}}+\lambda_3 L_{\mathrm{msrp}}+\lambda_4 L_{\mathrm{bwd}}, \end{equation} where $\lambda_1,\lambda_2,\lambda_3,\lambda_4>0$ are tuned on a validation set. After training, the forward predictor is \begin{equation} \Delta \hat{x}_i(t_{n+1}) = \Delta x_i(t_n) + \mathcal{N}_i\!\Big( \Delta x_i(t_n),\,\Delta x_{Z_i}(t_n),\, \Gamma_{i,n},\,\delta_n;\,\Theta_i^* \Big), \label{eq:pred_clean} \end{equation} and multi-step prediction is obtained by recursive rollout of \eqref{eq:pred_clean}. This learned predictor is the internal model used by the MPC optimizer in the next section. Finally, network parameters are optimized using Adam: \begin{equation} \Theta_{i,t+1} = \Theta_{i,t} - \alpha \frac{\hat{m}_{i,t}}{\sqrt{\hat{v}_{i,t}} + \varepsilon}, \end{equation} where $\alpha$ is the learning rate (we use $\alpha$ to avoid conflict with other symbols), $\hat{m}_{i,t}$ and $\hat{v}_{i,t}$ are bias-corrected moment estimates, and $\varepsilon>0$ is a small constant for numerical stability. Figure~\ref{fig:rnn_logic} illustrates the overall structure. \begin{figure}[htbp] \centering \includegraphics[scale=0.85]{picture/x6.pdf} \caption{Logic diagram of the residual neural network.} \label{fig:rnn_logic} \end{figure} \subsection{Explainability of the residual network} Although the five-stand cold rolling mill involves complex rolling deformation and tension-transport coupling, its stand-wise deviation dynamics can be abstractly described by a coupled nonlinear ODE: \begin{equation} \frac{d}{dt}\Delta x_i(t) = f_i\!\Big(\Delta x_i(t),\,\Delta x_{Z_i}(t),\,u_i(t),\,d_i(t)\Big), \qquad i=1,\ldots,5, \label{eq:mech_ode} \end{equation} where $f_i(\cdot)$ summarizes the mechanistic effects. For a given sampling interval length $\delta_n$ and a within-interval input trajectory $u_i(t_n+\tau)$, the state transition over one interval can be written through an evolution operator: \begin{equation} \Delta x_i(t_{n+1})=\Phi_{i,\delta_n}\Big(\Delta x_i(t_n),\,\Delta x_{Z_i}(t_n),\,u_i([t_n,t_{n+1}]),\,d_i([t_n,t_{n+1}])\Big). \label{eq:evolution_operator_mill} \end{equation} By the fundamental theorem of calculus, \eqref{eq:mech_ode} implies the increment form \begin{equation} \Delta x_i(t_{n+1}) = \Delta x_i(t_n) + \underbrace{\int_{0}^{\delta_n} f_i\!\Big(\Delta x_i(t_n+\tau),\,\Delta x_{Z_i}(t_n+\tau),\,u_i(t_n+\tau),\,d_i(t_n+\tau)\Big)\,d\tau}_{\triangleq\ \varphi_{i,n}}, \label{eq:increment_integral} \end{equation} where $\varphi_{i,n}$ is the one-interval state increment generated by the mechanistic dynamics. Our learned model \eqref{eq:learned_dyn_clean} adopts the same increment form as \eqref{eq:increment_integral}: \begin{equation} \Delta x_i(t_{n+1}) \approx \Delta x_i(t_n) + \mathcal{N}_i\!\Big(\Delta x_i(t_n),\,\Delta x_{Z_i}(t_n),\,\Gamma_{i,n},\,\delta_n;\Theta_i\Big). \end{equation} Here, $\mathcal{N}_i(\cdot)$ plays the role of a data-driven approximation of the integral increment $\varphi_{i,n}$, i.e., it approximates the accumulated effect of the mechanistic dynamics over $[t_n,t_{n+1}]$. This is consistent with the well-known interpretation that a residual network behaves like a one-step time integrator: the identity path propagates the current state, while the residual branch represents the increment over the time lag. Inside each interval, we do not optimize point-wise $u_i(t)$ but parameterize the increment trajectory by $\Delta u_{i,n}(\tau;\Gamma_{i,n})$. Hence, the mechanistic increment $\varphi_{i,n}$ in \eqref{eq:increment_integral} depends on the \emph{whole} within-interval trajectory. Feeding $(\Gamma_{i,n},\delta_n)$ into $\mathcal{N}_i$ is therefore a compact way to represent how different candidate gap/speed trajectories change the integral effect and thus the next thickness--tension state. When $\delta_n$ varies and is not very small, directly learning $\varphi_{i,n}$ may be harder. Motivated by the generalized residual idea, we decompose the increment predictor into two parts in \eqref{eq:aux_clean}: \begin{equation} \mathcal{N}_i(\cdot)=\psi_i(\cdot)+\rho_i(\cdot). \end{equation} Conceptually, $\psi_i(\cdot)$ captures low-frequency and scale effects strongly related to $\delta_n$, while $\rho_i(\cdot)$ captures the remaining nonlinear coupling corrections. This provides a mechanism-consistent interpretation: a baseline increment plus a residual correction that compensates unmodeled nonlinearities and inter-stand coupling. %======================== \section{Nash Equilibrium-Based RNE-DMPC} %======================== The five-stand tandem cold rolling system is strongly coupled through inter-stand tension propagation. As a result, changes in control actions (roll gap and stand speed) at one stand can affect both upstream and downstream stands, making centralized online optimization over all stands' decision variables computationally demanding. To mitigate this issue, we decompose the global predictive-control problem into $N=5$ local subproblems associated with individual stands. Each local controller optimizes its own decision variables while accounting for coupling via limited information exchange with neighboring controllers. Motivated by game-theoretic coordination \citep{rawlings2008coordinating}, we formulate distributed coordination as a Nash-equilibrium-seeking iteration. Based on the trained residual neural network surrogate model, we construct a Nash-equilibrium-based distributed MPC method (RNE-DMPC) for coordinated thickness--tension regulation and tracking. The overall control structure is shown in Figure~\ref{4}. \begin{figure*}[htbp] \centering \includegraphics[width=\linewidth]{picture/x2.pdf} \caption{Schematic diagram of the control architecture for a tandem cold rolling mill.}\label{4} \end{figure*} At sampling time $t_n$, stand $i$ chooses the polynomial-parameter sequence $\mathbf{\Gamma}_i(t_n)\in\mathbb{R}^{pN_c}$, where $p=(n+1)n_u$. Let $\mathbf{\Gamma}(t_n)\triangleq \mathrm{col}\{\mathbf{\Gamma}_1(t_n),\ldots,\mathbf{\Gamma}_5(t_n)\}$ denote the joint strategy profile, and let $\mathbf{\Gamma}_{-i}(t_n)$ denote the collection of all strategies except stand $i$. Given the current measured/estimated deviation state $\Delta x_i(t_n)$ and the strategies $(\mathbf{\Gamma}_i(t_n),\mathbf{\Gamma}_{Z_i}(t_n))$, the multi-step prediction used by stand $i$ is written explicitly as \begin{equation} \begin{aligned} \Delta \hat{x}_i(t_{n+s+1};\mathbf{\Gamma}_i,\mathbf{\Gamma}_{Z_i}) &= \Delta \hat{x}_i(t_{n+s};\mathbf{\Gamma}_i,\mathbf{\Gamma}_{Z_i}) + \mathcal{N}_i\!\Big( \Delta \hat{x}_i(t_{n+s};\cdot),\, \Delta \hat{x}_{Z_i}(t_{n+s};\mathbf{\Gamma}_{Z_i}),\\ &\qquad \Gamma_{i,n+s},\, \delta_{n+s};\Theta_i^* \Big), \end{aligned} \label{eq:rollout_mpc_game} \end{equation} for $s=0,\ldots,N_p-1$, with initialization $\Delta \hat{x}_i(t_n;\cdot)=\Delta x_i(t_n)$. Here the neighbor stack $\Delta \hat{x}_{Z_i}(t_{n+s};\mathbf{\Gamma}_{Z_i})$ is generated from neighbors' strategies via the same learned predictors. Over $[t_{n+s},t_{n+s+1}]$ with length $\delta_{n+s}$, \begin{equation} \Delta u_{i,n+s}(\tau;\Gamma_{i,n+s}) = \sum_{m=0}^{n}\Gamma_{i,n+s,m}\tau^{m}, \qquad \tau \in [0,\delta_{n+s}], \end{equation} and the interval-averaged increment is \begin{equation} \Delta u_i(t_{n+s}) = \sum_{m=0}^{n}\Gamma_{i,n+s,m}\frac{\delta_{n+s}^{m}}{m+1}. \label{eq:du_avg_clean} \end{equation} \begin{remark} Because inter-stand tension $T_i$ is jointly affected by the adjacent stands $i$ and $i+1$, the predicted evolution of $\Delta x_i$ depends on neighbors' future actions, hence the MPC problems are not independent but form a coupled dynamic game. \end{remark} At time $t_n$, the local strategy of stand $i$ is \begin{equation} \mathbf{\Gamma}_i(t_n) = \mathrm{col}\{\Gamma_{i,n},\Gamma_{i,n+1},\ldots,\Gamma_{i,n+N_c-1}\} \in \mathbb{R}^{pN_c}. \end{equation} In deviation coordinates, the regulation/tracking objective is $\Delta x_i(t)\rightarrow 0$, i.e. \begin{equation} \Delta x_{i,\mathrm{ref}}(t_{n+s})\equiv 0\in\mathbb{R}^{d},\qquad d=3. \end{equation} Recall $\Delta x_i=[\Delta h_i,\Delta T_{i-1},\Delta T_i]^\top$. Define the row selectors \begin{equation} C^- \triangleq [0\ \ 1\ \ 0]\in\mathbb{R}^{1\times 3},\qquad C^+ \triangleq [0\ \ 0\ \ 1]\in\mathbb{R}^{1\times 3}, \end{equation} so that $C^- \Delta x_i=\Delta T_{i-1}$ (upstream interface) and $C^+\Delta x_i=\Delta T_i$ (downstream interface). For the interface between stands $i$ and $i+1$, stand $i$'s prediction provides $C^+\Delta \hat{x}_i$, while stand $i+1$'s prediction provides $C^- \Delta \hat{x}_{i+1}$. Their mismatch measures coupling inconsistency: \begin{equation} e_{i}^{\mathrm{sh}}(t_{n+s};\mathbf{\Gamma}) \triangleq C^+\Delta \hat{x}_i(t_{n+s};\mathbf{\Gamma}) - C^-\Delta \hat{x}_{i+1}(t_{n+s};\mathbf{\Gamma}),\qquad i=1,\ldots,4. \label{eq:shared_tension_mismatch} \end{equation} We define the stage cost of stand $i$ as a function of all players' strategies: \begin{equation} J_i(\mathbf{\Gamma}_i;\mathbf{\Gamma}_{-i}) = \sum_{s=1}^{N_p} \left\| \Delta \hat{x}_i(t_{n+s};\mathbf{\Gamma}_i,\mathbf{\Gamma}_{Z_i}) \right\|_{Q_i}^{2} + \sum_{s=0}^{N_c-1} \left\|\Gamma_{i,n+s}\right\|_{R_i}^{2} + J_i^{\mathrm{cpl}}(\mathbf{\Gamma}_i;\mathbf{\Gamma}_{-i}) \label{eq:Ji_game} \end{equation} where $Q_i\succeq 0$ weights thickness and tension deviations, and $R_i\succeq 0$ penalizes actuation magnitudes. The coupling term $J_i^{\mathrm{cpl}}$ explicitly reflects the game/coordination requirement on shared tensions. A simple and effective choice is to penalize the interface mismatches adjacent to stand $i$: \begin{equation} \begin{aligned} J_i^{\mathrm{cpl}} &= \mu_i\sum_{s=1}^{N_p} \Big( \mathbb{I}_{\{i\ge 2\}} \big\| C^-\Delta \hat{x}_i(t_{n+s};\mathbf{\Gamma}) - C^+\Delta \hat{x}_{i-1}(t_{n+s};\mathbf{\Gamma}) \big\|^2 \\ &\qquad\quad + \mathbb{I}_{\{i\le 4\}} \big\| C^+\Delta \hat{x}_i(t_{n+s};\mathbf{\Gamma}) - C^-\Delta \hat{x}_{i+1}(t_{n+s};\mathbf{\Gamma}) \big\|^2 \Big), \end{aligned} \label{eq:coupling_cost} \end{equation} with $\mu_i>0$ and indicator $\mathbb{I}_{\{\cdot\}}$. This term makes the coupling conflict explicit: unilateral actions that locally reduce thickness error may worsen shared-tension compatibility and thus increase $J_i$, and also affect neighbors' objectives. We enforce the absolute-input bounds and within-interval increment bounds. Absolute input bounds (roll gap and speed): \begin{equation} u_{i,\min}\le u_i(t_{n+s})\le u_{i,\max}, \qquad s=0,\ldots,N_p-1, \label{eq:u_abs_game} \end{equation} where $u_i(t)=[s_i(t),\,v_i(t)]^\top$. Within-interval increment-trajectory bounds: \begin{equation} \Delta u_{i,\min}\le \Delta u_{i,n+s}(\tau;\Gamma_{i,n+s})\le \Delta u_{i,\max}, \qquad \forall\tau\in[0,\delta_{n+s}],\ s=0,\ldots,N_p-1, \label{eq:du_traj_game} \end{equation} where $\Delta u_{i,n+s}(\tau;\Gamma_{i,n+s})$ is given by the polynomial parameterization. Consistency with discrete execution: Define the interval-averaged increment \begin{equation} \Delta u_i(t_{n+s}) = \frac{1}{\delta_{n+s}}\int_{0}^{\delta_{n+s}}\Delta u_{i,n+s}(\tau;\Gamma_{i,n+s})\,d\tau = \sum_{m=0}^{n}\Gamma_{i,n+s,m}\frac{\delta_{n+s}^{m}}{m+1}, \label{eq:du_avg_game} \end{equation} and propagate the absolute input along the horizon by \begin{equation} u_i(t_n)=u_i(t_{n-1})+\Delta u_i(t_n),\qquad u_i(t_{n+s})=u_i(t_{n+s-1})+\Delta u_i(t_{n+s}),\ s=1,\ldots,N_p-1, \label{eq:u_prop_game} \end{equation} where $u_i(t_{n-1})$ is the applied (measured) input from the previous sampling instant. Compact feasible set: \begin{equation} \Omega_i \triangleq \Big\{\mathbf{\Gamma}_i\ \Big|\ \eqref{eq:rollout_mpc_game}\ \text{holds and}\ \eqref{eq:u_abs_game},\eqref{eq:du_traj_game},\eqref{eq:u_prop_game}\ \text{are satisfied} \Big\}. \label{eq:Omega_i_game} \end{equation} Given neighbors' current strategies, stand $i$ solves the differentiable NLP: \begin{equation} \mathbf{\Gamma}_i^{\mathrm{BR}} = \arg\min_{\mathbf{\Gamma}_i\in\Omega_i}\ J_i(\mathbf{\Gamma}_i;\mathbf{\Gamma}_{-i}). \label{eq:local_BR} \end{equation} Because the learned surrogate is differentiable, \eqref{eq:local_BR} can be solved by standard gradient-based NLP solvers. At each sampling time $t_n$, the distributed MPC coordination induces a finite-horizon dynamic game: players are stands $i=1,\ldots,5$; strategy sets are $\Omega_i$; and payoff (cost) functions are $J_i(\mathbf{\Gamma}_i;\mathbf{\Gamma}_{-i})$ defined in \eqref{eq:Ji_game}--\eqref{eq:coupling_cost}. A joint strategy profile $\mathbf{\Gamma}^*=\mathrm{col}\{\mathbf{\Gamma}_1^*,\ldots,\mathbf{\Gamma}_5^*\}$ is a Nash equilibrium if \begin{equation} \forall i\in\{1,\ldots,5\},\qquad \mathbf{\Gamma}_i^*\in \arg\min_{\mathbf{\Gamma}_i\in\Omega_i} J_i(\mathbf{\Gamma}_i;\mathbf{\Gamma}_{-i}^*). \label{eq:NE_def} \end{equation} This definition explicitly characterizes the strategic coupling: each player's optimal decision depends on neighbors' decisions through the shared-tension dynamics and the coupling term. To compute an NE online with limited communication, we employ a relaxed best-response iteration. Let $l$ denote the Nash-iteration index. Given $\mathbf{\Gamma}^{(l-1)}$, each stand computes a best response $\mathbf{\Gamma}_i^{\mathrm{BR},(l)}$ by solving \eqref{eq:local_BR}, and then updates with relaxation: \begin{equation} \mathbf{\Gamma}_i^{(l)} = (1-\omega)\mathbf{\Gamma}_i^{(l-1)} + \omega\,\mathbf{\Gamma}_i^{\mathrm{BR},(l)}, \qquad \omega\in(0,1]. \label{eq:relaxed_BR} \end{equation} The relaxation factor $\omega$ mitigates oscillations caused by strong coupling and improves practical convergence. The Nash equilibrium is computed through distributed best-response iterations, summarized in Table ~\ref{tab:nash_iter_en}. The convergence metric in Step F is defined as \begin{equation} \varsigma^{(l)} = \max_i \frac{\left\| \mathbf{\Gamma}_i^{(l)}-\mathbf{\Gamma}_i^{(l-1)} \right\|_2}{ \left\| \mathbf{\Gamma}_i^{(l-1)} \right\|_2+\epsilon}, \end{equation} with $\epsilon>0$ small. \begin{table}[t] \centering \small \renewcommand{\arraystretch}{1.12} \setlength{\tabcolsep}{3.5pt} \caption{Relaxed distributed Nash best-response iteration for RNE-DMPC (five-stand).} \label{tab:nash_iter_en} \begin{tabularx}{\linewidth}{>{\centering\arraybackslash}p{0.11\linewidth} X} \toprule \textbf{Step} & \textbf{Description} \\ \midrule A & Initialize $l=1$ and warm-start $\mathbf{\Gamma}_i^{(0)}$ (e.g., from the previous sampling time). \\ B & Communicate $\mathbf{\Gamma}_i^{(l-1)}$ (or the induced predicted trajectories) among neighbors; form $\mathbf{\Gamma}_{Z_i}^{(l-1)}$. \\ C & Given $\mathbf{\Gamma}_{-i}^{(l-1)}$, solve the best-response NLP \eqref{eq:local_BR} to obtain $\mathbf{\Gamma}_i^{\mathrm{BR},(l)}$. \\ D & Update the relaxed strategy using \eqref{eq:relaxed_BR} and compute the induced predictions $\Delta \hat{x}_i^{(l)}(\cdot;\mathbf{\Gamma}^{(l)})$ via \eqref{eq:rollout_mpc_game}. \\ E & Broadcast $\mathbf{\Gamma}_i^{(l)}$ and the predicted interface quantities needed in \eqref{eq:coupling_cost} (e.g., $C^\pm\Delta \hat{x}_i^{(l)}$) to neighbors. \\ F & Compute $\varsigma^{(l)}$; if $\varsigma^{(l)}\le \varsigma_{\mathrm{tol}}$, stop and set $\mathbf{\Gamma}_i^*=\mathbf{\Gamma}_i^{(l)}$; otherwise $l\leftarrow l+1$ and repeat. \\ \bottomrule \end{tabularx} \end{table} After convergence, only the first-interval parameters $\Gamma_{i,n}^*$ are applied. The within-interval increment is $\Delta u_{i,n}(\tau;\Gamma_{i,n}^*)$ and the discrete applied increment uses the interval average: \begin{equation} \Delta u_i(t_n) = \sum_{m=0}^{n}\Gamma_{i,n,m}^*\frac{\delta_n^{m}}{m+1}, \end{equation} followed by $u_i(t_n)=u_i(t_{n-1})+\Delta u_i(t_n)$. The overall control flow chart is shown in Fig~\ref{liu}. \begin{figure} \centering \includegraphics[width=\linewidth]{picture/x5.pdf} \caption{The overall system control flow chart}\label{liu} \end{figure} \begin{remark} Equations \eqref{eq:Ji_game}--\eqref{eq:coupling_cost} and the NE definition \eqref{eq:NE_def} explicitly show that each player's objective depends on others' strategies through the coupled prediction dynamics and the shared-tension interface mismatch penalty, thus making the multi-stand strategic interaction and coordination mechanism transparent. \end{remark}
指出来都改了哪,都怎么改的
我这次只动了“时变信号的多项式参数化”相关链条,把原来固定的二阶(常数+一次+二次)形式,统一改成“一般 阶”,并把所有依赖它的地方做了一致性联动更新。具体改动点如下(按你文中出现顺序):
(你原文的 \eqref{eq:du_poly_vec_clean})
仍然用同一个标签 \eqref{eq:du_poly_vec_clean} 承接上下文。
标签仍为 \eqref{eq:du_components_clean}。
(\eqref{eq:Gamma_clean})
并把解释文字从“线性/二次变化率”改成“高阶变化率()”。
(\eqref{eq:avg_closed_clean})
标签仍为 \eqref{eq:avg_closed_clean},保证前后引用不乱。
不改结构,只改 的定义来源。
这里是你要求的“所有时变信号参数化都这样改”的关键:我把 MPC 章节里所有沿用二阶形式的地方都改成通式,保持逻辑闭环。
并且在“Consistency with discrete execution”那条等式链里也同步替换,保证等号链仍成立。
如果你愿意,我也可以把“文中新增的那句 ‘仿真取 ’”改成你更偏好的中文/英文表述风格(但不改数学结构)。
新加的变量m原来没有吧,没产生冲突吧
新加的变量m原来没有吧,没产生冲突吧
不会产生“数学/LaTeX 级别的冲突”,原因有两点:
里, 只在该求和符号的作用域内有效,出了这个式子就不“携带意义”,不会和别处的符号发生 LaTeX 编译冲突。
可以把求和下标 换成别的字母,比如 或 。需要同步改动的只是在所有出现这类求和的地方(我改过的这些式子里):
但就你现在用 的版本来说:不冲突、逻辑一致、能正常使用。
k使用过么,没使用过的话给m换成k
k 已经用过了,所以不建议把我新加的求和哑变量换成 k,否则会造成符号复用(读者层面的冲突/混淆,LaTeX 也容易看错含义):
k 明确表示“邻居机架索引”。K 表示多步段长度(rollout horizon),k/K 混在一起更容易误读。在不改其他地方(尤其是不动 \eqref{eq:xZi_clean} 这类已存在定义)的前提下,最稳妥的做法是把我引入的 m 换成 \ell(或 q),避免和你已有的 k、K 产生语义混淆。
m 统一替换成 \ell 的改法(推荐)你需要做的就是把所有我新增的求和指标 m 改成 \ell,对应系数下标也一起改:
把
latex\sum_{m=0}^{n}\Gamma_{i,nm}\tau^{m}
改成
latex\sum_{\ell=0}^{n}\Gamma_{i,n\ell}\tau^{\ell}
把
latex\sum_{m=0}^{n}\gamma^{(s)}_{i,nm}\tau^{m},\quad \sum_{m=0}^{n}\gamma^{(v)}_{i,nm}\tau^{m}
改成
latex\sum_{\ell=0}^{n}\gamma^{(s)}_{i,n\ell}\tau^{\ell},\quad \sum_{\ell=0}^{n}\gamma^{(v)}_{i,n\ell}\tau^{\ell}
把
latex\sum_{m=0}^{n}\Gamma_{i,nm}\frac{\delta_n^{m}}{m+1}
改成
latex\sum_{\ell=0}^{n}\Gamma_{i,n\ell}\frac{\delta_n^{\ell}}{\ell+1}
例如把
latex\sum_{m=0}^{n}\Gamma_{i,n+s,m}\tau^{m} \quad\text{和}\quad \sum_{m=0}^{n}\Gamma_{i,n+s,m}\frac{\delta_{n+s}^{m}}{m+1}
统一改成
latex\sum_{\ell=0}^{n}\Gamma_{i,n+s,\ell}\tau^{\ell} \quad\text{和}\quad \sum_{\ell=0}^{n}\Gamma_{i,n+s,\ell}\frac{\delta_{n+s}^{\ell}}{\ell+1}
那就必须把你原来用作邻居索引的 k(\eqref{eq:xZi_clean})改成别的字母(比如 j),这会牵动你说“不改其他地方”的要求,所以我不建议这么做。
你要是同意我用 \ell,我可以把你那份“完整 LaTeX”版本再发一遍(仅把 m→\ell 这一个点改掉,其他不动)。
还是用m吧,adam那的m换一个字母,换什么好
用 做多项式求和下标没问题;为了避免和 Adam 里“动量一阶矩”记号 混淆,Adam 那里的 我建议换成 (含义直观:gradient 的一阶动量/动量项),改动也最小、基本不和你文中其他符号抢意义。
你只需要把 Adam 这段里所有的 (一阶矩)替换掉即可:
把原来这句
latex\Theta_{i,t+1} = \Theta_{i,t} - \alpha \frac{\hat{m}_{i,t}}{\sqrt{\hat{v}_{i,t}} + \varepsilon},
改成
latex\Theta_{i,t+1} = \Theta_{i,t} - \alpha \frac{\hat{g}_{i,t}}{\sqrt{\hat{v}_{i,t}} + \varepsilon},
并把后面解释文字里对应这一句
latex$\hat{m}_{i,t}$ and $\hat{v}_{i,t}$ are bias-corrected moment estimates,
改成
latex$\hat{g}_{i,t}$ and $\hat{v}_{i,t}$ are bias-corrected moment estimates,
补充一句:其实 Adam 里的 和你控制量里的速度 也“长得像”,但一个带帽且下标是 ,另一个是速度 ,通常读者还能区分;如果你也想彻底消除这种视觉复用,我也可以顺手把 Adam 的 一并换成别的(比如 )。
\section{Simulations}
To verify the effectiveness of the proposed residual-neural-network and Nash-equilibrium based distributed MPC (RNE-DMPC) for coordinated thickness–tension regulation in a five-stand cold tandem rolling system, a distributed control architecture composed of multiple local controllers and a communication/coordination module is considered. Each stand is treated as a subsystem, and information exchange through the communication system enforces coordination among subsystems, enabling distributed thickness–tension control for the interconnected plant.
In this study, the simulation data were generated using the mathematical model reported by Hu et al. \citep{hu2021nash}.
\subsection{Tracking under multiple reference changes}
\begin{figure*}
\centering
\includegraphics[scale=0.5]{picture/simulation.pdf}
\caption{Tension–thickness tracking responses under multiple reference changes: (a) shows the tension tracking response; (b) shows the thickness tracking response.}\label{simulation}
\end{figure*}
In Figure~\ref{simulation}, three reference targets were modified sequentially during the simulation for the five-stand thickness--tension coupled system, and the resulting closed-loop responses were recorded.
It can be observed that the tension and thickness outputs of all stands were driven toward the updated references promptly, while overall stability was maintained.
At each switching instant, short-lived deviations of the tracking errors were induced due to the strong inter-stand coupling; however, these deviations were rapidly attenuated and kept at small levels afterwards.
Such behavior indicates that effective dynamic tracking and coordination were achieved under multivariable coupling and repeated setpoint switching.
Moreover, the contribution of the distributed architecture was reflected by the information exchange enabled by the communication mechanism, through which consistency across subsystems was enforced and coordinated regulation was maintained throughout the successive reference updates.
\subsection{Thickness and tension step change response}
In Figure~\ref{thickness}, a single step change was imposed on the thickness reference, and both the thickness trajectories and the tracking errors were evaluated.
Following the step, the thickness outputs were guided toward the new reference levels with a fast transition, and the tracking errors were suppressed quickly and maintained within a small range, implying strong transient regulation and high steady-state accuracy in the thickness channel.
In particular, after the responses reached steady state, the absolute tracking error of the outlet thickness (target value ) was , corresponding to a relative error of 2.28%.
This response is consistent with the objective design in which reference tracking and input regularization are jointly penalized over the prediction horizon.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{picture/thickness.pdf}
\caption{Thickness responses and tracking errors under a step change of the thickness reference: (a) shows the thickness response; Figure (b) shows the tracking error.}\label{thickness}
\end{figure*}
In Figure~\ref{tension}, a single step change was applied to the tension reference, and the corresponding tension responses and error evolutions were examined.
After the reference update, the tension outputs were pulled toward the new setpoints rapidly, and the errors were driven to small values without being accompanied by sustained oscillations, indicating satisfactory damping and closed-loop stability.
After steady state, the absolute tracking error of the tension (target value ) was , corresponding to a relative error of 1.91%.
\begin{figure*}
\centering
\includegraphics[width=\linewidth]{picture/tension.pdf}
\caption{Tension responses and tracking errors under a step change of the tension reference: (a) shows the tension response; (b) shows the tracking error}\label{tension}
\end{figure*}
\subsection{Computation rate comparison}
Through thickness and tension simulations, the computation time of each control cycle was recorded .
Figure~\ref{rate} compares the computation time of the proposed RNE-DMPC with a Nash-equilibrium-based distributed predictive control baseline \citep{hu2021nash}.
Statistical results show that the per-cycle computation time of both Nash-type distributed controllers remains small, which is suitable for online receding-horizon implementation.
More importantly, the proposed RNE-DMPC consistently yields a noticeably lower per-cycle computation time than the Nash baseline in Ref.~\citep{hu2021nash}, indicating faster equilibrium seeking and coordination in online operation.
Overall, the results indicate that RNE-DMPC offers excellent computational performance and can effectively reduce system load and implementation cost.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/rate.pdf}
\caption{Comparison of calculation rates for different control strategies: (a) Step disturbance in outlet thickness; (b) Step disturbance in tension.}\label{rate}
\end{figure*}
这个是我目前的仿真部分,在这个基础上,我还要加两个实验。一个实验是我的残差网络与其他网络预测的一个比对,一个实验是我用nash均衡和不用nash均衡的跟踪精度的比对,两个实验的结果都是我的更好。帮我把整个仿真部分理清一下思路,看看应该怎么写最合适,尽量分点不要太多
下面给你一个更顺、也更容易让审稿人接受的仿真章节写法思路:把“效果验证”按两条主线展开——模型(残差网络)是否好、控制(Nash 协调)是否好。你现在已有的三组结果(多参考切换、厚度/张力阶跃、计算时间)很适合放在“控制效果+工程可用性”这条主线里;新增的两个实验刚好分别对应两条主线各补一块。
在开头段落后加一句类似:
这样后面新增两组实验就“有名分”,不会显得硬塞。
把这一小节放在所有闭环控制结果之前,逻辑最顺:模型好 → MPC 预测可靠 → 控制好。
\subsection{Prediction performance comparison of surrogate models}\subsection{Residual network prediction accuracy comparison}这部分写完,用一两句“因此后续控制全部采用该 surrogate”把话接到控制实验。
这个实验的目的不是证明“DMPC能用”,而是证明Nash 协调机制是必要且更优的,尤其在强耦合张力共享界面上。
\subsection{Effect of Nash-equilibrium coordination}\subsection{Tracking comparison with and without Nash coordination}你可以选一种最自然、最容易实现且可复现的 baseline,二选一:
建议你选并行一次更新,更贴合“有没有 Nash equilibrium seeking”的主题。
图也建议一张多子图就够:
这一节最后一句话把结论扣回理论:Nash 迭代实质上在在线上逼近耦合一致的策略剖面,因此在强耦合下能显著提升协调质量与跟踪精度。
你现在三小节其实很完整,只需要把它们放到“控制层验证”之后,形成一个自然递进:
这一块建议轻微改一句:说明这些结果是在“采用 Nash 协调的 RNE-DMPC”下得到的,与上一小节对比形成闭环。
你可以把仿真部分改成下面这个框架(保持小节数量不爆炸):
\subsection{Prediction performance comparison of surrogate models}(新增:残差网络 vs 其他网络)\subsection{Tracking comparison with and without Nash coordination}(新增:有 Nash vs 无 Nash)\subsection{Tracking under multiple reference changes}(你已有)\subsection{Thickness and tension step change response}(你已有)\subsection{Computation rate comparison}(你已有)这样一共 5 个小节,不算多,但主线非常清晰:模型优 → 协调优 → 跟踪好 → 计算快。
如果你愿意,我也可以直接按这个结构,把你现有三段文字原句尽量不动,只做“承上启下”的桥接句,并给你新增两小节的可直接粘贴的英文 LaTeX 文本模板(含推荐图注、指标表达、和你论文符号一致的 用法)。
把第三点和第四点合成一点,先写第一点,在写第三点和第四点,然后写写第二点,最后写第五点。网络的仿真我只有一个预测轨迹叠图,指标你看看用什么;nash均衡的仿真我有一个跟踪误差曲线(Nash 更小)的图片,指标你看看用什么。帮我写一个完整的仿真部分,中英文都要
下面给你一份**“可直接粘贴”的完整仿真部分 LaTeX(中英文各一套)**,结构顺序按你要求:
另外:
picture/pred_compare.pdf 和 picture/nash_error_compare.pdf,你按自己文件名替换即可。latex\section{仿真验证} 为验证本文提出的残差神经网络与基于 Nash 均衡的分布式模型预测控制(RNE-DMPC)在五机架冷连轧厚度--张力协调控制中的有效性, 本文采用由多个局部控制器与通信/协调模块构成的分布式控制架构。每个机架视为一个子系统,邻近子系统通过通信机制交换必要信息, 从而在强耦合的互联系统上实现分布式厚度--张力协调调节与跟踪。 本文仿真数据由 Hu 等人报道的五机架冷连轧数学模型生成 \citep{hu2021nash}。 \subsection{预测模型精度对比:残差网络 vs 其他网络} 为验证所提出残差网络在一阶预测与多步滚动预测中的有效性, 将本文模型与两类代表性基线模型进行比较:\emph{(i) 普通前馈网络(MLP)}(相近参数规模,采用直接一阶回归); \emph{(ii) 序列模型(如 LSTM/GRU)}(以历史序列为输入的预测器)。 所有模型均在相同训练/测试划分下训练,并在测试集上评估预测性能。 图~\ref{fig:pred_compare} 给出了一个代表性工况下的预测轨迹叠加结果。可以观察到,本文残差网络预测轨迹与真实轨迹吻合程度更高, 且在多步滚动预测时漂移更小。为定量比较,本文采用两项指标:\textbf{one-step MAE}(一步预测平均绝对误差)以及 \textbf{$K$-step rollout RMSE}($K$ 步滚动预测的均方根误差)。相应统计结果汇总于表~\ref{tab:pred_metrics}, 可以看出本文残差网络在上述指标上均优于对比模型。 \begin{figure*}[htbp] \centering \includegraphics[width=\linewidth]{picture/pred_compare.pdf} \caption{预测轨迹叠加对比(代表性工况):真实轨迹与不同网络模型预测结果对比。}\label{fig:pred_compare} \end{figure*} \begin{table}[t] \centering \small \renewcommand{\arraystretch}{1.15} \caption{不同预测模型在测试集上的误差指标对比(数值由你填入)。} \label{tab:pred_metrics} \begin{tabular}{lcc} \toprule \textbf{模型} & \textbf{one-step MAE} & \textbf{$K$-step rollout RMSE} \\ \midrule 本文残差网络(RNN) & -- & -- \\ MLP 基线 & -- & -- \\ 序列模型基线(LSTM/GRU) & -- & -- \\ \bottomrule \end{tabular} \end{table} \subsection{闭环跟踪性能:多参考切换与阶跃响应} 本小节在闭环条件下综合验证 RNE-DMPC 的动态跟踪能力。首先考虑多参考目标连续切换的情形, 其次考察厚度与张力参考分别发生单次阶跃变化时的瞬态与稳态性能。 如图~\ref{simulation} 所示,在五机架厚度--张力强耦合系统中,仿真过程中依次改变三组参考目标并记录闭环响应。 可以观察到,各机架厚度与张力均能较快跟踪新的参考值,同时整体稳定性保持良好。 由于机架间强耦合,在每次切换瞬间会出现短暂的误差扰动,但扰动可快速衰减并维持在较小范围内, 表明在多变量耦合与反复设定值切换下仍能实现有效的动态跟踪与协调。 \begin{figure*}[htbp] \centering \includegraphics[scale=0.5]{picture/simulation.pdf} \caption{多参考切换下的厚度--张力跟踪响应:(a) 张力跟踪;(b) 厚度跟踪。}\label{simulation} \end{figure*} 进一步地,图~\ref{thickness} 给出了厚度参考发生单次阶跃变化时的厚度响应与误差曲线。 阶跃发生后,厚度输出可快速过渡至新参考附近,跟踪误差迅速抑制并维持在较小范围内,表明厚度通道具有较强的瞬态调节能力与稳态精度。 尤其是在稳态后,出口厚度(目标值 $1.27~\mathrm{mm}$)的绝对跟踪误差为 $0.029~\mathrm{mm}$,相对误差为 2.28\%。 \begin{figure*}[htbp] \centering \includegraphics[width=\linewidth]{picture/thickness.pdf} \caption{厚度参考阶跃下的厚度响应与跟踪误差:(a) 厚度响应;(b) 跟踪误差。}\label{thickness} \end{figure*} 图~\ref{tension} 给出了张力参考发生单次阶跃变化时的张力响应与误差曲线。 参考更新后,张力输出可快速收敛至新设定值,误差在无持续振荡的情况下被压制到较小水平,表明闭环具有良好的阻尼特性与稳定性。 稳态后张力(目标值 $1.1~\mathrm{kN}$)的绝对跟踪误差为 $0.021~\mathrm{kN}$,相对误差为 1.91\%。 \begin{figure*}[htbp] \centering \includegraphics[width=\linewidth]{picture/tension.pdf} \caption{张力参考阶跃下的张力响应与跟踪误差:(a) 张力响应;(b) 跟踪误差。}\label{tension} \end{figure*} \subsection{Nash 协调机制的作用:使用 Nash vs 不使用 Nash} 为突出 Nash 均衡迭代在强耦合系统分布式协调中的作用, 本文进一步比较“使用 Nash 均衡协调”的 RNE-DMPC 与“不使用 Nash 协调”的分布式 MPC 基线的跟踪精度。 其中,“不使用 Nash 协调”指各机架在每个采样时刻仅进行一次局部优化更新, 并将邻居机架策略(或界面预测量)固定为上一时刻/上一轮结果,不进行在线 best-response 迭代协调。 图~\ref{fig:nash_compare} 给出了两种策略下的跟踪误差曲线对比,可以观察到使用 Nash 协调时误差整体更小、收敛更快。 为进行简洁量化,本文采用 \textbf{RMSE($e$)} 与 \textbf{Max$|e|$} 两项指标分别衡量误差能量与峰值误差, 统计结果可汇总于表~\ref{tab:nash_metrics},由此可验证 Nash 协调机制能显著提升耦合系统的在线协调质量与跟踪精度。 \begin{figure*}[htbp] \centering \includegraphics[width=\linewidth]{picture/nash_error_compare.pdf} \caption{使用 Nash 协调与不使用 Nash 协调的跟踪误差对比(本文方法误差更小)。}\label{fig:nash_compare} \end{figure*} \begin{table}[t] \centering \small \renewcommand{\arraystretch}{1.15} \caption{使用/不使用 Nash 协调的跟踪误差指标对比(数值由你填入)。} \label{tab:nash_metrics} \begin{tabular}{lcc} \toprule \textbf{方法} & \textbf{RMSE($e$)} & \textbf{Max$|e|$} \\ \midrule RNE-DMPC(使用 Nash 协调) & -- & -- \\ 分布式 MPC 基线(不使用 Nash 协调) & -- & -- \\ \bottomrule \end{tabular} \end{table} \subsection{计算时间对比} 在厚度与张力跟踪仿真中,记录每个控制周期的计算时间。 图~\ref{rate} 对比了本文 RNE-DMPC 与文献 \citep{hu2021nash} 的基于 Nash 的分布式预测控制方法的计算耗时。 统计结果表明,两种 Nash 类型分布式控制器的单周期计算时间均较小,适用于在线滚动优化实现。 更重要的是,本文 RNE-DMPC 的单周期计算时间整体更低,体现出更快的在线协调与均衡求解效率,从而降低系统负载与实现成本。 \begin{figure*}[htbp] \centering \includegraphics[width=\linewidth]{picture/rate.pdf} \caption{不同控制策略的计算速度对比:(a) 出口厚度阶跃扰动;(b) 张力阶跃扰动。}\label{rate} \end{figure*}
latex\section{Simulations} To verify the effectiveness of the proposed residual-neural-network and Nash-equilibrium based distributed MPC (RNE-DMPC) for coordinated thickness--tension regulation in a five-stand cold tandem rolling system, a distributed control architecture composed of multiple local controllers and a communication/coordination module is considered. Each stand is treated as a subsystem, and information exchange through the communication system enforces coordination among subsystems, enabling distributed thickness--tension control for the interconnected plant. In this study, the simulation data were generated using the mathematical model reported by Hu et al.~\citep{hu2021nash}. \subsection{Surrogate-model prediction comparison: residual network vs other networks} To justify the use of the proposed residual predictor in MPC, we first compare its prediction accuracy against representative baseline models, including \emph{(i) a plain feedforward network (MLP)} with comparable model size trained by one-step regression, and \emph{(ii) a sequence model (e.g., LSTM/GRU)} trained for trajectory prediction. All models are trained and tested under the same dataset split. Figure~\ref{fig:pred_compare} shows an illustrative trajectory overlay on a representative test case. It can be observed that the proposed residual network produces a closer match to the ground truth and exhibits less drift in multi-step rollouts. For a concise quantitative assessment, we report two metrics: \textbf{one-step MAE} and \textbf{$K$-step rollout RMSE}. The aggregated results can be summarized in Table~\ref{tab:pred_metrics}, where the proposed residual network achieves the lowest prediction errors among the compared models. \begin{figure*}[htbp] \centering \includegraphics[width=\linewidth]{picture/pred_compare.pdf} \caption{Prediction trajectory overlay (representative case): ground truth versus different network predictors.}\label{fig:pred_compare} \end{figure*} \begin{table}[t] \centering \small \renewcommand{\arraystretch}{1.15} \caption{Prediction error metrics on the test set (fill in your numbers).} \label{tab:pred_metrics} \begin{tabular}{lcc} \toprule \textbf{Model} & \textbf{one-step MAE} & \textbf{$K$-step rollout RMSE} \\ \midrule Proposed residual network (RNN) & -- & -- \\ MLP baseline & -- & -- \\ Sequence baseline (LSTM/GRU) & -- & -- \\ \bottomrule \end{tabular} \end{table} \subsection{Closed-loop tracking performance: multiple reference changes and step responses} This subsection evaluates the closed-loop tracking performance of RNE-DMPC under (i) multiple sequential reference changes and (ii) single step changes in thickness and tension references. In Figure~\ref{simulation}, three reference targets were modified sequentially during the simulation for the five-stand thickness--tension coupled system, and the resulting closed-loop responses were recorded. It can be observed that the tension and thickness outputs of all stands were driven toward the updated references promptly, while overall stability was maintained. At each switching instant, short-lived deviations of the tracking errors were induced due to the strong inter-stand coupling; however, these deviations were rapidly attenuated and kept at small levels afterwards. Such behavior indicates that effective dynamic tracking and coordination were achieved under multivariable coupling and repeated setpoint switching. Moreover, the contribution of the distributed architecture was reflected by the information exchange enabled by the communication mechanism, through which consistency across subsystems was enforced and coordinated regulation was maintained throughout the successive reference updates. \begin{figure*}[htbp] \centering \includegraphics[scale=0.5]{picture/simulation.pdf} \caption{Tension--thickness tracking responses under multiple reference changes: (a) shows the tension tracking response; (b) shows the thickness tracking response.}\label{simulation} \end{figure*} In Figure~\ref{thickness}, a single step change was imposed on the thickness reference, and both the thickness trajectories and the tracking errors were evaluated. Following the step, the thickness outputs were guided toward the new reference levels with a fast transition, and the tracking errors were suppressed quickly and maintained within a small range, implying strong transient regulation and high steady-state accuracy in the thickness channel. In particular, after the responses reached steady state, the absolute tracking error of the outlet thickness (target value $1.27~\mathrm{mm}$) was $0.029~\mathrm{mm}$, corresponding to a relative error of 2.28\%. \begin{figure*}[htbp] \centering \includegraphics[width=\linewidth]{picture/thickness.pdf} \caption{Thickness responses and tracking errors under a step change of the thickness reference: (a) shows the thickness response; (b) shows the tracking error.}\label{thickness} \end{figure*} In Figure~\ref{tension}, a single step change was applied to the tension reference, and the corresponding tension responses and error evolutions were examined. After the reference update, the tension outputs were pulled toward the new setpoints rapidly, and the errors were driven to small values without being accompanied by sustained oscillations, indicating satisfactory damping and closed-loop stability. After steady state, the absolute tracking error of the tension (target value $1.1~\mathrm{kN}$) was $0.021~\mathrm{kN}$, corresponding to a relative error of 1.91\%. \begin{figure*}[htbp] \centering \includegraphics[width=\linewidth]{picture/tension.pdf} \caption{Tension responses and tracking errors under a step change of the tension reference: (a) shows the tension response; (b) shows the tracking error.}\label{tension} \end{figure*} \subsection{Impact of Nash coordination: with Nash vs without Nash} To highlight the role of Nash-equilibrium coordination in strongly coupled distributed control, we compare the proposed RNE-DMPC (with relaxed best-response iterations) against a distributed MPC baseline without Nash coordination. Specifically, the baseline performs only a single local update per sampling instant while keeping neighbors' strategies (or interface predictions) fixed from the previous update, i.e., no online best-response iterations are carried out. Figure~\ref{fig:nash_compare} shows the tracking error comparison, where the Nash-coordinated scheme yields consistently smaller errors and faster convergence. For a compact quantitative summary, we report \textbf{RMSE($e$)} and \textbf{Max$|e|$}, which capture the error energy and peak error, respectively. The results can be summarized in Table~\ref{tab:nash_metrics}, confirming that Nash coordination improves the tracking accuracy in the coupled multi-stand system. \begin{figure*}[htbp] \centering \includegraphics[width=\linewidth]{picture/nash_error_compare.pdf} \caption{Tracking error comparison with and without Nash coordination (the proposed Nash-coordinated scheme yields smaller errors).}\label{fig:nash_compare} \end{figure*} \begin{table}[t] \centering \small \renewcommand{\arraystretch}{1.15} \caption{Tracking error metrics with and without Nash coordination (fill in your numbers).} \label{tab:nash_metrics} \begin{tabular}{lcc} \toprule \textbf{Method} & \textbf{RMSE($e$)} & \textbf{Max$|e|$} \\ \midrule RNE-DMPC (with Nash coordination) & -- & -- \\ Distributed MPC baseline (without Nash) & -- & -- \\ \bottomrule \end{tabular} \end{table} \subsection{Computation rate comparison} Through thickness and tension simulations, the computation time of each control cycle was recorded. Figure~\ref{rate} compares the computation time of the proposed RNE-DMPC with a Nash-equilibrium-based distributed predictive control baseline~\citep{hu2021nash}. Statistical results show that the per-cycle computation time of both Nash-type distributed controllers remains small, which is suitable for online receding-horizon implementation. More importantly, the proposed RNE-DMPC consistently yields a noticeably lower per-cycle computation time than the Nash baseline in Ref.~\citep{hu2021nash}, indicating faster equilibrium seeking and coordination in online operation. Overall, the results indicate that RNE-DMPC offers excellent computational performance and can effectively reduce system load and implementation cost. \begin{figure*}[htbp] \centering \includegraphics[width=\linewidth]{picture/rate.pdf} \caption{Comparison of calculation rates for different control strategies: (a) step disturbance in outlet thickness; (b) step disturbance in tension.}\label{rate} \end{figure*}
如果你把两张新增图的真实文件名告诉我(或直接贴 \includegraphics{...} 你现在用的路径),我可以把占位名改成你项目里一致的命名;另外你如果不想放表格,我也能把两处指标改成“文中一句话给出数值”的写法(更短)。
我要引用一下MLP和LSTM参考文献,帮我找一下
下面给你两类最常用、最“标准”的引用:MLP(前馈多层感知机)和LSTM,都附上可直接用的 BibTeX。
bibtex@article{Hornik1989MLPUniversal, title = {Multilayer feedforward networks are universal approximators}, author = {Hornik, Kurt and Stinchcombe, Maxwell and White, Halbert}, journal = {Neural Networks}, volume = {2}, number = {5}, pages = {359--366}, year = {1989}, doi = {10.1016/0893-6080(89)90020-8} } @article{Rumelhart1986Backprop, title = {Learning representations by back-propagating errors}, author = {Rumelhart, David E. and Hinton, Geoffrey E. and Williams, Ronald J.}, journal = {Nature}, volume = {323}, pages = {533--536}, year = {1986}, doi = {10.1038/323533a0} } @article{Hochreiter1997LSTM, title = {Long Short-Term Memory}, author = {Hochreiter, Sepp and Schmidhuber, J{\"u}rgen}, journal = {Neural Computation}, volume = {9}, number = {8}, pages = {1735--1780}, year = {1997}, doi = {10.1162/neco.1997.9.8.1735} }
对应来源信息我核对过:Rumelhart(1986) 的卷期与页码在 Nature 页面可见,Hornik(1989) 的卷期页码与 DOI 在 ScienceDirect/BibTeX 页面可见,LSTM(1997) 的卷期页码在论文 PDF 首页可见。 (Nature)
如果你论文里“MLP baseline”只想放一个引用:更推荐用 Hornik1989MLPUniversal(直接对应“MLP 作为通用逼近器”这个语境);若强调“用 BP 训练”,再加 Rumelhart1986Backprop。
多步的RMSE一般是多少
没有一个“通用标准值”。多步 rollout 的 RMSE 大小主要取决于:你预测的量(厚度/张力、是否是偏差量)、采样周期 、预测步长 、数据扰动强度、以及你有没有做归一化/反归一化。
更常用、也更容易写进论文的做法是把它规范化成百分比(NRMSE),给出“通常认为不错/很好”的区间:
在工程系统(尤其强耦合、滚动预测)里,常见的经验是:
你只要写“在 xx 步时 NRMSE 为 x%(更低)”就很清楚、也不依赖绝对单位。
你文里闭环稳态跟踪误差大概是 厚度 0.029 mm、张力 0.021 kN。多步预测 RMSE 通常会比一步误差更大一些(随 增长),所以很多工作里会出现类似量级:
你现在只有一张轨迹叠图,表格里我建议填:
这样读者不会纠结“到底 RMSE 应该是多少”,只会看到:你的更小、而且是跨变量可比的百分比。
下面是按你截图那种样式改好的 LaTeX 表格代码(只保留 RMSE,3 个 state:)。
(需要 \usepackage{booktabs,multirow})
latex% in preamble: % \usepackage{booktabs} % \usepackage{multirow} \begin{table}[t] \centering \small \renewcommand{\arraystretch}{1.15} \setlength{\tabcolsep}{8pt} \caption{RMSE comparison of multi-step prediction for each state.} \label{tab:pred_rmse_states} \begin{tabular}{lccc} \toprule \multirow{2}{*}{Methods} & \multicolumn{3}{c}{RMSEs of each state} \\ \cmidrule(lr){2-4} & State $x_1$ & State $x_2$ & State $x_3$ \\ \midrule Proposed method & 0.0411 & 0.0832 & 0.0967 \\ Method in~\cite{ref27} & 0.2658 & 0.6463 & 0.1424 \\ Method in~\cite{ref30} & 0.0442 & 0.0981 & 0.1397 \\ \bottomrule \end{tabular} \end{table}
如果你不想用 \cite{}(就想写死成 [27] [30]),把两行改成下面这样即可:
latexMethod in [27] & 0.2658 & 0.6463 & 0.1424 \\ Method in [30] & 0.0442 & 0.0981 & 0.1397 \\
\multirow这个报错了
\\multirow 报错最常见原因是:没加载 multirow 宏包,于是会出现 Undefined control sequence \multirow。
把下面两行放到 \documentclass... 后、\begin{document} 前:
latex\usepackage{booktabs} \usepackage{multirow}
然后你的表格就能正常编译。
如果你不想引入 multirow(或模板不兼容),用下面这个版本:完全不用 \multirow,版式仍和你截图一致。
latex% \usepackage{booktabs} % 需要 booktabs \begin{table}[t] \centering \small \renewcommand{\arraystretch}{1.15} \setlength{\tabcolsep}{8pt} \caption{RMSE comparison of multi-step prediction for each state.} \label{tab:pred_rmse_states} \begin{tabular}{lccc} \toprule Methods & \multicolumn{3}{c}{RMSEs of each state} \\ \cmidrule(lr){2-4} & State $x_1$ & State $x_2$ & State $x_3$ \\ \midrule Proposed method & 0.0411 & 0.0832 & 0.0967 \\ Method in~\cite{ref27} & 0.2658 & 0.6463 & 0.1424 \\ Method in~\cite{ref30} & 0.0442 & 0.0981 & 0.1397 \\ \bottomrule \end{tabular} \end{table}
如果你不用
\cite{}、想显示[27] [30],把那两行的\cite{ref27}改成[27]即可。
你把报错信息(哪一行、完整错误)贴一下的话,我也能帮你判断是不是“缺包”之外的情况(比如模板禁用某些命令)。
methods能不能居中
可以。你现在第一列用的是 l(左对齐),把它改成 c(居中)就行。
下面给你两种常用写法(都能让 Methods 这一列居中):
把 \begin{tabular}{lccc} 改成 \begin{tabular}{cccc}:
latex\begin{table}[t] \centering \small \renewcommand{\arraystretch}{1.15} \setlength{\tabcolsep}{8pt} \caption{RMSE comparison of multi-step prediction for each state.} \label{tab:pred_rmse_states} \begin{tabular}{cccc} \toprule Methods & \multicolumn{3}{c}{RMSEs of each state} \\ \cmidrule(lr){2-4} & State $x_1$ & State $x_2$ & State $x_3$ \\ \midrule Proposed method & 0.0411 & 0.0832 & 0.0967 \\ Method in~\cite{ref27} & 0.2658 & 0.6463 & 0.1424 \\ Method in~\cite{ref30} & 0.0442 & 0.0981 & 0.1397 \\ \bottomrule \end{tabular} \end{table}
把 lccc 改成 c c c c(效果同上),或者更明确一点写:
latex\begin{tabular}{c c c c}
如果你后面觉得第一列太宽、居中后不好看,我也可以给你一个“固定宽度且居中”的第一列版本(用 p{} + 居中),但一般上面这两种就够了。
分布式 MPC 基线(不使用 Nash 协调),帮我找一个文献,我要引用
可以引用一篇非常经典、被广泛用来代表“非博弈/非 Nash 的分布式 MPC(DMPC)架构”的综述论文:
Riccardo Scattolini, Architectures for distributed and hierarchical Model Predictive Control – A review, Journal of Process Control, 19(5):723–731, 2009, doi:10.1016/j.jprocont.2009.02.003. (ScienceDirect)
对应 BibTeX(可直接粘贴):
bibtex@article{Scattolini2009DMPCReview, title = {Architectures for distributed and hierarchical Model Predictive Control -- A review}, author = {Scattolini, Riccardo}, journal = {Journal of Process Control}, volume = {19}, number = {5}, pages = {723--731}, year = {2009}, doi = {10.1016/j.jprocont.2009.02.003} }
你在文中写“Distributed MPC baseline (without Nash coordination)”时,就可以在第一次出现 baseline 的地方加 \cite{Scattolini2009DMPCReview}。
有没有有仿真的那种,在帮我找一篇
有的。给你一篇带仿真例子、而且是典型“分布式/协同 MPC(非 Nash 均衡)”的参考文献,拿来当你文中“distributed MPC baseline (without Nash coordination)”的引用很合适:
Stewart, Venkat, Rawlings, Wright, Pannocchia, “Cooperative distributed model predictive control”, Systems & Control Letters, 59(8):460–469, 2010. DOI: 10.1016/j.sysconle.2010.06.005. (dblp)
(文末有示例仿真,对比 cooperative 与 noncooperative / decentralized 策略。)(UW Computer Sciences)
BibTeX:
bibtex@article{Stewart2010CoopDMPC, title = {Cooperative distributed model predictive control}, author = {Stewart, Brett T. and Venkat, Aswin N. and Rawlings, James B. and Wright, Stephen J. and Pannocchia, Gabriele}, journal = {Systems \& Control Letters}, volume = {59}, number = {8}, pages = {460--469}, year = {2010}, doi = {10.1016/j.sysconle.2010.06.005} }
Camponogara et al., “Distributed Model Predictive Control”, IEEE Control Systems Magazine, 22(1):44–52, Feb. 2002(文中包含两区域负荷频率控制的示例仿真)。(ScienceDirect)
如果你更希望 baseline 是“非迭代/一次通信的 DMPC”那类,我也可以再帮你找一篇更贴近你“无 Nash(单次更新)”定义的论文(带仿真)。
latexThe root mean square error (RMSE) is adopted, which is shown as follows: \begin{equation} \mathrm{RMSE} = \sqrt{\frac{1}{N}\sum_{i=1}^{N}\left(x_i-\hat{x}_i\right)^2 }. \end{equation} where $N$ is the number of samples, $x_i$ and $\hat{x}_i$ are the actual and predicted values, respectively. Then, the prediction results are given in Table~II.
\begin{figure*}[htbp]
\centering
\subfigure[State x1]{\includegraphics[scale=0.4]{picture/pred1.pdf}\label{F1_pred}}
\subfigure[State x2]{\includegraphics[scale=0.4]{picture/pred2.pdf}\label{F2_pred}}
\subfigure[State x3]{\includegraphics[scale=0.4]{picture/pred3.pdf}\label{F3_pred}}
\caption{Prediction trajectory overlay.}\label{fig:pred_compare}
\end{figure*}我这个为啥报错
你这段最常见的报错原因是:模板里没有定义 \subfigure(也就是你没加载 subfigure/subfig 相关宏包,或模板禁用了旧的 subfigure 包)。于是会出现类似:
Undefined control sequence \subfigure\subfigure 与某些类(IEEEtran 等)冲突另外一个小问题是标题里写 x1 不算错误,但更规范是写成数学下标 。
subcaption(不要用 \subfigure)导言区加:
latex\usepackage{graphicx} \usepackage{subcaption}
然后把你的图改成:
latex\begin{figure*}[htbp] \centering \begin{subfigure}[b]{0.32\textwidth} \centering \includegraphics[width=\linewidth]{picture/pred1.pdf} \caption{State $x_1$} \label{F1_pred} \end{subfigure} \hfill \begin{subfigure}[b]{0.32\textwidth} \centering \includegraphics[width=\linewidth]{picture/pred2.pdf} \caption{State $x_2$} \label{F2_pred} \end{subfigure} \hfill \begin{subfigure}[b]{0.32\textwidth} \centering \includegraphics[width=\linewidth]{picture/pred3.pdf} \caption{State $x_3$} \label{F3_pred} \end{subfigure} \caption{Prediction trajectory overlay.} \label{fig:pred_compare} \end{figure*}
\subfigure)导言区至少要有:
latex\usepackage{graphicx} \usepackage{subfigure} % 老包,部分模板会冲突/不推荐
但很多期刊/IEEE 模板会直接不建议或不兼容 subfigure,所以还是建议用上面的 subcaption。
如果报错是 File 'picture/pred1.pdf' not found,那就是路径/文件名问题(跟 subfigure 无关)。
如果你把具体报错信息贴一行出来,我也能精准告诉你是哪一种情况。
还是有问题You have written a length, but have not specified the appropriate units (pt, mm, cm etc.). If you have not written a length, check that you have not witten a linebreak \ followed by square brackets […] anywhere.
这个报错 99% 是因为某处写了“长度但没单位”——最典型就是把
0.32\textwidth 写成了 0.32textwidth(少了反斜杠),或写成 width=0.32(没有 \linewidth/\textwidth 或 cm/pt 等单位)。你现在这段用的是旧的 \subfigure,我建议直接换成更稳的 subfig 的 \subfloat(IEEE/期刊模板兼容性最好),并且我把长度写成 0.32\linewidth,只要你别漏掉反斜杠就不会再触发这个单位错误。
导言区加(如果模板是 IEEEtran,通常要求 caption=false):
latex\usepackage{graphicx} \usepackage[caption=false,font=footnotesize]{subfig}
正文用:
latex\begin{figure*}[htbp] \centering \subfloat[State $x_1$]{% \includegraphics[width=0.32\linewidth]{picture/pred1.pdf}% \label{F1_pred}% }\hfill \subfloat[State $x_2$]{% \includegraphics[width=0.32\linewidth]{picture/pred2.pdf}% \label{F2_pred}% }\hfill \subfloat[State $x_3$]{% \includegraphics[width=0.32\linewidth]{picture/pred3.pdf}% \label{F3_pred}% } \caption{Prediction trajectory overlay.} \label{fig:pred_compare} \end{figure*}
关键点:
width=0.32\linewidth里 必须有\linewidth的反斜杠,否则就会报你那个 “no appropriate units” 的错。
导言区:
latex\usepackage{graphicx} \usepackage{subcaption}
正文(注意 {0.32\textwidth} 里也必须有 \textwidth 的反斜杠):
latex\begin{figure*}[htbp] \centering \begin{subfigure}[b]{0.32\textwidth} \centering \includegraphics[width=\linewidth]{picture/pred1.pdf} \caption{State $x_1$} \label{F1_pred} \end{subfigure}\hfill \begin{subfigure}[b]{0.32\textwidth} \centering \includegraphics[width=\linewidth]{picture/pred2.pdf} \caption{State $x_2$} \label{F2_pred} \end{subfigure}\hfill \begin{subfigure}[b]{0.32\textwidth} \centering \includegraphics[width=\linewidth]{picture/pred3.pdf} \caption{State $x_3$} \label{F3_pred} \end{subfigure} \caption{Prediction trajectory overlay.} \label{fig:pred_compare} \end{figure*}
如果你愿意,把你完整报错前后两三行贴出来(尤其是那一行附近有没有 0.32textwidth 这种),我还能帮你精确定位是哪一处漏了单位/反斜杠。
RNE-DMPC跟踪精度的RMSE通常是多少
没有一个固定“通常值”,RMSE 会随下面这些因素变化很大:参考变化幅度、耦合强度、采样周期/预测步长、噪声/扰动、以及你算 RMSE 的时间窗口(全程 vs 去掉切换瞬间)。
更稳妥的经验说法是用**相对 RMSE(百分比)**来给“通常水平”,论文里也更好写:
把误差 做
工程/文献里常见的经验区间(给审稿人直觉):
这里“reference magnitude”你可以用:设定值 、量程 、或测试集标准差 (三选一,写清楚即可)。
在你这种五机架厚度–张力跟踪里,文献里“看起来比较合理”的量级往往是:
但这两条只是“量级直觉”,不如用 rRMSE/NRMSE 更靠谱。
如果你把你那张 “Nash 更小” 的误差曲线对应的误差定义(是厚度误差?张力误差?还是综合误差?)和参考值量级(比如 1.27 mm、1.1 kN)告诉我,我可以帮你选最合适的 rRMSE 归一化方式,并给一段可以直接写进论文的英文表述。
我现在有一张图里有五张子图,帮我漂亮的排版,给出latex代码
下面给你一个五张子图的“3+2”排版(第一行 3 张,第二行 2 张居中),在双栏/单栏都比较好看、也不容易报长度单位错误。推荐用 subfig 的 \subfloat(兼容性通常最好)。
latex\usepackage{graphicx} \usepackage[caption=false,font=footnotesize]{subfig} % 若非IEEE模板,可去掉 caption=false
latex\begin{figure*}[htbp] \centering % -------- Row 1: three subfigures -------- \subfloat[Subfigure 1]{% \includegraphics[width=0.32\linewidth]{picture/fig1.pdf}% \label{fig:five_a}% }\hfill \subfloat[Subfigure 2]{% \includegraphics[width=0.32\linewidth]{picture/fig2.pdf}% \label{fig:five_b}% }\hfill \subfloat[Subfigure 3]{% \includegraphics[width=0.32\linewidth]{picture/fig3.pdf}% \label{fig:five_c}% }\\[2mm] % -------- Row 2: two subfigures (centered) -------- \subfloat[Subfigure 4]{% \includegraphics[width=0.48\linewidth]{picture/fig4.pdf}% \label{fig:five_d}% }\hfill \subfloat[Subfigure 5]{% \includegraphics[width=0.48\linewidth]{picture/fig5.pdf}% \label{fig:five_e}% } \caption{Your overall caption here.} \label{fig:five_subplots} \end{figure*}
你只需要把 picture/fig1.pdf … picture/fig5.pdf 换成你的真实文件名,把子图标题 Subfigure 1 等换成比如 State $x_1$、Stand 1 之类即可。
关键避免报错点:
width=0.32\linewidth、0.48\linewidth里的\linewidth一定要有反斜杠。
可以把“厚度”和“张力”分成两组指标列,用分组表头最清晰、也最像你截图那种风格。下面给你一个推荐版(两行表头、四个指标列),直接把 “--” 换成你的数值即可。
需要
booktabs(\usepackage{booktabs})
latex% Preamble: % \usepackage{booktabs} \begin{table}[t] \centering \small \renewcommand{\arraystretch}{1.15} \setlength{\tabcolsep}{10pt} \caption{Tracking accuracy comparison (thickness vs tension).} \label{tab:track_acc_ht} \begin{tabular}{lcccc} \toprule \textbf{Method} & \multicolumn{2}{c}{\textbf{Thickness error}} & \multicolumn{2}{c}{\textbf{Tension error}} \\ \cmidrule(lr){2-3}\cmidrule(lr){4-5} & $\mathrm{RMSE}(e_h)$ & $\max|e_h|$ & $\mathrm{RMSE}(e_T)$ & $\max|e_T|$ \\ \midrule RNE-DMPC (with Nash) & -- & -- & -- & -- \\ Distributed MPC (w/o Nash) & -- & -- & -- & -- \\ \bottomrule \end{tabular} \end{table}
如果你想在表头里把单位也写清楚(更规范),可以这样写(按你实际单位改):
latex& $\mathrm{RMSE}(e_h)$ (mm) & $\max|e_h|$ (mm) & $\mathrm{RMSE}(e_T)$ (kN) & $\max|e_T|$ (kN) \\
需要我按你论文里到底用的是出口厚度 还是某一机架厚度、张力是**** 还是选某一个关键张力,帮你把 的定义写到 caption 里吗?
我这里面的精度是五个机架的还是说某个单独机架的
你现在这个表格(每种方法只给 两列厚度 + 两列张力)没法同时代表 5 个机架各自的精度,除非你把它定义成“汇总指标”(跨机架/跨张力的聚合)。所以这里有两种写法,二选一(建议用第 1 种,最稳、最不容易被审稿人挑)。
适用于你要证明“用 Nash 整体更准”,也最符合分布式协调的主旨。
你在文中把误差定义清楚:
然后表里的两类指标建议这样“汇总”:
这样你表里四个数就是“全局统计”,逻辑完全闭合。
适用于你只关心“出口厚度(5 号机架)”和“某一段张力(比如 )”。
那你必须在表标题/正文写清楚:
否则审稿人会问:你表里的厚度到底是哪一架?张力到底是哪一段?
既然你说要“区分张力和厚度的跟踪精度”并比较 RNE-DMPC vs 不用 Nash,推荐方案 1(整体汇总):它能体现“全系统协调更好”,也不会因为你挑了某一架/某一段而被质疑“挑结果”。
如果你愿意,我也可以按你最终选的方案,把表格 caption和RMSE 定义段落一起写成可直接粘贴的 LaTeX(中英文各一份)。
我现在这个表格里要写五个机架的均值稳态误差,帮我修改表格,并且修改下面相应的文字:Figure~\ref{nash_compare_thickness} and Figure~\ref{nash_compare_tension} present the comparison of tracking errors for thickness and tension, where the Nash-coordinated scheme yields consistently smaller errors and faster convergence.
For a compact quantitative summary, we report \textbf{RMSE()} and \textbf{Max}, which capture the error energy and peak error, respectively.
The results can be summarized in Table~\ref{tab:nash_metrics}, confirming that Nash coordination improves the tracking accuracy in the coupled multi-stand system.
下面给你一套改好的表格 + 对应文字(把原来的 RMSE/Max 指标改为“五个机架的均值稳态误差”,并区分厚度与张力)。
说明:厚度误差做 5 个机架平均;张力误差做 4 段机架间张力平均()。
你图里原来的数我先原样放进来;如果这些数不是稳态均值,你再替换即可。
latex% Preamble: % \usepackage{booktabs} \begin{table}[t] \centering \small \renewcommand{\arraystretch}{1.15} \setlength{\tabcolsep}{10pt} \caption{Mean steady-state tracking errors averaged over all stands (thickness: 5 stands; tension: 4 inter-stand spans).} \label{tab:nash_metrics} \begin{tabular}{lcc} \toprule \textbf{Method} & \textbf{Thickness error} & \textbf{Tension error} \\ \cmidrule(lr){2-2}\cmidrule(lr){3-3} & $\overline{|e_h|}_{\mathrm{ss}}$ (mm) & $\overline{|e_T|}_{\mathrm{ss}}$ (kN) \\ \midrule RNE-DMPC (with Nash) & 0.0411 & 0.0714 \\ Distributed MPC (w/o Nash) & 0.0503 & 0.0706 \\ \bottomrule \end{tabular} \end{table}
如果你想在表头明确“5 stands / 4 spans”,也可以把第二行改成:
latex& $\overline{|e_h|}_{\mathrm{ss}}$ (mm, avg. over 5 stands) & $\overline{|e_T|}_{\mathrm{ss}}$ (kN, avg. over 4 spans) \\
把你原来那段:
For a compact quantitative summary, we report RMSE(e) and Max|e|...
替换为下面这段(直接粘贴用):
latexFigure~\ref{nash_compare_thickness} and Figure~\ref{nash_compare_tension} present the comparison of tracking errors for thickness and tension, where the Nash-coordinated scheme yields consistently smaller errors and faster convergence. For a compact quantitative summary, we report the \textbf{mean steady-state absolute tracking errors} averaged over all stands, denoted by $\overline{|e_h|}_{\mathrm{ss}}$ for thickness and $\overline{|e_T|}_{\mathrm{ss}}$ for tension. Specifically, $\overline{|e_h|}_{\mathrm{ss}}$ is obtained by averaging $|e_{h,i}|$ over the steady-state window and then taking the mean across the five stands ($i=1,\ldots,5$), while $\overline{|e_T|}_{\mathrm{ss}}$ is obtained similarly by averaging $|e_{T,i}|$ and then taking the mean across the four inter-stand tensions ($i=1,\ldots,4$). The results are summarized in Table~\ref{tab:nash_metrics}, confirming that Nash coordination improves the tracking accuracy in the coupled multi-stand system.
如果你告诉我“稳态窗口”你是怎么取的(比如最后 5 秒/最后 200 个采样点/去掉切换后前 1 秒),我还可以把上面那句 “steady-state window” 写得更具体、更像论文里的严谨表述。
误差进入±5%阈值后到结束
好的,那“稳态窗口”就可以明确写成:误差首次进入并保持在 阈值内之后,到仿真结束。我帮你把表格 caption(可选)和下面英文文字一起改得更严谨、可直接粘贴。
latex% Preamble: % \usepackage{booktabs} \begin{table}[t] \centering \small \renewcommand{\arraystretch}{1.15} \setlength{\tabcolsep}{10pt} \caption{Mean steady-state absolute tracking errors averaged over all stands, where the steady-state window is defined from the first time the tracking error enters and remains within the $\pm5\%$ band until the end of the simulation (thickness: 5 stands; tension: 4 inter-stand spans).} \label{tab:nash_metrics} \begin{tabular}{lcc} \toprule \textbf{Method} & \textbf{Thickness error} & \textbf{Tension error} \\ \cmidrule(lr){2-2}\cmidrule(lr){3-3} & $\overline{|e_h|}_{\mathrm{ss}}$ (mm) & $\overline{|e_T|}_{\mathrm{ss}}$ (kN) \\ \midrule RNE-DMPC (with Nash) & 0.0411 & 0.0714 \\ Distributed MPC (w/o Nash) & 0.0503 & 0.0706 \\ \bottomrule \end{tabular} \end{table}
说明:
- :先在稳态窗口内对每个机架 的 做时间平均,再对 5 个机架取平均。
- :同理,对 4 段张力 处理。
把你原来那段替换为下面这段:
latexFigure~\ref{nash_compare_thickness} and Figure~\ref{nash_compare_tension} present the comparison of tracking errors for thickness and tension, where the Nash-coordinated scheme yields consistently smaller errors and faster convergence. For a compact quantitative summary, we report the \textbf{mean steady-state absolute tracking errors} averaged over all stands, denoted by $\overline{|e_h|}_{\mathrm{ss}}$ for thickness and $\overline{|e_T|}_{\mathrm{ss}}$ for tension. The steady-state window is defined from the first time the tracking error enters and remains within the $\pm5\%$ band until the end of the simulation. Specifically, $\overline{|e_h|}_{\mathrm{ss}}$ is computed by time-averaging $|e_{h,i}|$ over the steady-state window and then taking the mean across the five stands ($i=1,\ldots,5$), while $\overline{|e_T|}_{\mathrm{ss}}$ is computed similarly and then averaged across the four inter-stand tensions ($i=1,\ldots,4$). The results are summarized in Table~\ref{tab:nash_metrics}, confirming that Nash coordination improves the tracking accuracy in the coupled multi-stand system.
如果你希望“±5%阈值”更严格(比如要求连续保持 个采样点才算进入稳态),我也可以把这句改成更工程化的版本。
把第二行两张子图的 width=0.48\linewidth 改成和第一行一样的 0.32\linewidth,然后用 \makebox[\linewidth][c]{...} 把这一行整体居中即可:
latex\begin{figure*}[htbp] \centering % -------- Row 1: three subfigures -------- \subfloat[Stand 1]{% \includegraphics[width=0.32\linewidth]{picture/fig1.pdf}% \label{fig:five_a}% }\hfill \subfloat[Stand 2]{% \includegraphics[width=0.32\linewidth]{picture/fig2.pdf}% \label{fig:five_b}% }\hfill \subfloat[Stand 3]{% \includegraphics[width=0.32\linewidth]{picture/fig3.pdf}% \label{fig:five_c}% }\\[2mm] % -------- Row 2: two subfigures (same size as row 1, centered) -------- \makebox[\linewidth][c]{% \subfloat[Stand 4]{% \includegraphics[width=0.32\linewidth]{picture/fig4.pdf}% \label{fig:five_d}% }\hspace{0.04\linewidth}% \subfloat[Stand 5]{% \includegraphics[width=0.32\linewidth]{picture/fig5.pdf}% \label{fig:five_e}% }% } \caption{Your overall caption here.} \label{fig:five_subplots} \end{figure*}
你只需要把
picture/fig*.pdf换成你的实际文件名即可。
第二行的\hspace{0.04\linewidth}也可以微调(比如 0.03 或 0.05)来控制两张图之间的间距。
平均稳态误差一般是多少
没有一个统一“标准值”,因为你现在把稳态窗口定义成“误差进入并保持在 ±5% 阈值后到结束”,那平均稳态误差会强烈依赖:
不过在你这种五机架厚度–张力跟踪场景里,论文里“看起来合理/常见”的量级通常可以按相对误差和绝对量纲两种方式给个直觉:
如果你希望表里的数字更“有说服力”,建议同时在文中给一句相对误差(例如把 除以厚度参考、 除以张力参考),读者马上就能判断这是 1% 还是 4%。
RNE-DMPC & 2.9 & 0.0411 & -- & 0.0714 \
Distributed MPC & -- & 0.0503 & -- & 0.0706 \这个2.9是2.9%,怎么打上去
在 LaTeX 里百分号要写成转义的 \%,所以 2.9% 写成 2.9\%。
你这一行就改成:
latexRNE-DMPC & 2.9\% & 0.0411 & -- & 0.0714 \\ Distributed MPC & -- & 0.0503 & -- & 0.0706 \\
如果你想让 “%” 也跟数字一样紧凑,可以用同样写法;不需要额外宏包。
% interactapasample.tex
% v1.05 - August 2017
\documentclass[]{interact}
% 让你仍然可以用 \citep / \citet 这类 natbib 命令
\usepackage[natbibapa,nodoi]{apacite}
\setlength\bibhang{12pt}
\renewcommand\bibliographytypesize{\fontsize{10}{12}\selectfont}
\usepackage{epstopdf}% To incorporate .eps illustrations using PDFLaTeX, etc.
\usepackage[caption=false]{subfig}% Support for small, sub' figures and tables %\usepackage[nolists,tablesfirst]{endfloat}% To separate' figures and tables from text if required
%\usepackage[doublespacing]{setspace}% To produce a `double spaced' document if required
%\setlength\parindent{24pt}% To increase paragraph indentation when line spacing is doubled
\usepackage{booktabs}
\usepackage[numbers,longnamesfirst,sort,authoryear,round]{natbib}
\bibpunct[, ]{(}{)}{;}{a}{,}{,}% Citation support using natbib.sty
\usepackage{tabularx}
\usepackage{array}
\usepackage{amsmath,amssymb}
\usepackage{graphicx}
\usepackage[caption=false,font=footnotesize]{subfig}
\usepackage{wrapfig}
\usepackage{multirow}
\usepackage[final]{graphicx} % 关键:即使 class 里是 draft,也强制显示图片
\bibpunct[, ]{(}{)}{;}{a}{,}{,}% Citation support using natbib.sty
\renewcommand\bibfont{\fontsize{10}{12}\selectfont}% To set the list of references in 10 point font using natbib.sty
%\usepackage[natbibapa,nodoi]{apacite}% Citation support using apacite.sty. Commands using natbib.sty MUST be deactivated first!
%\setlength\bibhang{12pt}% To set the indentation in the list of references using apacite.sty. Commands using natbib.sty MUST be deactivated first!
%\renewcommand\bibliographytypesize{\fontsize{10}{12}\selectfont}% To set the list of references in 10 point font using apacite.sty. Commands using natbib.sty MUST be deactivated first!
\theoremstyle{plain}% Theorem-like structures provided by amsthm.sty
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{proposition}[theorem]{Proposition}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{example}[theorem]{Example}
\theoremstyle{remark}
\newtheorem{remark}{Remark}
\newtheorem{notation}{Notation}
\begin{document}
\articletype{ARTICLE TEMPLATE}% Specify the article type or omit as appropriate
\title{Distributed MPC Control for Tandem Cold Rolling Based on Residual Network and Nash Equilibrium}
\author{
\name{Chenhao Guo\textsuperscript{a}, Jiahao Nie\textsuperscript{a}, Xiaojian Li\textsuperscript{a}\textsuperscript{b}\thanks{CONTACT Xiaojian Li. Email: [email protected]}}
\affil{\textsuperscript{a}College of Information Science and Engineering, Northeastern University, Shenyang, China; \textsuperscript{b}State Key Laboratory of Synthetical Automation for Process
Industries, Northeastern University, Shenyang, China}
}
\maketitle
\begin{abstract}
With the rapid development of multi-stand tandem cold rolling technology, high-precision and high-efficiency thickness–tension regulation has become increasingly important for product quality and process stability. However, the multi-stand cold rolling system has a complex structure. Not only is it difficult to establish the mechanism model, but there are also coupling relationships among the stands. This coupling relationships have led to the problem of optimization conflicts among multiple racks. Given that the complexity and coupling of the system, this study proposes a distributed model predictive control strategy based on residual neural network and Nash equilibrium(RNE-DMPC). A ResNet is trained to capture the input-output behaviors of the system, avoiding reliance on an explicit mechanism model. For online constrained multivariable optimization, a Nash-equilibrium-based distributed MPC is developed to coordinate multi-objective tension–thickness regulation across stands. The performance of the proposed method is verified by a series numerical simulations.
\end{abstract}
\begin{keywords}
Residual neural network; Nash equilibrium; Distributed model predictive control; Tandem cold rolling; Tension control; Thickness control
\end{keywords}
\section{Introduction}
Cold-rolled strip steel is a crucial intermediate product in the strip-steel manufacturing chain and significantly affects downstream process performance and final product quality \citep{mancini2012effects}. It is primarily produced by tandem cold rolling mills. During the process, hot-rolled strip passes sequentially through multiple stands. In each stand, a pair of work rolls applies a rolling force to the material, reducing the strip thickness progressively from stand to stand. Figure~\ref{AutoReg} illustrates a typical layout of a five-stand tandem cold rolling mill.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/x3.pdf}
\caption{Equipment layout of a five-stand tandem cold rolling mill.}\label{AutoReg}
\end{figure*}
Tandem cold rolling is a fundamental process in modern strip production, where the final product quality is largely determined by the capability to regulate exit thickness and inter-stand tensions. From a control perspective, thickness–tension regulation is a typical large-scale multivariable problem featuring strong nonlinearities, constraints, time-varying parameters, transport delays, and disturbances. These characteristics not only complicate accurate prediction of rolling dynamics but also make high-performance closed-loop control difficult to sustain over changing operating conditions.
Owing to its long-established industrial use and ease of implementation, PID control remains prevalent in practice. However, the fixed parameter mechanism of PID leads to insufficient flexibility and often struggles when dealing with strongly coupled multivariable processes \citep{gupta2025grid,liu2025data,hasilci2025superior}.
To overcome the limitations of traditional PID control, various advanced control strategies have been investigated \citep{tezuka2001application,raftery2010online,geddes2002improvements,dong2025mpc}. Model Predictive Control (MPC) is adopted as the control strategy. MPC can uniformly handle multi-variable coupling and constraint tracking, and has been widely reported in cold rolling applications \citep{yaghini2024linear,luis2019trajectory,han2021hierarchical,song2025model,sun2017dynamic}. Depending on system scale and information-interaction requirements, MPC can be classified into three typical structures: centralized, decentralized, and distributed architectures \citep{song2025model}.
Distributed MPC decomposes a large-scale coupled optimization problem into several smaller local MPC problems solved in parallel with limited information exchange, thereby reducing computational burden while improving scalability and coordination with reduced computational burden for interconnected systems. Scattolini systematically reviewed the distributed and hierarchical MPC architecture, clarifying that through local optimization and information interaction, global performance can be maintained while reducing computational and communication costs \citep{scattolini2009architectures}. Multi-stand tandem cold rolling is naturally an interconnected large-scale system, making distributed optimization particularly suitable. \citep{zhao2025multi,ji2022coordinate,wang2025distributed}.
However, mechanistic models are typically adopted for prediction in most MPC systems, and extensive parameter identification and calibration are usually required. When the type of steel, the dynamic friction coefficient, speed, or thermal state change, model mismatch is likely to occur, and such errors can significantly undermine the effectiveness of the predictive controller.
Meanwhile, as the availability of industrial data has improved, data-driven modeling has advanced rapidly in the rolling field. Recent reviews indicate that machine learning has been applied to various tasks in tandem cold rolling, while highlighting challenges in interpretability, robustness, and deployment. The application of machine learning in cold continuous rolling has been thoroughly examined, with recent progress and industrial implementation issues summarized \citep{zong2025machine}. In the direction of predictive control, Ma and Li proposed a data-driven predictive control strategy for cold continuous rolling, using GRU modeling and providing analysis of model convergence and control stability within the framework \citep{ma2025data}. Nevertheless, pure black-box models are difficult to integrate with rolling knowledge, and their generalization across operating conditions and practical reliability can be limited.
Along this line, neural-network models identified from input--output data have been explored \citep{xu2024novel,cheng2015neural,xu2022control,han2019data,feng2025model}. At present, Niu et al. have developed a distributed nonlinear model predictive control based on RBF neural network for tandem cold rolling process \citep{niu2025rbf}; Huang et al. proposed the LSTM-AGD multi-mode adaptive MPC framework, which reduces overshoot and improves accuracy in a steady state \citep{huang2022lstm}; Xia et al. used BFGS to optimize the ANN topology and data-driven high-precision prediction of key parameters in cold continuous rolling \citep{xia2023using}. On the other hand, many existing network structures are relatively shallow or unidirectional, with poor network interpretability and which may lead to error accumulation. Additionally, when dealing with the multi-dimensional data of five stands, training and online computation can become challenging due to high-dimensional signals. Therefore, to address these issues, this paper integrates the residual neural network into MPC for prediction. This data-based deep residual network can model dynamic evolution processes,aiming to improve prediction accuracy and interpretability for thickness and tension regulation.
However, in multi-stand scenarios, the strong coupling caused by the tension transmission between stands can lead to mutual interference in local control. The action of one stand will transmit and affect other stands, thereby limiting the further improvement of performance. Therefore, it is necessary to introduce an online coordination optimization mechanism at the system level. From an optimization perspective, game theory offers a principled framework for handling multi-agent decision-making with coupled objectives and constraints, making it well-suited for cooperative optimal control. Hu et al. proposed a thickness-tension distributed predictive control strategy based on Nash equilibrium, decomposing the global control task into interacting sub-problems and achieving coordination through Nash equilibrium \citep{hu2021nash}. At the algorithmic theoretical level, Ye and Hu studied the optimal search for full-distributed Nash equilibrium in network games and established stability stability guarantees under standard assumptions, providing theoretical support for equilibrium coordination without centralized computing \citep{ye2021adaptive}. Although some studies have introduced Nash equilibrium into the distributed predictive control of thickness tension, this iterative process is highly dependent on the accuracy of the prediction model and the shared information. Once the on-site operating conditions change and lead to an increase in prediction errors, the equilibrium iterative effect increased prediction error may slow convergence or lead to termination at an approximate equilibrium, thereby degrading coordination quality and constraint satisfaction.
Therefore, this paper proposes a distributed model predictive control strategy based on residual neural networks and Nash equilibrium (RNE-DMPC).
The main contributions of this article are as follows:
\begin{enumerate}
\item A residual neural network model is developed to capture the input--output relationship, aiming to improve prediction accuracy while enhancing model interpretability.
\item A Nash-equilibrium-based distributed coordination mechanism is integrated with the residual nerual network prediction model to construct RNE-DMPC, aiming to alleviate degraded equilibrium iterations, slow convergence, or convergence to an approximate equilibrium under varying operating conditions.
\end{enumerate}
The rest of this article is organized as follows: Section 2 describes the construction of the dataset. Section 3 builds the residual neural network. Section 4 discusses the specific content of RNE-DMPC. Section 5 presents the simulation results of the thickness and tension of the five-rack system. Section 6 presents the conclusion of the full text.
%========================
\section{Construction of Dataset}
%========================
Consider a five-stand tandem cold rolling mill indexed by .
Let denote the -th sampling instant and the corresponding sampling interval length.
Define the local within-interval time variable .
Let denote the exit thickness of stand (),
and let denote the inter-stand strip tension between stand and ().
The system is strongly coupled because the inter-stand tensions propagate along the mill line and are affected by neighboring stands' actions.
Let and be the desired references (setpoints) given by process requirements. We define deviation variables
\begin{equation}
\Delta h_i(t)\triangleq h_i(t)-h_i^{\mathrm{ref}}(t),\qquad
\Delta T_i(t)\triangleq T_i(t)-T_i^{\mathrm{ref}}(t).
\label{eq:dev_def}
\end{equation}
Throughout this paper, the symbol ``'' attached to states always means deviation from reference.
For each stand , we choose the local deviation state as
\begin{equation}
\Delta x_i(t)\triangleq
\begin{bmatrix}
\Delta h_i(t)\
\Delta T_{i-1}(t)\
\Delta T_i(t)
\end{bmatrix}\in\mathbb{R}^{d},\qquad d=3,
\label{eq:xi_def_clean}
\end{equation}
with the boundary convention and to keep a unified dimension for all stands.
For a five-stand tandem mill, the dominant coupling is between adjacent stands, hence we define
\begin{equation}
Z_1={2},\quad
Z_i={i-1,i+1}\ (i=2,3,4),\quad
Z_5={4}.
\label{eq:Zi_clean}
\end{equation}
Define the neighbor-state stack
\begin{equation}
\Delta x_{Z_i}(t_n)=\mathrm{col}{\Delta x_k(t_n),|,k\in Z_i}.
\label{eq:xZi_clean}
\end{equation}
\begin{bmatrix}
\Delta s_i(t_n)\
\Delta v_i(t_n)
\end{bmatrix}.
\label{eq:du_discrete_clean}
\end{equation}
Throughout this paper, the symbol ``'' attached to inputs means sample-to-sample increment.
Thus, and are conceptually different, and this is fixed by definition.
Let denote exogenous disturbances.
We denote the interval-level equivalent disturbance by .
denotes the identity matrix; denotes the zero matrix.
M_d,\Delta x_i(t_n)
+
N_d,\Delta u_i(t_n)
+
F_d,\Delta d_i(t_n),
\label{eq:linear_form_concept}
\end{equation}
where represent equivalent discrete-time matrices around operating conditions.
In a practical five-stand cold rolling mill, accurately deriving and identifying these matrices and disturbance models from first principles is difficult,
due to strong coupling, unmodeled nonlinearities, and time-varying operating regimes.
Therefore, this paper aims to learn a high-fidelity approximation of the interval evolution from data and then embed it into distributed MPC.
\begin{remark}
In fact, due to the existence of complex coupling relationships, it is difficult to directly and accurately establish \eqref{eq:linear_form_concept}
based on first principles. Therefore, in this paper, we learn an approximate mapping of \eqref{eq:true_mapping_clean} from data.
\end{remark}
Although decisions are updated at discrete instants , the hydraulic gap and drive systems evolve continuously inside each interval,
and abrupt within-interval changes may excite tension oscillations and deteriorate thickness stability.
Thus, parameterizing the within-interval increment trajectory by a low-order polynomial:
(i) yields a compact finite-dimensional decision representation;
(ii) enforces smooth profiles inside the interval;
(iii) enables enforcing increment constraints for all .
This is appropriate when is not excessively large relative to actuator bandwidth and the within-interval evolution is well approximated by a low-order basis.
\sum_{m=0}^{n}\Gamma_{i,nm}\tau^{m},
\qquad \tau\in[0,\delta_n],
\label{eq:du_poly_vec_clean}
\end{equation}
where are coefficient vectors (), and is the chosen polynomial order.
In simulations, one may select (quadratic) for a good accuracy--complexity trade-off.
Component-wise, \eqref{eq:du_poly_vec_clean} corresponds to
\begin{equation}
\begin{aligned}
\Delta s_{i,n}(\tau) &= \sum_{m=0}^{n}\gamma^{(s)}{i,nm}\tau^{m},\
\Delta v{i,n}(\tau) &= \sum_{m=0}^{n}\gamma^{(v)}{i,nm}\tau^{m}.
\end{aligned}
\label{eq:du_components_clean}
\end{equation}
Define the stacked parameter vector
\begin{equation}
\Gamma{i,n}\triangleq
\big[
(\Gamma_{i,n0})^\top,,
(\Gamma_{i,n1})^\top,,
\ldots,,
(\Gamma_{i,nn})^\top
\big]^\top
\in\mathbb{R}^{p},
\qquad
p=(n+1)n_u.
\label{eq:Gamma_clean}
\end{equation}
Here, is the baseline increment at , while () describe the higher-order variation rates.
Define the interval-averaged equivalent increments as
\begin{equation}
\begin{aligned}
\Delta u_i(t_n) &\triangleq \frac{1}{\delta_n}\int_0^{\delta_n}\Delta u_{i,n}(\tau),d\tau,\
\Delta d_i(t_n) &\triangleq \frac{1}{\delta_n}\int_0^{\delta_n}\Delta d_i(\tau),d\tau.
\end{aligned}
\label{eq:avg_def_clean}
\end{equation}
With \eqref{eq:du_poly_vec_clean}, the input average has a closed form:
\begin{equation}
\Delta u_i(t_n)=
\sum_{m=0}^{n}\Gamma_{i,nm}\frac{\delta_n^{m}}{m+1}.
\label{eq:avg_closed_clean}
\end{equation}
Let denote the sampling domain of deviation states and neighbor stacks ,
and let denote the sampling domain of polynomial parameters .
These domains specify the operating envelope used to generate supervised training data.
Given the above parameterization, one training sample is generated on each interval .
In addition to the local deviation state, the neighbor deviation states are included to represent inter-stand coupling.
The process is summarized in Table~\ref{tab:interval_sample_generation_en}.
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.15}
\caption{Procedure for generating one interval-level sample on (five-stand coupled mill).}
\label{tab:interval_sample_generation_en}
\begin{tabularx}{\linewidth}{>{\centering\arraybackslash}p{0.09\linewidth} X}
\toprule
\textbf{Step} & \textbf{Operation} \
\midrule
1 & \textbf{State sampling:} sample and from . \
2 & \textbf{Parameter sampling:} draw (coefficients for both and ). \
3 & \textbf{Control construction:} compute via \eqref{eq:du_poly_vec_clean}. \
4 & \textbf{State propagation:} integrate the \emph{five-stand coupled} mill model on (e.g., RK4) using the within-interval control trajectory, and record . \
\bottomrule
\end{tabularx}
\end{table}
Accordingly, an interval sample for subsystem can be represented as
\begin{equation}
\mathcal{D}{i,n}=\big{\Delta x_i(t_n),\ \Delta x{Z_i}(t_n),\ \Delta u_{i,n}(\tau),\ \Delta x_i(t_{n+1})\big}.
\label{eq:interval_sample_clean}
\end{equation}
Note that is fully determined by via \eqref{eq:du_poly_vec_clean},
therefore it is sufficient to store as the learning input.
For each subsystem , by repeating the above procedure across multiple intervals and randomized draws,
the local one-step training dataset is formed as
\begin{equation}
\begin{split}
S_i=\Big{&
\big(\Delta x_i^{(j)}(t_n),,\Delta x_{Z_i}^{(j)}(t_n),,\Delta x_i^{(j)}(t_{n+1});,
\Gamma_{i,n}^{(j)},,\delta_n^{(j)}\big)
\ \Big|\ j=1,\ldots,J
\Big}.
\end{split}
\label{eq:S_i_clean}
\end{equation}
Here is the number of one-step samples for subsystem .
The overall dataset for the five-stand mill is denoted by .
The point-cloud visualization of the training dataset is shown in Figure~\ref{2}.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=0.5]{picture/Fig2.pdf}
\caption{Point cloud map of the training dataset.}\label{2}
\end{figure*}
The one-step set is sufficient for one-step regression, but it is not sufficient for training with multi-step rollout loss
and reciprocal-consistency regularization, because these objectives require ground-truth deviation-state trajectories over a horizon of consecutive intervals.
Therefore, without changing the single-interval sampling mechanism above, we additionally organize the offline-simulated samples
into -step trajectory segments.
Specifically, for each starting time we generate a segment of length by consecutively sampling
(and the corresponding inputs/disturbances),
and integrating the five-stand coupled mill model over for .
Hence, we obtain the deviation-state sequence as well as the neighbor stacks
.
Define a -step segment sample for subsystem as
\begin{equation}
\begin{aligned}
\mathcal{W}{i,n}=
\Big{&
\big(\Delta x_i(t{n+s}),,\Delta x_{Z_i}(t_{n+s}),,\Gamma_{i,n+s},,\delta_{n+s}\big){s=0}^{K-1}; \
&\big(\Delta x_i(t{n+s+1})\big){s=0}^{K-1}
\Big}.
\end{aligned}
\label{eq:segment_clean}
\end{equation}
By repeating the above segment generation, we form the multi-step training set
\begin{equation}
S_i^{(K)}=\Big{\mathcal{W}{i,n}^{(j)}\ \Big|\ j=1,\ldots,J_K\Big},
\label{eq:S_i_K_clean}
\end{equation}
where is the number of -step segment samples.
Note that can be viewed as the marginal one-step projection of (keeping only ),
thus the original dataset design is preserved, and only an additional \emph{segment organization} is introduced for multi-step training.
%========================
\section{Construction of Residual Neural Network}
%========================
\subsection{Residual Neural Network Structure Construction and Training Method}
Given the dataset, the neural network model is trained to learn a stand-wise, control-dependent one-step evolution law of deviation states:
\begin{equation}
\Delta x_i(t_{n+1})
\approx
\Delta x_i(t_n)+
\mathcal{N}i!\Big(\Delta x_i(t_n),,\Delta x{Z_i}(t_n),,\Gamma_{i,n},,\delta_n;,\Theta_i\Big),
\label{eq:learned_dyn_clean}
\end{equation}
where outputs the one-step deviation-state change and are trainable parameters.
\begin{remark}
If does not take control information as input (here and ),
the predictor becomes an autoregressive model that only reproduces trajectories under the training input patterns
and cannot answer the counterfactual question: ``what will happen if we choose a different roll gap and speed trajectory?''
Since MPC optimizes over candidate decisions, a control-dependent predictor \eqref{eq:learned_dyn_clean} is necessary
to evaluate the predicted thickness and tension behavior under different candidate actuator trajectories.
\end{remark}
Let (state dimension), be the number of neighbors of stand in \eqref{eq:Zi_clean}, and in \eqref{eq:Gamma_clean}.
Define the input vector
\begin{equation}
X_{i,\text{in}} \triangleq
\big[
\Delta x_i(t_n)^\top,,
\Delta x_{Z_i}(t_n)^\top,,
\Gamma_{i,n}^\top,,
\delta_n
\big]^\top
\in \mathbb{R}^{d(1+|Z_i|)+p+1}.
\label{eq:X_in_clean}
\end{equation}
The network mapping is
\begin{equation}
\mathcal{N}_i:\mathbb{R}^{d(1+|Z_i|)+p+1}\rightarrow\mathbb{R}^{d}.
\end{equation}
To improve training stability and long-horizon rollout robustness, we use a residual form.
Let be a selection matrix extracting the local state block:
\begin{equation}
\hat{I}i = [I_d,, 0{d\times(d|Z_i|+p+1)}].
\label{eq:Ihat_clean}
\end{equation}
Then the one-step predictor is written as
\begin{equation}
X_{i,\text{out}} = \hat{I}i X{i,\text{in}} + \mathcal{N}i(X{i,\text{in}}; \Theta_i),
\label{eq:res_predict_clean}
\end{equation}
where represents the predicted .
This structure implements a baseline-plus-correction interpretation:
the shortcut propagates the current deviation state , while the network learns the correction capturing
unmodeled nonlinearities and inter-stand coupling (via ) under varying operating conditions.
To improve robustness when varies, we introduce an auxiliary branch inside :
\begin{equation}
\mathcal{N}i(X{i,\text{in}};\Theta_i)\triangleq
\psi_i(X_{i,\text{in}};\Theta_{\psi_i}) + \rho_i(X_{i,\text{in}};\theta_i),
\label{eq:aux_clean}
\end{equation}
where is a lightweight feedforward branch that captures low-frequency/scale effects strongly related to ,
and captures the remaining nonlinear coupling corrections.
When , the model reduces to a standard residual network.
For the -th sample in \eqref{eq:S_i_clean}, define
\begin{equation}
X_{i,\text{in}}^{(j)} =
\big[
\Delta x_i^{(j)}(t_n),\ \Delta x_{Z_i}^{(j)}(t_n),\
\Gamma_{i,n}^{(j)},\ \delta_n^{(j)}
\big]^{\top},
\end{equation}
and the supervised residual target
\begin{equation}
\Delta r_i^{(j)}=\Delta x_i^{(j)}(t_{n+1})-\Delta x_i^{(j)}(t_n).
\label{eq:target_clean}
\end{equation}
To suppress accumulation drift induced by long-horizon recursion and to improve long-term predictive stability,
we train the forward predictor jointly with an auxiliary backward residual model
and impose a multi-step reciprocal-consistency regularization over a -step segment from .
Construct a backward residual network
\begin{equation}
\mathcal{B}i:\mathbb{R}^{d(1+|Z_i|)+p+1}\rightarrow\mathbb{R}^{d},
\end{equation}
parameterized by . For the backward step associated with interval , define
\begin{equation}
\begin{aligned}
X{i,\mathrm{in}}^{b}
&=
\big[
\Delta x_i(t_{n+1}),\ \Delta x_{Z_i}(t_{n+1}),
\Gamma_{i,n},\ \delta_n
\big]^{\top},\
X_{i,\mathrm{out}}^{b}
&=
\hat{I}i X{i,\mathrm{in}}^{b} + \mathcal{B}i(X{i,\mathrm{in}}^{b};\bar{\Theta}i),
\end{aligned}
\label{eq:back_clean}
\end{equation}
where represents the backward estimate of .
The supervised backward residual target is
\begin{equation}
\Delta r_i^{b}=\Delta x_i(t_n)-\Delta x_i(t{n+1}).
\end{equation}
Given a segment sample , initialize
\begin{equation}
\Delta \hat{x}i(t_n)=\Delta x_i(t_n),
\end{equation}
and recursively apply the forward predictor for steps:
\begin{equation}
\begin{aligned}
\Delta \hat{x}i(t{n+s+1})
&=
\Delta \hat{x}i(t{n+s})
+
\mathcal{N}i!\Big(
\Delta \hat{x}i(t{n+s}),,\Delta \hat{x}{Z_i}(t{n+s}),,
\Gamma_{i,n+s},,\delta_{n+s};,\Theta_i
\Big),\
&\qquad s=0,\ldots,K-1.
\end{aligned}
\label{eq:fwd_roll_clean}
\end{equation}
Set the terminal condition
\begin{equation}
\Delta \bar{x}i(t{n+K})=\Delta \hat{x}i(t{n+K}),
\end{equation}
and roll back using :
\begin{equation}
\begin{aligned}
\Delta \bar{x}i(t{n+s})
&=
\hat{I}i X{i,\mathrm{in}}^{b}(t_{n+s})
+
\mathcal{B}i!\Big(X{i,\mathrm{in}}^{b}(t_{n+s});,\bar{\Theta}i\Big),
\quad s=K-1,\ldots,0,
\end{aligned}
\label{eq:bwd_roll_clean}
\end{equation}
where
\begin{equation}
X{i,\mathrm{in}}^{b}(t_{n+s})=
\big[
\Delta \bar{x}i(t{n+s+1}),\ \Delta \hat{x}{Z_i}(t{n+s+1}),
\Gamma_{i,n+s},\ \delta_{n+s}
\big]^{\top}.
\end{equation}
\sum_{s=0}^{K}
\left|
\Delta \hat{x}i(t{n+s})-\Delta \bar{x}i(t{n+s})
\right|^2.
\end{equation}
We jointly minimize:
\begin{equation}
\begin{aligned}
L_{\mathrm{1step}}(\Theta_i)
&= \frac{1}{J_K}\sum_{j=1}^{J_K}\frac{1}{K}\sum_{s=0}^{K-1}
\Big|
\big(\Delta x_i^{(j)}(t_{n+s+1})-\Delta x_i^{(j)}(t_{n+s})\big)
-\mathcal{N}i!\left(
X{i,\mathrm{in}}^{(j)}(t_{n+s});\Theta_i
\right)
\Big|^2,\[2mm]
L_{\mathrm{bwd}}(\bar{\Theta}i)
&= \frac{1}{J_K}\sum{j=1}^{J_K}\frac{1}{K}\sum_{s=0}^{K-1}
\Big|
\big(\Delta x_i^{(j)}(t_{n+s})-\Delta x_i^{(j)}(t_{n+s+1})\big)
-\mathcal{B}i!\left(
X{i,\mathrm{in}}^{b,(j)}(t_{n+s});\bar{\Theta}i
\right)
\Big|^2,\[2mm]
L{\mathrm{msrp}}(\Theta_i,\bar{\Theta}i)
&= \frac{1}{J_K}\sum{j=1}^{J_K} E_i^{(j)}(t_n),\[2mm]
L_{\mathrm{roll}}(\Theta_i)
&= \frac{1}{J_K}\sum_{j=1}^{J_K}\sum_{s=1}^{K}
\Big|
\Delta x_i^{(j)}(t_{n+s})-\Delta \hat{x}i^{(j)}(t{n+s})
\Big|^2.
\end{aligned}
\label{eq:loss_clean}
\end{equation}
Here, enforces one-step accuracy; explicitly suppresses long-horizon drift under recursion;
regularizes the learned dynamics by enforcing reciprocal consistency between forward and backward rollouts;
and trains the backward model for the consistency regularization.
In implementation, these terms are combined as
\begin{equation}
L_{\mathrm{total}}=\lambda_1 L_{\mathrm{1step}}+\lambda_2 L_{\mathrm{roll}}+\lambda_3 L_{\mathrm{msrp}}+\lambda_4 L_{\mathrm{bwd}},
\end{equation}
where are tuned on a validation set.
\Delta x_i(t_n)
+
\mathcal{N}i!\Big(
\Delta x_i(t_n),,\Delta x{Z_i}(t_n),,
\Gamma_{i,n},,\delta_n;,\Theta_i^*
\Big),
\label{eq:pred_clean}
\end{equation}
and multi-step prediction is obtained by recursive rollout of \eqref{eq:pred_clean}.
This learned predictor is the internal model used by the MPC optimizer in the next section.
Finally, network parameters are optimized using Adam:
\begin{equation}
\Theta_{i,t+1} = \Theta_{i,t} - \alpha \frac{\hat{g}{i,t}}{\sqrt{\hat{r}{i,t}} + \varepsilon},
\end{equation}
where is the learning rate (we use to avoid conflict with other symbols),
and are bias-corrected moment estimates, and is a small constant for numerical stability.
Figure~\ref{fig:rnn_logic} illustrates the overall structure.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.85]{picture/x6.pdf}
\caption{Logic diagram of the residual neural network.}
\label{fig:rnn_logic}
\end{figure}
\subsection{Explainability of the residual network}
\Delta x_i(t_n)
+
\underbrace{\int_{0}^{\delta_n}
f_i!\Big(\Delta x_i(t_n+\tau),,\Delta x_{Z_i}(t_n+\tau),,u_i(t_n+\tau),,d_i(t_n+\tau)\Big),d\tau}{\triangleq\ \varphi{i,n}},
\label{eq:increment_integral}
\end{equation}
where is the one-interval state increment generated by the mechanistic dynamics.
Our learned model \eqref{eq:learned_dyn_clean} adopts the same increment form as \eqref{eq:increment_integral}:
\begin{equation}
\Delta x_i(t_{n+1})
\approx
\Delta x_i(t_n)
+
\mathcal{N}i!\Big(\Delta x_i(t_n),,\Delta x{Z_i}(t_n),,\Gamma_{i,n},,\delta_n;\Theta_i\Big).
\end{equation}
Here, plays the role of a data-driven approximation of the integral increment ,
i.e., it approximates the accumulated effect of the mechanistic dynamics over .
This is consistent with the well-known interpretation that a residual network behaves like a one-step time integrator:
the identity path propagates the current state, while the residual branch represents the increment over the time lag.
Inside each interval, we do not optimize point-wise but parameterize the increment trajectory by
.
Hence, the mechanistic increment in \eqref{eq:increment_integral} depends on the \emph{whole} within-interval trajectory.
Feeding into is therefore a compact way to represent how different candidate
gap/speed trajectories change the integral effect and thus the next thickness--tension state.
When varies and is not very small, directly learning may be harder.
Motivated by the generalized residual idea,
we decompose the increment predictor into two parts in \eqref{eq:aux_clean}:
\begin{equation}
\mathcal{N}_i(\cdot)=\psi_i(\cdot)+\rho_i(\cdot).
\end{equation}
Conceptually, captures low-frequency and scale effects strongly related to ,
while captures the remaining nonlinear coupling corrections.
This provides a mechanism-consistent interpretation: a baseline increment plus a residual correction
that compensates unmodeled nonlinearities and inter-stand coupling.
%========================
\section{Nash Equilibrium-Based RNE-DMPC}
%========================
The five-stand tandem cold rolling system is strongly coupled through inter-stand tension propagation.
As a result, changes in control actions (roll gap and stand speed) at one stand can affect both upstream and downstream stands,
making centralized online optimization over all stands' decision variables computationally demanding.
To mitigate this issue, we decompose the global predictive-control problem into local subproblems associated with individual stands.
Each local controller optimizes its own decision variables while accounting for coupling via limited information exchange with neighboring controllers.
Motivated by game-theoretic coordination \citep{rawlings2008coordinating}, we formulate distributed coordination as a Nash-equilibrium-seeking iteration.
Based on the trained residual neural network surrogate model, we construct a Nash-equilibrium-based distributed MPC method (RNE-DMPC)
for coordinated thickness--tension regulation and tracking. The overall control structure is shown in Figure~\ref{4}.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/x2.pdf}
\caption{Schematic diagram of the control architecture for a tandem cold rolling mill.}\label{4}
\end{figure*}
At sampling time , stand chooses the polynomial-parameter sequence
, where .
Let
denote the joint strategy profile, and let denote the collection of all strategies except stand .
Given the current measured/estimated deviation state and the strategies
,
the multi-step prediction used by stand is written explicitly as
\begin{equation}
\begin{aligned}
\Delta \hat{x}i(t{n+s+1};\mathbf{\Gamma}i,\mathbf{\Gamma}{Z_i})
&=
\Delta \hat{x}i(t{n+s};\mathbf{\Gamma}i,\mathbf{\Gamma}{Z_i})
+
\mathcal{N}i!\Big(
\Delta \hat{x}i(t{n+s};\cdot),,
\Delta \hat{x}{Z_i}(t_{n+s};\mathbf{\Gamma}{Z_i}),\
&\qquad
\Gamma{i,n+s},,
\delta_{n+s};\Theta_i^*
\Big),
\end{aligned}
\label{eq:rollout_mpc_game}
\end{equation}
for , with initialization .
Here the neighbor stack is generated from neighbors' strategies via the same learned predictors.
\sum_{m=0}^{n}\Gamma_{i,n+s,m}\frac{\delta_{n+s}^{m}}{m+1}.
\label{eq:du_avg_clean}
\end{equation}
\begin{remark}
Because inter-stand tension is jointly affected by the adjacent stands and ,
the predicted evolution of depends on neighbors' future actions,
hence the MPC problems are not independent but form a coupled dynamic game.
\end{remark}
\mathrm{col}{\Gamma_{i,n},\Gamma_{i,n+1},\ldots,\Gamma_{i,n+N_c-1}}
\in \mathbb{R}^{pN_c}.
\end{equation}
In deviation coordinates, the regulation/tracking objective is , i.e.
\begin{equation}
\Delta x_{i,\mathrm{ref}}(t_{n+s})\equiv 0\in\mathbb{R}^{d},\qquad d=3.
\end{equation}
Recall .
Define the row selectors
\begin{equation}
C^- \triangleq [0\ \ 1\ \ 0]\in\mathbb{R}^{1\times 3},\qquad
C^+ \triangleq [0\ \ 0\ \ 1]\in\mathbb{R}^{1\times 3},
\end{equation}
so that (upstream interface) and (downstream interface).
C^-\Delta \hat{x}{i+1}(t{n+s};\mathbf{\Gamma}),\qquad i=1,\ldots,4.
\label{eq:shared_tension_mismatch}
\end{equation}
\sum_{s=1}^{N_p}
\left|
\Delta \hat{x}i(t{n+s};\mathbf{\Gamma}i,\mathbf{\Gamma}{Z_i})
\right|{Q_i}^{2}
+
\sum{s=0}^{N_c-1}
\left|\Gamma_{i,n+s}\right|_{R_i}^{2}
+
J_i^{\mathrm{cpl}}(\mathbf{\Gamma}i;\mathbf{\Gamma}{-i})
\label{eq:Ji_game}
\end{equation}
where weights thickness and tension deviations, and penalizes actuation magnitudes.
C^-\Delta \hat{x}{i+1}(t{n+s};\mathbf{\Gamma})
\big|^2
\Big),
\end{aligned}
\label{eq:coupling_cost}
\end{equation}
with and indicator .
This term makes the coupling conflict explicit: unilateral actions that locally reduce thickness error may worsen shared-tension
compatibility and thus increase , and also affect neighbors' objectives.
We enforce the absolute-input bounds and within-interval increment bounds.
Absolute input bounds (roll gap and speed):
\begin{equation}
u_{i,\min}\le u_i(t_{n+s})\le u_{i,\max},
\qquad s=0,\ldots,N_p-1,
\label{eq:u_abs_game}
\end{equation}
where .
Within-interval increment-trajectory bounds:
\begin{equation}
\Delta u_{i,\min}\le \Delta u_{i,n+s}(\tau;\Gamma_{i,n+s})\le \Delta u_{i,\max},
\qquad \forall\tau\in[0,\delta_{n+s}],\ s=0,\ldots,N_p-1,
\label{eq:du_traj_game}
\end{equation}
where is given by the polynomial parameterization.
\sum_{m=0}^{n}\Gamma_{i,n+s,m}\frac{\delta_{n+s}^{m}}{m+1},
\label{eq:du_avg_game}
\end{equation}
and propagate the absolute input along the horizon by
\begin{equation}
u_i(t_n)=u_i(t_{n-1})+\Delta u_i(t_n),\qquad
u_i(t_{n+s})=u_i(t_{n+s-1})+\Delta u_i(t_{n+s}),\ s=1,\ldots,N_p-1,
\label{eq:u_prop_game}
\end{equation}
where is the applied (measured) input from the previous sampling instant.
Compact feasible set:
\begin{equation}
\Omega_i \triangleq
\Big{\mathbf{\Gamma}_i\ \Big|\
\eqref{eq:rollout_mpc_game}\ \text{holds and}
\eqref{eq:u_abs_game},\eqref{eq:du_traj_game},\eqref{eq:u_prop_game}\ \text{are satisfied}
\Big}.
\label{eq:Omega_i_game}
\end{equation}
\arg\min_{\mathbf{\Gamma}_i\in\Omega_i}\
J_i(\mathbf{\Gamma}i;\mathbf{\Gamma}{-i}).
\label{eq:local_BR}
\end{equation}
Because the learned surrogate is differentiable, \eqref{eq:local_BR} can be solved by standard gradient-based NLP solvers.
At each sampling time , the distributed MPC coordination induces a finite-horizon dynamic game:
players are stands ; strategy sets are ; and payoff (cost) functions are
defined in \eqref{eq:Ji_game}--\eqref{eq:coupling_cost}.
A joint strategy profile
is a Nash equilibrium if
\begin{equation}
\forall i\in{1,\ldots,5},\qquad
\mathbf{\Gamma}i^*\in
\arg\min{\mathbf{\Gamma}_i\in\Omega_i}
J_i(\mathbf{\Gamma}i;\mathbf{\Gamma}{-i}^*).
\label{eq:NE_def}
\end{equation}
This definition explicitly characterizes the strategic coupling:
each player's optimal decision depends on neighbors' decisions through the shared-tension dynamics and the coupling term.
(1-\omega)\mathbf{\Gamma}_i^{(l-1)}
+
\omega,\mathbf{\Gamma}_i^{\mathrm{BR},(l)},
\qquad \omega\in(0,1].
\label{eq:relaxed_BR}
\end{equation}
The relaxation factor mitigates oscillations caused by strong coupling and improves practical convergence.
The Nash equilibrium is computed through distributed best-response iterations, summarized in Table ~\ref{tab:nash_iter_en}.
\max_i
\frac{\left|
\mathbf{\Gamma}_i^{(l)}-\mathbf{\Gamma}_i^{(l-1)}
\right|_2}{
\left|
\mathbf{\Gamma}_i^{(l-1)}
\right|_2+\epsilon},
\end{equation}
with small.
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.12}
\setlength{\tabcolsep}{3.5pt}
\caption{Relaxed distributed Nash best-response iteration for RNE-DMPC (five-stand).}
\label{tab:nash_iter_en}
\begin{tabularx}{\linewidth}{>{\centering\arraybackslash}p{0.11\linewidth} X}
\toprule
\textbf{Step} & \textbf{Description} \
\midrule
A &
Initialize and warm-start (e.g., from the previous sampling time). \
B &
Communicate (or the induced predicted trajectories) among neighbors; form . \
C &
Given , solve the best-response NLP \eqref{eq:local_BR} to obtain . \
D &
Update the relaxed strategy using \eqref{eq:relaxed_BR} and compute the induced predictions
via \eqref{eq:rollout_mpc_game}. \
E &
Broadcast and the predicted interface quantities needed in \eqref{eq:coupling_cost}
(e.g., ) to neighbors. \
F &
Compute ; if , stop and set ; otherwise and repeat. \
\bottomrule
\end{tabularx}
\end{table}
\sum_{m=0}^{n}\Gamma_{i,n,m}^*\frac{\delta_n^{m}}{m+1},
\end{equation}
followed by .
The overall control flow chart is shown in Fig~\ref{liu}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{picture/x5.pdf}
\caption{The overall system control flow chart}\label{liu}
\end{figure}
\begin{remark}
Equations \eqref{eq:Ji_game}--\eqref{eq:coupling_cost} and the NE definition \eqref{eq:NE_def} explicitly show that each player's objective
depends on others' strategies through the coupled prediction dynamics and the shared-tension interface mismatch penalty,
thus making the multi-stand strategic interaction and coordination mechanism transparent.
\end{remark}
\section{Simulations}
To verify the effectiveness of the proposed residual-neural-network and Nash-equilibrium based distributed MPC (RNE-DMPC) for coordinated thickness--tension regulation in a five-stand cold tandem rolling system, a distributed control architecture composed of multiple local controllers and a communication/coordination module is considered. Each stand is treated as a subsystem, and information exchange through the communication system enforces coordination among subsystems, enabling distributed thickness--tension control for the interconnected plant.
In this study, the simulation data were generated using the mathematical model reported by Hu et al.~\citep{hu2021nash}.
\sqrt{\frac{1}{N}\sum_{i=1}^{N}\left(x_i-\hat{x}_i\right)^2 }.
\end{equation}
where is the number of samples, and are the actual and predicted values, respectively.
\subsection{Surrogate-model prediction comparison}
In this paper, all of the spatial dynamics network and temporal evolution networks are constructed
based on the ResNet architecture and trained using the loss function in \eqref{eq:loss_clean}. Specifically, one-block ResNet method is chosen, with 5 hidden layers in each block, and each layer contains 160 neurons. And the optimizer is chosen as Adam with betas equal (0.9,0.99), initial learning rate is 0.001 and will decay to 0.00001 gradually. All networks are trained in the PyTorch
environment for 500 epochs. Meanwhile, the multi-order parameterization in the simulation part uses the second-order approach.
To justify the use of the proposed residual predictor in MPC, we first compare its prediction accuracy against representative baseline models, including
a plain feedforward network (MLP) with comparable model size trained by one-step regression\citep{Hornik1989MLPUniversal}, and a sequence model (LSTM) trained for trajectory prediction\citep{Hochreiter1997LSTM}.
All models are trained and tested under the same dataset split.
Figure~\ref{fig:pred_compare} shows an illustrative trajectory overlay on a representative test case.
It can be observed that the proposed residual network produces a closer match to the ground truth and exhibits less drift in multi-step rollouts.
The aggregated results can be summarized in Table~\ref{tab:pred_rmse_states}, where the proposed residual network achieves the lowest prediction errors among the compared models.
\begin{figure*}[htbp]
\centering
\subfloat[State ]{%
\includegraphics[width=0.32\linewidth]{picture/pred1.pdf}%
\label{F1_pred}%
}\hfill
\subfloat[State ]{%
\includegraphics[width=0.32\linewidth]{picture/pred2.pdf}%
\label{F2_pred}%
}\hfill
\subfloat[State ]{%
\includegraphics[width=0.32\linewidth]{picture/pred3.pdf}%
\label{F3_pred}%
}
\caption{Prediction trajectory overlay.}
\label{fig:pred_compare}
\end{figure*}
% in preamble:
% \usepackage{booktabs}
% \usepackage{multirow}
% \usepackage{booktabs} % 需要 booktabs
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.15}
\setlength{\tabcolsep}{8pt}
\caption{RMSE comparison of multi-step prediction for each state.}
\label{tab:pred_rmse_states}
\begin{tabular}{cccc}
\toprule
Methods & \multicolumn{3}{c}{RMSEs of each state} \
\cmidrule(lr){2-4}
& State & State & State \
\midrule
Proposed method & 0.0311 & 0.0798 & 0.1010 \
LSTM & 0.0376 & 0.0678 & 0.1312 \
MLP & 0.0418 & 0.0975 & 0.1368 \
\bottomrule
\end{tabular}
\end{table}
\subsection{Closed-loop tracking performance}
This subsection evaluates the closed-loop tracking performance of RNE-DMPC under (i) multiple sequential reference changes and (ii) single step changes in thickness and tension references.
In Figure~\ref{simulation}, three reference targets were modified sequentially during the simulation for the five-stand thickness--tension coupled system, and the resulting closed-loop responses were recorded.
It can be observed that the tension and thickness outputs of all stands were driven toward the updated references promptly, while overall stability was maintained.
At each switching instant, short-lived deviations of the tracking errors were induced due to the strong inter-stand coupling; however, these deviations were rapidly attenuated and kept at small levels afterwards.
Such behavior indicates that effective dynamic tracking and coordination were achieved under multivariable coupling and repeated setpoint switching.
Moreover, the contribution of the distributed architecture was reflected by the information exchange enabled by the communication mechanism, through which consistency across subsystems was enforced and coordinated regulation was maintained throughout the successive reference updates.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/simulation.pdf}
\caption{Tension--thickness tracking responses under multiple reference changes: (a) shows the tension tracking response; (b) shows the thickness tracking response.}\label{simulation}
\end{figure*}
In Figure~\ref{thickness}, a single step change was imposed on the thickness reference, and both the thickness trajectories and the tracking errors were evaluated.
Following the step, the thickness outputs were guided toward the new reference levels with a fast transition, and the tracking errors were suppressed quickly and maintained within a small range, implying strong transient regulation and high steady-state accuracy in the thickness channel.
In particular, after the responses reached steady state, the absolute tracking error of the outlet thickness (target value ) was , corresponding to a relative error of 2.28%.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/thickness.pdf}
\caption{Thickness responses and tracking errors under a step change of the thickness reference: (a) shows the thickness response; (b) shows the tracking error.}\label{thickness}
\end{figure*}
In Figure~\ref{tension}, a single step change was applied to the tension reference, and the corresponding tension responses and error evolutions were examined.
After the reference update, the tension outputs were pulled toward the new setpoints rapidly, and the errors were driven to small values without being accompanied by sustained oscillations, indicating satisfactory damping and closed-loop stability.
After steady state, the absolute tracking error of the tension (target value ) was , corresponding to a relative error of 1.91%.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/tension.pdf}
\caption{Tension responses and tracking errors under a step change of the tension reference: (a) shows the tension response; (b) shows the tracking error.}\label{tension}
\end{figure*}
\subsection{Impact of Nash coordination}
To highlight the role of Nash-equilibrium coordination in strongly coupled distributed control, we compare the proposed RNE-DMPC against a distributed MPC baseline without Nash coordination\citep{camponogara2002distributed}.
Specifically, the baseline performs only a single local update per sampling instant while keeping neighbors' strategies fixed from the previous update, i.e., no online best-response iterations are carried out.
Figure~\ref{fig:five_subplots1} and Figure~\ref{fig:five_subplots2} present the comparison of tracking errors for thickness and tension, where the Nash-coordinated scheme yields consistently smaller errors and faster convergence.
For a compact quantitative summary, we report the mean steady-state absolute tracking errors averaged over all stands, and the peak absolute error.The average steady-state absolute tracking error is denoted as for thickness and for tension.
The steady-state window is defined from the first time the tracking error enters and remains within the band until the end of the simulation.
Specifically, is computed by time-averaging over the steady-state window and then taking the mean across the five stands (), while is computed similarly and then averaged across the four inter-stand tensions ().
The results are summarized in Table~\ref{tab:nash_metrics}, confirming that Nash coordination improves the tracking accuracy in the coupled multi-stand system.
\begin{figure*}[htbp]
\centering
% -------- Row 1: three subfigures --------
\subfloat[Stand 1]{%
\includegraphics[width=0.32\linewidth]{picture/1.1.pdf}%
\label{fig:five_a}%
}\hfill
\subfloat[Stand 2]{%
\includegraphics[width=0.32\linewidth]{picture/1.2.pdf}%
\label{fig:five_b}%
}\hfill
\subfloat[Stand 3]{%
\includegraphics[width=0.32\linewidth]{picture/1.3.pdf}%
\label{fig:five_c}%
}\[2mm]
% -------- Row 2: two subfigures (same size as row 1, centered) --------
\makebox[\linewidth][c]{%
\subfloat[Stand 4]{%
\includegraphics[width=0.32\linewidth]{picture/1.4.pdf}%
\label{fig:five_d}%
}\hspace{0.04\linewidth}%
\subfloat[Stand 5]{%
\includegraphics[width=0.32\linewidth]{picture/1.5.pdf}%
\label{fig:five_e}%
}%
}
\caption{Thickness tracking error comparison with and without Nash coordination.}
\label{fig:five_subplots1}
\end{figure*}
\begin{figure*}[htbp]
\centering
% -------- Row 1: three subfigures --------
\subfloat[Stand 1]{%
\includegraphics[width=0.32\linewidth]{picture/2.1.pdf}%
\label{fig:five_a}%
}\hfill
\subfloat[Stand 2]{%
\includegraphics[width=0.32\linewidth]{picture/2.2.pdf}%
\label{fig:five_b}%
}\hfill
\subfloat[Stand 3]{%
\includegraphics[width=0.32\linewidth]{picture/2.3.pdf}%
\label{fig:five_c}%
}\[2mm]
% -------- Row 2: two subfigures (same size as row 1, centered) --------
\makebox[\linewidth][c]{%
\subfloat[Stand 4]{%
\includegraphics[width=0.32\linewidth]{picture/2.4.pdf}%
\label{fig:five_d}%
}\hspace{0.04\linewidth}%
\subfloat[Stand 5]{%
\includegraphics[width=0.32\linewidth]{picture/2.5.pdf}%
\label{fig:five_e}%
}%
}
\caption{Tension tracking error comparison with and without Nash coordination.}
\label{fig:five_subplots2}
\end{figure*}
% Preamble:
% \usepackage{booktabs}
% Preamble:
% \usepackage{booktabs}
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.15}
\setlength{\tabcolsep}{10pt}
\caption{Steady-state tracking accuracy comparison (mean steady-state absolute error and peak error).}
\label{tab:nash_metrics}
\begin{tabular}{lcccc}
\toprule
\textbf{Method} &
\multicolumn{2}{c}{\textbf{Thickness error}} &
\multicolumn{2}{c}{\textbf{Tension error}} \
\cmidrule(lr){2-3}\cmidrule(lr){4-5}
& & (mm)
& & (kN) \
\midrule
RNE-DMPC & 2.9% & 0.0411 & 3.5% & 0.0714 \
Distributed MPC & 3.2% & 0.0503 & 4.1% & 0.0706 \
\bottomrule
\end{tabular}
\end{table}
\subsection{Computation rate comparison}
Through thickness and tension simulations, the computation time of each control cycle was recorded.
Figure~\ref{rate} compares the computation time of the proposed RNE-DMPC with a Nash-equilibrium-based distributed predictive control baseline~\citep{hu2021nash}.
Statistical results show that the per-cycle computation time of both Nash-type distributed controllers remains small, which is suitable for online receding-horizon implementation.
More importantly, the proposed RNE-DMPC consistently yields a noticeably lower per-cycle computation time than the Nash baseline in Ref.~\citep{hu2021nash}, indicating faster equilibrium seeking and coordination in online operation.
Overall, the results indicate that RNE-DMPC offers excellent computational performance and can effectively reduce system load and implementation cost.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/rate.pdf}
\caption{Comparison of calculation rates for different control strategies: (a) step disturbance in outlet thickness; (b) step disturbance in tension.}\label{rate}
\end{figure*}
\section{Conclusion}
This paper proposes a distributed model predictive control strategy based on residual neural networks and Nash equilibrium (RNE-DMPC) to achieve coordinated and high-precision control of thickness and tension in the cold rolling process. By using residual neural networks to conduct data-driven modeling of the input-output relationship of the system, the problem of difficulty in interpretation of traditional neural networks has been successfully overcome. Further, by combining the distributed model predictive control method based on Nash equilibrium, the constraint optimization and coordinated control of multi-variable coupled systems have been achieved, thereby alleviating the problem of slow convergence speed and improving the calculation speed. Numerical simulation results show that the proposed strategy exhibits excellent comprehensive performance in thickness and tension control, and has practical significance for improving the quality and production efficiency of cold-rolled steel strips. Future work can further consider shortening the system stability time in actual noisy environments and enhancing the control robustness of the system.
\section*{Disclosure statement}
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
\section*{Funding}
An unnumbered section, e.g.\ \verb"\section*{Funding}", may be used for grant details, etc.\ if required and included \emph{in the non-anonymous version} before any Notes or References.
\section*{Notes on contributors}
\noindent
\begin{minipage}[t]{0.28\textwidth}
\vspace{0pt}
\includegraphics[width=\linewidth]{picture/guochenhao.jpg} % ←务必写清扩展名
\end{minipage}\hspace{0.03\textwidth}
\begin{minipage}[t]{0.69\textwidth}
\vspace{0pt}
\textit{Chenhao Guo} is pursuing the B.E. degree at Northeastern University, Shenyang, China. His research interests include the control of multi-agent systems, safety control, and adaptive control.
\end{minipage}
\par\medskip
\noindent
\begin{minipage}[t]{0.28\textwidth}
\vspace{0pt}
\includegraphics[width=\linewidth]{picture/niejiahao.jpg} % ←务必写清扩展名
\end{minipage}\hspace{0.03\textwidth}
\begin{minipage}[t]{0.69\textwidth}
\vspace{0pt}
\textit{Jiahao Nie} is currently pursuing the B.S. degree at the College of Information Science and Engineering, Northeastern University, Shenyang, China. His research interests include intelligent sensing and gas detection.
\end{minipage}
\par\medskip
\noindent
\begin{minipage}[t]{0.28\textwidth}
\vspace{0pt}
\includegraphics[width=\linewidth]{} % ←务必写清扩展名
\end{minipage}\hspace{0.03\textwidth}
\begin{minipage}[t]{0.69\textwidth}
\vspace{0pt}
\textit{Chenhao Guo} is pursuing a bachelor’s degree at North-eastern University,
Shenyang, China. His research interests include the control of multi-agent systems,
safety control, and adaptive control.
\end{minipage}
\par\medskip
\bibliographystyle{apacite}
\bibliography{interactapasample}
\end{document}
这是我的完整的一篇论文 。根据所有内容,帮我改一下introduction的两点贡献部分
下面给你一版**更贴合你全文内容、且更“审稿友好”**的两点贡献(直接替换你 Introduction 里 The main contributions... 的 enumerate 部分即可)。
latexThe main contributions of this article are as follows: \begin{enumerate} \item A control-dependent residual neural surrogate model is developed for the five-stand thickness--tension coupled dynamics. By parameterizing within-interval input increments using a low-order polynomial and training the network with one-step loss, multi-step rollout loss, and reciprocal-consistency regularization, the proposed predictor improves long-horizon robustness and alleviates error accumulation under recursive prediction. \item A Nash-equilibrium-based distributed MPC scheme (RNE-DMPC) is proposed to coordinate multi-stand thickness--tension regulation under strong inter-stand coupling. The coordination is formulated via explicit shared-tension consistency penalties and solved online using a relaxed best-response iteration, yielding improved tracking performance and reduced computation time compared with non-Nash distributed MPC and Nash-type baselines in simulations. \end{enumerate}
如果你希望贡献2里把“更低计算耗时”只保留为“验证了实时性/计算效率”,我也可以给你一个更保守的版本(有些期刊更偏好弱化结论措辞)。
在贡献二中加入在使用神经网络的时候使用了nash均衡
下面是把你要求的点(“在使用神经网络预测模型时引入/采用 Nash 均衡协调”)明确写进贡献二后的版本,直接替换你 Introduction 里的两点贡献即可。
latexThe main contributions of this article are as follows: \begin{enumerate} \item A control-dependent residual neural surrogate model is developed for the five-stand thickness--tension coupled dynamics. By parameterizing within-interval input increments using a low-order polynomial and training the network with one-step loss, multi-step rollout loss, and reciprocal-consistency regularization, the proposed predictor improves long-horizon robustness and alleviates error accumulation under recursive prediction. \item A Nash-equilibrium-based distributed MPC scheme integrated with the neural surrogate model (RNE-DMPC) is proposed for coordinated multi-stand thickness--tension regulation under strong inter-stand coupling. Specifically, the trained residual network is embedded as the prediction model in each local MPC, and Nash-equilibrium coordination is introduced to resolve strategy conflicts among stands via shared-tension consistency penalties, which is solved online using a relaxed best-response iteration. Simulations demonstrate improved tracking performance and reduced computation time compared with non-Nash distributed MPC and Nash-type baselines. \end{enumerate}
如果你想把最后一句“reduced computation time”说得更稳一点(避免审稿人抓“所有工况都更快吗”),可以把那句改成:Simulations demonstrate improved tracking performance with competitive computation time... 我也可以顺手帮你替换成更保守措辞。
For the multi-stand thickness-tension coordination adjustment problem under strong coupling conditions, a distributed model predictive control scheme based on Nash equilibrium (RNE-DMPC) is proposed. Specifically, the trained residual network is embedded into each local model predictive control as the prediction model, and the Nash equilibrium coordination mechanism is introduced. The strategy conflicts among each station are resolved by sharing the tension consistency penalty, and this process is solved online through the iterative best response relaxation.缩短点,太长了
可以缩成一句到两句,信息不丢:
版本1(两句,最稳):
For strongly coupled multi-stand thickness--tension coordination, we propose a Nash-equilibrium-based distributed MPC (RNE-DMPC). The trained residual network is embedded as the local predictor, and Nash coordination is achieved online via a shared-tension consistency penalty and a relaxed best-response iteration.
版本2(一句,更短):
A Nash-equilibrium-based distributed MPC (RNE-DMPC) is developed for strongly coupled multi-stand thickness--tension coordination, where a residual-network predictor is embedded locally and Nash coordination is enforced via a shared-tension penalty with relaxed best-response iterations.
你想更“学术保守”一点(少用 achieved/enforced),我也能再压一版。
你看看我的introduction和abstract还有没有什么需要改的地方,能不能写的更好
可以改得更紧、更“像论文”,主要问题集中在:语法/用词不统一、重复表达、贡献点没把你全文里最硬的技术点写出来、Abstract 没量化亮点。下面我给你一套“可直接替换”的 Abstract 和 Introduction(开头+过渡+贡献段),并列出你原文里建议修的关键点。
latex\begin{abstract} High-precision and efficient thickness--tension regulation in multi-stand tandem cold rolling is challenging due to strong inter-stand coupling, constraints, and the difficulty of maintaining accurate mechanistic prediction models under varying operating conditions. This paper proposes a residual-neural-network and Nash-equilibrium integrated distributed model predictive control scheme (RNE-DMPC) for coordinated thickness--tension tracking in a five-stand cold tandem mill. A control-dependent residual neural surrogate is trained to capture stand-wise deviation dynamics, where within-interval input increments are parameterized by a low-order polynomial and the network is trained with one-step loss, multi-step rollout loss, and reciprocal-consistency regularization to mitigate error accumulation in recursive prediction. The learned surrogate is embedded into each local MPC, and online coordination is achieved via a Nash-equilibrium seeking mechanism with shared-tension consistency penalties solved by a relaxed best-response iteration. Numerical simulations on a benchmark five-stand model demonstrate improved multi-step prediction accuracy over MLP and LSTM baselines, enhanced closed-loop tracking under reference changes and step disturbances, and competitive computational efficiency compared with Nash-type distributed MPC baselines. \end{abstract}
如果你愿意更“硬”,我可以把你表里的 RMSE 和跟踪误差数字直接塞进摘要,摘要会更有说服力。
latex\section{Introduction} Cold-rolled strip steel is a crucial intermediate product in the strip-steel manufacturing chain and significantly affects downstream process performance and final product quality \citep{mancini2012effects}. It is commonly produced by tandem cold rolling mills, where the strip passes sequentially through multiple stands and is progressively reduced in thickness by rolling forces. In such mills, regulating the exit thickness and inter-stand tensions is essential for product quality and process stability. Figure~\ref{AutoReg} shows a typical layout of a five-stand tandem cold rolling mill. From a control perspective, thickness--tension regulation is a large-scale multivariable problem featuring strong coupling, nonlinearities, constraints, time-varying parameters, and disturbances. These characteristics complicate reliable prediction of rolling dynamics and make high-performance closed-loop control difficult to sustain over changing operating conditions.
latexPID control remains prevalent in practice due to its simplicity, yet fixed-parameter tuning often struggles in strongly coupled multivariable rolling processes \citep{gupta2025grid,liu2025data,hasilci2025superior}. Model predictive control (MPC) provides a systematic framework to handle multivariable coupling and constraints and has been widely investigated in rolling applications \citep{yaghini2024linear,luis2019trajectory,han2021hierarchical,song2025model,sun2017dynamic}. Depending on the system scale and information interaction, MPC implementations can be centralized, decentralized, or distributed \citep{song2025model}. In particular, distributed MPC decomposes the large-scale optimization into local problems solved in parallel with limited information exchange, improving scalability and reducing computational burden for interconnected systems \citep{scattolini2009architectures}.
latexHowever, most MPC implementations still rely on mechanistic prediction models, which typically require extensive parameter identification and calibration. In tandem cold rolling, variations in steel grade, friction, speed, and thermal conditions can induce significant model mismatch, which may degrade prediction accuracy and consequently undermine closed-loop performance. With increasing availability of industrial data, data-driven modeling has gained attention in rolling processes \citep{zong2025machine}. Neural-network predictors identified from input--output data have been explored to improve prediction and facilitate predictive control \citep{xu2024novel,cheng2015neural,xu2022control,han2019data,feng2025model,niu2025rbf,huang2022lstm,xia2023using}. Nevertheless, purely black-box predictors may suffer from long-horizon error accumulation under recursive rollout, and their reliability can deteriorate when embedded in online optimization for strongly coupled multi-stand systems.
latexIn multi-stand rolling, inter-stand tension propagation introduces strong coupling among local control objectives: actions at one stand affect neighboring tensions and thickness responses, leading to potential strategy conflicts among local MPC controllers. Game-theoretic coordination provides a principled way to resolve such conflicts. Nash-equilibrium-based distributed predictive control has been studied for thickness--tension coordination in rolling mills \citep{hu2021nash,ye2021adaptive}. However, the effectiveness and convergence of online equilibrium seeking can be sensitive to prediction accuracy, especially when operating conditions vary and modeling errors increase.
把你原来的两点贡献替换成下面这两点(更贴合你全文):
latexThe main contributions of this article are as follows: \begin{enumerate} \item A control-dependent residual neural surrogate model is developed for five-stand thickness--tension deviation dynamics. By parameterizing within-interval input increments using a low-order polynomial and training the network with one-step loss, multi-step rollout loss, and reciprocal-consistency regularization, the proposed predictor improves long-horizon robustness and alleviates error accumulation in recursive prediction. \item A Nash-equilibrium-based distributed MPC scheme integrated with the neural surrogate model (RNE-DMPC) is proposed for strongly coupled multi-stand thickness--tension coordination. The trained residual network is embedded as the local prediction model, and Nash coordination is introduced via shared-tension consistency penalties solved online using a relaxed best-response iteration, leading to improved coordination and tracking performance in simulations. \end{enumerate}
你原摘要中这些地方建议直接替换:
This coupling relationships have led to → These coupling relationships lead tomultiple racks → multiple standsa series numerical simulations → a series of numerical simulationsNash equilibrium(RNE-DMPC) → Nash equilibrium (RNE-DMPC)residual nerual network → residual neural network如果你把期刊对 Abstract 字数要求(比如 150/200/250 words)告诉我,我可以把上面的 Abstract 再压缩/扩展到刚好卡线;另外如果你愿意放数字进摘要,我也能给你一版“带关键数值”的更强摘要。
术语统一
全文用 stand,不要混用 rack(你摘要里写了 racks)。
“Residual neural network / ResNet / residual network” 建议统一成 residual neural network (ResNet-style),第一次出现解释一次即可。
“Nash equilibrium(RNE-DMPC)” 中间要空格:Nash equilibrium (RNE-DMPC)。
少用过强的“可解释性”表述 你正文确实给了“残差=积分增量”的解释,但不建议写“successfully overcome interpretability”,更稳的写法是:
offers a mechanism-consistent increment interpretation / improves interpretability via an increment form。
Abstract 需要数字亮点 你现在的摘要最后一句太泛。建议至少放 2 组数字(你文中已有):
跟踪:0.029 mm、0.021 kN(或稳态百分比 2.28%、1.91%)
预测:你表里 ResNet vs LSTM/MLP 的 RMSE(可给一组代表性 state)
这些问题都在那里出现,帮我指出来
下面把你提到的 4 类问题,逐条指出在你这份稿子里具体出现在哪里(原句片段),并给出建议替换写法(保持你原意不变)。
rack / stand 混用(需要统一成 stand)原句:
... have led to the problem of optimization conflicts among multiple racks.racks 改为 stands:... among multiple stands.原句:
... optimization conflicts among multiple racks.... among multiple stands.原句:
Section 5 presents the simulation results of the thickness and tension of the five-rack system.... of the five-stand system.原句:
The strategy conflicts among each station are resolved ...station 改为 stand:... among stands are resolved ...你现在三种写法都出现了,建议统一为 residual neural network (ResNet-style),并在第一次出现时解释一次。
原句:
... based on residual neural network and Nash equilibrium(RNE-DMPC). A ResNet is trained ...建议(统一写法示例):
a residual neural network (ResNet-style)the residual neural network(不要再切换成 ResNet)原句:
... integrates the residual neural network into MPC for prediction. This data-based deep residual network ...residual neural network 和 deep residual network。建议:
residual neural network (ResNet-style)(第一次)/ residual neural network(后续)原句:
... Nash equilibrium(RNE-DMPC).... Nash equilibrium (RNE-DMPC).(同类的括号前后空格问题,你正文里基本写得不错,主要是摘要这句。)
你提到的“不要写得太满”,这句不是在摘要里,而是在 Conclusion(结论)。
原句:
... the problem of difficulty in interpretation of traditional neural networks has been successfully overcome.建议改成更稳的两种之一:
... which offers a mechanism-consistent increment interpretation.... which improves interpretability via an increment-form modeling structure.这两种都更不容易被审稿人抓“你怎么证明完全可解释”。
原句:
The performance of the proposed method is verified by a series numerical simulations.问题是:没有任何量化结果。
你说要放两组数字,你文中现成能用的就在仿真小节里:
0.029 mm (2.28%),张力稳态绝对误差 0.021 kN (1.91%)0.0311,LSTM 0.0376,MLP 0.0418(或你选最能体现优势的 state)建议你把摘要最后一句换成类似这种结构(只示例,不强迫你用这句):
Simulations show steady-state tracking errors of 0.029 mm (2.28%) in thickness and 0.021 kN (1.91%) in tension, and lower multi-step prediction RMSEs than MLP and LSTM baselines.如果你愿意,我也可以直接把你的 Abstract 按“术语统一 + 语法修正 + 塞入两组数字”完整重写一版(保持你原长度不变或更短),你只要告诉我:预测 RMSE 你想在摘要里报 哪个 state(x1/x2/x3) 或者报 三者平均。
“机理模型难、数据驱动兴起”这一段:把“痛点”说透、说清
这一部分,我的思路是数据驱动-机器学习-神经网络这样下来的,你根据我原本的思路,整合一下,给出一个更清晰的
下面给你一段按你原本“数据驱动 → 机器学习 → 神经网络(进而到预测控制)”思路串起来的更清晰版本(英文,可直接替换你 Introduction 里对应那几段)。我把“痛点”压在两件事上:** 机理模型维护难**、纯黑箱在多步滚动与在线优化里容易累积误差/不稳,正好和你后面 ResNet + 多步一致性训练、再到 Nash 协调自然衔接。
latexHowever, most MPC implementations still rely on mechanistic prediction models, which typically require extensive parameter identification and calibration. In tandem cold rolling, the rolling dynamics vary with steel grade, friction, rolling speed, and thermal conditions. Such operating-condition changes often induce model mismatch, leading to degraded prediction accuracy and consequently undermining closed-loop MPC performance. With the increasing availability of industrial data, data-driven methods have attracted growing interest in rolling processes. Recent progress in machine learning has enabled effective mapping from measured process variables to key quality-related outputs, providing an alternative to repeatedly re-identifying mechanistic parameters under varying regimes \citep{zong2025machine}. In particular, neural networks have been widely explored due to their strong nonlinear approximation capability and flexible modeling of coupled multivariable dynamics \citep{xu2024novel,cheng2015neural,xu2022control,han2019data,feng2025model}. Neural predictors have also been integrated with predictive control frameworks for rolling applications, aiming to improve prediction fidelity and tracking performance under constraints \citep{niu2025rbf,huang2022lstm,xia2023using,ma2025data}. Nevertheless, when neural predictors are embedded into receding-horizon optimization, two practical issues become critical. First, multi-step prediction is typically obtained by recursive rollout, where small one-step errors may accumulate and cause drift over the horizon, especially under strong inter-stand coupling. Second, purely black-box models may lack a structure consistent with the incremental evolution of rolling dynamics, which can reduce robustness and reliability when operating conditions change. These challenges motivate the development of a control-dependent residual neural surrogate with improved long-horizon robustness for distributed MPC in multi-stand cold rolling.
如果你想更贴你论文后文(你确实做了“区间内增量多项式参数化”),可以在最后一句再补半句(可选):
... where within-interval input increments are compactly parameterized for smooth actuation and efficient optimization.“可解释性”表述太硬(你文中确实有 “successfully overcome … interpretability”)
这个帮我把那一整句改一下
你原句是:
latexBy using residual neural networks to conduct data-driven modeling of the input-output relationship of the system, the problem of difficulty in interpretation of traditional neural networks has been successfully overcome.
可以改成更稳、更符合你正文“残差=增量/积分”解释的表述(保留原意但不夸张):
推荐替换句(最贴你论文逻辑):
latexBy using a residual neural network to model the input--output behavior of the system in an increment form, a mechanism-consistent interpretation of the learned state evolution can be provided.
如果你想更保守一点(不说“mechanism-consistent”,只说“improves interpretability”):
latexBy using a residual neural network to model the input--output behavior of the system in an increment form, the interpretability of the learned dynamics is improved compared with a purely black-box predictor.
% interactapasample.tex
% v1.05 - August 2017
\documentclass[]{interact}
% 让你仍然可以用 \citep / \citet 这类 natbib 命令
\usepackage[natbibapa,nodoi]{apacite}
\setlength\bibhang{12pt}
\renewcommand\bibliographytypesize{\fontsize{10}{12}\selectfont}
\usepackage{epstopdf}% To incorporate .eps illustrations using PDFLaTeX, etc.
\usepackage[caption=false]{subfig}% Support for small, sub' figures and tables %\usepackage[nolists,tablesfirst]{endfloat}% To separate' figures and tables from text if required
%\usepackage[doublespacing]{setspace}% To produce a `double spaced' document if required
%\setlength\parindent{24pt}% To increase paragraph indentation when line spacing is doubled
\usepackage{booktabs}
\usepackage[numbers,longnamesfirst,sort,authoryear,round]{natbib}
\bibpunct[, ]{(}{)}{;}{a}{,}{,}% Citation support using natbib.sty
\usepackage{tabularx}
\usepackage{array}
\usepackage{amsmath,amssymb}
\usepackage{graphicx}
\usepackage[caption=false,font=footnotesize]{subfig}
\usepackage{wrapfig}
\usepackage{multirow}
\usepackage[final]{graphicx} % 关键:即使 class 里是 draft,也强制显示图片
\bibpunct[, ]{(}{)}{;}{a}{,}{,}% Citation support using natbib.sty
\renewcommand\bibfont{\fontsize{10}{12}\selectfont}% To set the list of references in 10 point font using natbib.sty
%\usepackage[natbibapa,nodoi]{apacite}% Citation support using apacite.sty. Commands using natbib.sty MUST be deactivated first!
%\setlength\bibhang{12pt}% To set the indentation in the list of references using apacite.sty. Commands using natbib.sty MUST be deactivated first!
%\renewcommand\bibliographytypesize{\fontsize{10}{12}\selectfont}% To set the list of references in 10 point font using apacite.sty. Commands using natbib.sty MUST be deactivated first!
\theoremstyle{plain}% Theorem-like structures provided by amsthm.sty
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{proposition}[theorem]{Proposition}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{example}[theorem]{Example}
\theoremstyle{remark}
\newtheorem{remark}{Remark}
\newtheorem{notation}{Notation}
\begin{document}
\articletype{ARTICLE TEMPLATE}% Specify the article type or omit as appropriate
\title{Distributed MPC Control for Tandem Cold Rolling Based on Residual Network and Nash Equilibrium}
\author{
\name{Chenhao Guo\textsuperscript{a}, Jiahao Nie\textsuperscript{a}, Xiaojian Li\textsuperscript{a}\textsuperscript{b}\thanks{CONTACT Xiaojian Li. Email: [email protected]}}
\affil{\textsuperscript{a}College of Information Science and Engineering, Northeastern University, Shenyang, China; \textsuperscript{b}State Key Laboratory of Synthetical Automation for Process
Industries, Northeastern University, Shenyang, China}
}
\maketitle
\begin{abstract}
With the rapid development of multi-stand tandem cold rolling technology, high-precision and high-efficiency thickness–tension regulation has become increasingly important for product quality and process stability. However, the multi-stand cold rolling system has a complex structure. Not only is it difficult to establish the mechanism model, but there are also coupling relationships among the stands. This coupling relationships have led to the problem of optimization conflicts among multiple stands. Given that the complexity and coupling of the system, this study proposes a distributed model predictive control strategy based on residual neural network (ResNet) and Nash equilibrium (RNE-DMPC). A ResNet is trained to capture the input-output behaviors of the system, avoiding reliance on an explicit mechanism model. For online constrained multivariable optimization, a Nash-equilibrium-based distributed MPC is developed to coordinate multi-objective tension–thickness regulation across stands. The performance of the proposed method is verified by a series numerical simulations.
\end{abstract}
\begin{keywords}
Residual neural network; Nash equilibrium; Distributed model predictive control; Tandem cold rolling; Tension control; Thickness control
\end{keywords}
\section{Introduction}
Cold-rolled strip steel is a crucial intermediate product in the strip-steel manufacturing chain and significantly affects downstream process performance and final product quality \citep{mancini2012effects}. It is commonly produced by tandem cold rolling mills, where the strip passes sequentially through multiple stands and is progressively reduced in thickness by rolling forces. In such mills, regulating the exit thickness and inter-stand tensions is essential for product quality and process stability. Figure~\ref{AutoReg} shows a typical layout of a five-stand tandem cold rolling mill.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/x3.pdf}
\caption{Equipment layout of a five-stand tandem cold rolling mill.}\label{AutoReg}
\end{figure*}
From a control perspective, thickness--tension regulation is a large-scale multivariable problem featuring strong coupling, nonlinearities, constraints, time-varying parameters, and disturbances. These characteristics complicate reliable prediction of rolling dynamics and make high-performance closed-loop control difficult to sustain over changing operating conditions.
PID control remains prevalent in practice due to its simplicity, yet fixed-parameter tuning often struggles in strongly coupled multivariable rolling processes \citep{gupta2025grid,liu2025data,hasilci2025superior}. Model predictive control (MPC) provides a systematic framework to handle multivariable coupling and constraints and has been widely investigated in rolling applications \citep{yaghini2024linear,luis2019trajectory,han2021hierarchical,song2025model,sun2017dynamic}. Depending on the system scale and information interaction, MPC implementations can be centralized, decentralized, or distributed \citep{song2025model}.
In particular, distributed MPC decomposes the large-scale optimization into local problems solved in parallel with limited information exchange, improving scalability and reducing computational burden for interconnected systems \citep{scattolini2009architectures}. Multi-stand tandem cold rolling is naturally an interconnected large-scale system, making distributed optimization particularly suitable. \citep{zhao2025multi,ji2022coordinate,wang2025distributed}.
However, most MPC implementations still rely on mechanistic prediction models, which typically require extensive parameter identification and calibration. In tandem cold rolling, the rolling dynamics vary with steel grade, friction, rolling speed, and thermal conditions. Such operating-condition changes often induce model mismatch, leading to degraded prediction accuracy and consequently undermining closed-loop MPC performance.
With the increasing availability of industrial data, data-driven methods have attracted growing interest in rolling processes. Recent progress in machine learning has enabled effective mapping from measured process variables to key quality-related outputs, providing an alternative to repeatedly re-identifying mechanistic parameters under varying regimes \citep{zong2025machine}.
In particular, neural networks have been widely explored due to their strong nonlinear approximation capability and flexible modeling of coupled multivariable dynamics \citep{xu2024novel,cheng2015neural,xu2022control,han2019data,feng2025model}. At present, Niu et al. have developed a distributed nonlinear model predictive control based on RBF neural network for tandem cold rolling process \citep{niu2025rbf}; Huang et al. proposed the LSTM-AGD multi-mode adaptive MPC framework, which reduces overshoot and improves accuracy in a steady state \citep{huang2022lstm}; Xia et al. used BFGS to optimize the ANN topology and data-driven high-precision prediction of key parameters in cold continuous rolling \citep{xia2023using}.
Nevertheless, when neural predictors are embedded into receding-horizon optimization, two practical issues become critical. First, multi-step prediction is typically obtained by recursive rollout, where small one-step errors may accumulate and cause drift over the horizon, especially under strong inter-stand coupling. Second, purely black-box models may lack a structure consistent with the incremental evolution of rolling dynamics, which can reduce robustness and reliability when operating conditions change. These challenges motivate the development of a control-dependent residual neural surrogate with improved long-horizon robustness for distributed MPC in multi-stand cold rolling.
Therefore, to address these issues, this paper integrates the ResNet into MPC for prediction. This data-based ResNet can model dynamic evolution processes,aiming to improve prediction accuracy and interpretability for thickness and tension regulation.
In multi-stand rolling, inter-stand tension propagation introduces strong coupling among local control objectives: actions at one stand affect neighboring tensions and thickness responses, leading to potential strategy conflicts among local MPC controllers. Game-theoretic coordination provides a principled way to resolve such conflicts. Nash-equilibrium-based distributed predictive control has been studied for thickness--tension coordination in rolling mills \citep{hu2021nash,ye2021adaptive}. However, the effectiveness and convergence of online equilibrium seeking can be sensitive to prediction accuracy, especially when operating conditions vary and modeling errors increase.Although some studies have introduced Nash equilibrium into the distributed predictive control of thickness tension, this iterative process is highly dependent on the accuracy of the prediction model and the shared information. Once the on-site operating conditions change and lead to an increase in prediction errors, the equilibrium iterative effect increased prediction error may slow convergence or lead to termination at an approximate equilibrium, thereby degrading coordination quality and constraint satisfaction.
Therefore, this paper proposes a distributed model predictive control strategy based on residual neural networks and Nash equilibrium (RNE-DMPC).
The main contributions of this article are as follows:
\begin{enumerate}
\item A residual neural network model is developed to capture the input--output relationship, aiming to improve prediction accuracy while enhancing model interpretability.
\item For strongly coupled multi-stand thickness--tension coordination, we propose a Nash-equilibrium-based distributed MPC (RNE-DMPC). The trained residual network is embedded as the local predictor, and Nash coordination is achieved online via consistency penalty and a relaxed best-response iteration.
\end{enumerate}
The rest of this article is organized as follows: Section 2 describes the construction of the dataset. Section 3 builds the residual neural network. Section 4 discusses the specific content of RNE-DMPC. Section 5 presents the simulation results of the thickness and tension of the five-stand system. Section 6 presents the conclusion of the full text.
%========================
\section{Construction of Dataset}
%========================
Consider a five-stand tandem cold rolling mill indexed by .
Let denote the -th sampling instant and the corresponding sampling interval length.
Define the local within-interval time variable .
Let denote the exit thickness of stand (),
and let denote the inter-stand strip tension between stand and ().
The system is strongly coupled because the inter-stand tensions propagate along the mill line and are affected by neighboring stands' actions.
Let and be the desired references (setpoints) given by process requirements. We define deviation variables
\begin{equation}
\Delta h_i(t)\triangleq h_i(t)-h_i^{\mathrm{ref}}(t),\qquad
\Delta T_i(t)\triangleq T_i(t)-T_i^{\mathrm{ref}}(t).
\label{eq:dev_def}
\end{equation}
Throughout this paper, the symbol ``'' attached to states always means deviation from reference.
For each stand , we choose the local deviation state as
\begin{equation}
\Delta x_i(t)\triangleq
\begin{bmatrix}
\Delta h_i(t)\
\Delta T_{i-1}(t)\
\Delta T_i(t)
\end{bmatrix}\in\mathbb{R}^{d},\qquad d=3,
\label{eq:xi_def_clean}
\end{equation}
with the boundary convention and to keep a unified dimension for all stands.
For a five-stand tandem mill, the dominant coupling is between adjacent stands, hence we define
\begin{equation}
Z_1={2},\quad
Z_i={i-1,i+1}\ (i=2,3,4),\quad
Z_5={4}.
\label{eq:Zi_clean}
\end{equation}
Define the neighbor-state stack
\begin{equation}
\Delta x_{Z_i}(t_n)=\mathrm{col}{\Delta x_k(t_n),|,k\in Z_i}.
\label{eq:xZi_clean}
\end{equation}
\begin{bmatrix}
\Delta s_i(t_n)\
\Delta v_i(t_n)
\end{bmatrix}.
\label{eq:du_discrete_clean}
\end{equation}
Throughout this paper, the symbol ``'' attached to inputs means sample-to-sample increment.
Thus, and are conceptually different, and this is fixed by definition.
Let denote exogenous disturbances.
We denote the interval-level equivalent disturbance by .
denotes the identity matrix; denotes the zero matrix.
M_d,\Delta x_i(t_n)
+
N_d,\Delta u_i(t_n)
+
F_d,\Delta d_i(t_n),
\label{eq:linear_form_concept}
\end{equation}
where represent equivalent discrete-time matrices around operating conditions.
In a practical five-stand cold rolling mill, accurately deriving and identifying these matrices and disturbance models from first principles is difficult,
due to strong coupling, unmodeled nonlinearities, and time-varying operating regimes.
Therefore, this paper aims to learn a high-fidelity approximation of the interval evolution from data and then embed it into distributed MPC.
\begin{remark}
In fact, due to the existence of complex coupling relationships, it is difficult to directly and accurately establish \eqref{eq:linear_form_concept}
based on first principles. Therefore, in this paper, we learn an approximate mapping of \eqref{eq:true_mapping_clean} from data.
\end{remark}
Although decisions are updated at discrete instants , the hydraulic gap and drive systems evolve continuously inside each interval,
and abrupt within-interval changes may excite tension oscillations and deteriorate thickness stability.
Thus, parameterizing the within-interval increment trajectory by a low-order polynomial:
(i) yields a compact finite-dimensional decision representation;
(ii) enforces smooth profiles inside the interval;
(iii) enables enforcing increment constraints for all .
This is appropriate when is not excessively large relative to actuator bandwidth and the within-interval evolution is well approximated by a low-order basis.
\sum_{m=0}^{n}\Gamma_{i,nm}\tau^{m},
\qquad \tau\in[0,\delta_n],
\label{eq:du_poly_vec_clean}
\end{equation}
where are coefficient vectors (), and is the chosen polynomial order.
In simulations, one may select (quadratic) for a good accuracy--complexity trade-off.
Component-wise, \eqref{eq:du_poly_vec_clean} corresponds to
\begin{equation}
\begin{aligned}
\Delta s_{i,n}(\tau) &= \sum_{m=0}^{n}\gamma^{(s)}{i,nm}\tau^{m},\
\Delta v{i,n}(\tau) &= \sum_{m=0}^{n}\gamma^{(v)}{i,nm}\tau^{m}.
\end{aligned}
\label{eq:du_components_clean}
\end{equation}
Define the stacked parameter vector
\begin{equation}
\Gamma{i,n}\triangleq
\big[
(\Gamma_{i,n0})^\top,,
(\Gamma_{i,n1})^\top,,
\ldots,,
(\Gamma_{i,nn})^\top
\big]^\top
\in\mathbb{R}^{p},
\qquad
p=(n+1)n_u.
\label{eq:Gamma_clean}
\end{equation}
Here, is the baseline increment at , while () describe the higher-order variation rates.
Define the interval-averaged equivalent increments as
\begin{equation}
\begin{aligned}
\Delta u_i(t_n) &\triangleq \frac{1}{\delta_n}\int_0^{\delta_n}\Delta u_{i,n}(\tau),d\tau,\
\Delta d_i(t_n) &\triangleq \frac{1}{\delta_n}\int_0^{\delta_n}\Delta d_i(\tau),d\tau.
\end{aligned}
\label{eq:avg_def_clean}
\end{equation}
With \eqref{eq:du_poly_vec_clean}, the input average has a closed form:
\begin{equation}
\Delta u_i(t_n)=
\sum_{m=0}^{n}\Gamma_{i,nm}\frac{\delta_n^{m}}{m+1}.
\label{eq:avg_closed_clean}
\end{equation}
Let denote the sampling domain of deviation states and neighbor stacks ,
and let denote the sampling domain of polynomial parameters .
These domains specify the operating envelope used to generate supervised training data.
Given the above parameterization, one training sample is generated on each interval .
In addition to the local deviation state, the neighbor deviation states are included to represent inter-stand coupling.
The process is summarized in Table~\ref{tab:interval_sample_generation_en}.
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.15}
\caption{Procedure for generating one interval-level sample on (five-stand coupled mill).}
\label{tab:interval_sample_generation_en}
\begin{tabularx}{\linewidth}{>{\centering\arraybackslash}p{0.09\linewidth} X}
\toprule
\textbf{Step} & \textbf{Operation} \
\midrule
1 & \textbf{State sampling:} sample and from . \
2 & \textbf{Parameter sampling:} draw (coefficients for both and ). \
3 & \textbf{Control construction:} compute via \eqref{eq:du_poly_vec_clean}. \
4 & \textbf{State propagation:} integrate the \emph{five-stand coupled} mill model on (e.g., RK4) using the within-interval control trajectory, and record . \
\bottomrule
\end{tabularx}
\end{table}
Accordingly, an interval sample for subsystem can be represented as
\begin{equation}
\mathcal{D}{i,n}=\big{\Delta x_i(t_n),\ \Delta x{Z_i}(t_n),\ \Delta u_{i,n}(\tau),\ \Delta x_i(t_{n+1})\big}.
\label{eq:interval_sample_clean}
\end{equation}
Note that is fully determined by via \eqref{eq:du_poly_vec_clean},
therefore it is sufficient to store as the learning input.
For each subsystem , by repeating the above procedure across multiple intervals and randomized draws,
the local one-step training dataset is formed as
\begin{equation}
\begin{split}
S_i=\Big{&
\big(\Delta x_i^{(j)}(t_n),,\Delta x_{Z_i}^{(j)}(t_n),,\Delta x_i^{(j)}(t_{n+1});,
\Gamma_{i,n}^{(j)},,\delta_n^{(j)}\big)
\ \Big|\ j=1,\ldots,J
\Big}.
\end{split}
\label{eq:S_i_clean}
\end{equation}
Here is the number of one-step samples for subsystem .
The overall dataset for the five-stand mill is denoted by .
The point-cloud visualization of the training dataset is shown in Figure~\ref{2}.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=0.5]{picture/Fig2.pdf}
\caption{Point cloud map of the training dataset.}\label{2}
\end{figure*}
The one-step set is sufficient for one-step regression, but it is not sufficient for training with multi-step rollout loss
and reciprocal-consistency regularization, because these objectives require ground-truth deviation-state trajectories over a horizon of consecutive intervals.
Therefore, without changing the single-interval sampling mechanism above, we additionally organize the offline-simulated samples
into -step trajectory segments.
Specifically, for each starting time we generate a segment of length by consecutively sampling
(and the corresponding inputs/disturbances),
and integrating the five-stand coupled mill model over for .
Hence, we obtain the deviation-state sequence as well as the neighbor stacks
.
Define a -step segment sample for subsystem as
\begin{equation}
\begin{aligned}
\mathcal{W}{i,n}=
\Big{&
\big(\Delta x_i(t{n+s}),,\Delta x_{Z_i}(t_{n+s}),,\Gamma_{i,n+s},,\delta_{n+s}\big){s=0}^{K-1}; \
&\big(\Delta x_i(t{n+s+1})\big){s=0}^{K-1}
\Big}.
\end{aligned}
\label{eq:segment_clean}
\end{equation}
By repeating the above segment generation, we form the multi-step training set
\begin{equation}
S_i^{(K)}=\Big{\mathcal{W}{i,n}^{(j)}\ \Big|\ j=1,\ldots,J_K\Big},
\label{eq:S_i_K_clean}
\end{equation}
where is the number of -step segment samples.
Note that can be viewed as the marginal one-step projection of (keeping only ),
thus the original dataset design is preserved, and only an additional \emph{segment organization} is introduced for multi-step training.
%========================
\section{Construction of Residual Neural Network}
%========================
\subsection{Residual Neural Network Structure Construction and Training Method}
Given the dataset, the neural network model is trained to learn a stand-wise, control-dependent one-step evolution law of deviation states:
\begin{equation}
\Delta x_i(t_{n+1})
\approx
\Delta x_i(t_n)+
\mathcal{N}i!\Big(\Delta x_i(t_n),,\Delta x{Z_i}(t_n),,\Gamma_{i,n},,\delta_n;,\Theta_i\Big),
\label{eq:learned_dyn_clean}
\end{equation}
where outputs the one-step deviation-state change and are trainable parameters.
\begin{remark}
If does not take control information as input (here and ),
the predictor becomes an autoregressive model that only reproduces trajectories under the training input patterns
and cannot answer the counterfactual question: ``what will happen if we choose a different roll gap and speed trajectory?''
Since MPC optimizes over candidate decisions, a control-dependent predictor \eqref{eq:learned_dyn_clean} is necessary
to evaluate the predicted thickness and tension behavior under different candidate actuator trajectories.
\end{remark}
Let (state dimension), be the number of neighbors of stand in \eqref{eq:Zi_clean}, and in \eqref{eq:Gamma_clean}.
Define the input vector
\begin{equation}
X_{i,\text{in}} \triangleq
\big[
\Delta x_i(t_n)^\top,,
\Delta x_{Z_i}(t_n)^\top,,
\Gamma_{i,n}^\top,,
\delta_n
\big]^\top
\in \mathbb{R}^{d(1+|Z_i|)+p+1}.
\label{eq:X_in_clean}
\end{equation}
The network mapping is
\begin{equation}
\mathcal{N}_i:\mathbb{R}^{d(1+|Z_i|)+p+1}\rightarrow\mathbb{R}^{d}.
\end{equation}
To improve training stability and long-horizon rollout robustness, we use a residual form.
Let be a selection matrix extracting the local state block:
\begin{equation}
\hat{I}i = [I_d,, 0{d\times(d|Z_i|+p+1)}].
\label{eq:Ihat_clean}
\end{equation}
Then the one-step predictor is written as
\begin{equation}
X_{i,\text{out}} = \hat{I}i X{i,\text{in}} + \mathcal{N}i(X{i,\text{in}}; \Theta_i),
\label{eq:res_predict_clean}
\end{equation}
where represents the predicted .
This structure implements a baseline-plus-correction interpretation:
the shortcut propagates the current deviation state , while the network learns the correction capturing
unmodeled nonlinearities and inter-stand coupling (via ) under varying operating conditions.
To improve robustness when varies, we introduce an auxiliary branch inside :
\begin{equation}
\mathcal{N}i(X{i,\text{in}};\Theta_i)\triangleq
\psi_i(X_{i,\text{in}};\Theta_{\psi_i}) + \rho_i(X_{i,\text{in}};\theta_i),
\label{eq:aux_clean}
\end{equation}
where is a lightweight feedforward branch that captures low-frequency/scale effects strongly related to ,
and captures the remaining nonlinear coupling corrections.
When , the model reduces to a standard residual network.
For the -th sample in \eqref{eq:S_i_clean}, define
\begin{equation}
X_{i,\text{in}}^{(j)} =
\big[
\Delta x_i^{(j)}(t_n),\ \Delta x_{Z_i}^{(j)}(t_n),\
\Gamma_{i,n}^{(j)},\ \delta_n^{(j)}
\big]^{\top},
\end{equation}
and the supervised residual target
\begin{equation}
\Delta r_i^{(j)}=\Delta x_i^{(j)}(t_{n+1})-\Delta x_i^{(j)}(t_n).
\label{eq:target_clean}
\end{equation}
To suppress accumulation drift induced by long-horizon recursion and to improve long-term predictive stability,
we train the forward predictor jointly with an auxiliary backward residual model
and impose a multi-step reciprocal-consistency regularization over a -step segment from .
Construct a backward residual network
\begin{equation}
\mathcal{B}i:\mathbb{R}^{d(1+|Z_i|)+p+1}\rightarrow\mathbb{R}^{d},
\end{equation}
parameterized by . For the backward step associated with interval , define
\begin{equation}
\begin{aligned}
X{i,\mathrm{in}}^{b}
&=
\big[
\Delta x_i(t_{n+1}),\ \Delta x_{Z_i}(t_{n+1}),
\Gamma_{i,n},\ \delta_n
\big]^{\top},\
X_{i,\mathrm{out}}^{b}
&=
\hat{I}i X{i,\mathrm{in}}^{b} + \mathcal{B}i(X{i,\mathrm{in}}^{b};\bar{\Theta}i),
\end{aligned}
\label{eq:back_clean}
\end{equation}
where represents the backward estimate of .
The supervised backward residual target is
\begin{equation}
\Delta r_i^{b}=\Delta x_i(t_n)-\Delta x_i(t{n+1}).
\end{equation}
Given a segment sample , initialize
\begin{equation}
\Delta \hat{x}i(t_n)=\Delta x_i(t_n),
\end{equation}
and recursively apply the forward predictor for steps:
\begin{equation}
\begin{aligned}
\Delta \hat{x}i(t{n+s+1})
&=
\Delta \hat{x}i(t{n+s})
+
\mathcal{N}i!\Big(
\Delta \hat{x}i(t{n+s}),,\Delta \hat{x}{Z_i}(t{n+s}),,
\Gamma_{i,n+s},,\delta_{n+s};,\Theta_i
\Big),\
&\qquad s=0,\ldots,K-1.
\end{aligned}
\label{eq:fwd_roll_clean}
\end{equation}
Set the terminal condition
\begin{equation}
\Delta \bar{x}i(t{n+K})=\Delta \hat{x}i(t{n+K}),
\end{equation}
and roll back using :
\begin{equation}
\begin{aligned}
\Delta \bar{x}i(t{n+s})
&=
\hat{I}i X{i,\mathrm{in}}^{b}(t_{n+s})
+
\mathcal{B}i!\Big(X{i,\mathrm{in}}^{b}(t_{n+s});,\bar{\Theta}i\Big),
\quad s=K-1,\ldots,0,
\end{aligned}
\label{eq:bwd_roll_clean}
\end{equation}
where
\begin{equation}
X{i,\mathrm{in}}^{b}(t_{n+s})=
\big[
\Delta \bar{x}i(t{n+s+1}),\ \Delta \hat{x}{Z_i}(t{n+s+1}),
\Gamma_{i,n+s},\ \delta_{n+s}
\big]^{\top}.
\end{equation}
\sum_{s=0}^{K}
\left|
\Delta \hat{x}i(t{n+s})-\Delta \bar{x}i(t{n+s})
\right|^2.
\end{equation}
We jointly minimize:
\begin{equation}
\begin{aligned}
L_{\mathrm{1step}}(\Theta_i)
&= \frac{1}{J_K}\sum_{j=1}^{J_K}\frac{1}{K}\sum_{s=0}^{K-1}
\Big|
\big(\Delta x_i^{(j)}(t_{n+s+1})-\Delta x_i^{(j)}(t_{n+s})\big)
-\mathcal{N}i!\left(
X{i,\mathrm{in}}^{(j)}(t_{n+s});\Theta_i
\right)
\Big|^2,\[2mm]
L_{\mathrm{bwd}}(\bar{\Theta}i)
&= \frac{1}{J_K}\sum{j=1}^{J_K}\frac{1}{K}\sum_{s=0}^{K-1}
\Big|
\big(\Delta x_i^{(j)}(t_{n+s})-\Delta x_i^{(j)}(t_{n+s+1})\big)
-\mathcal{B}i!\left(
X{i,\mathrm{in}}^{b,(j)}(t_{n+s});\bar{\Theta}i
\right)
\Big|^2,\[2mm]
L{\mathrm{msrp}}(\Theta_i,\bar{\Theta}i)
&= \frac{1}{J_K}\sum{j=1}^{J_K} E_i^{(j)}(t_n),\[2mm]
L_{\mathrm{roll}}(\Theta_i)
&= \frac{1}{J_K}\sum_{j=1}^{J_K}\sum_{s=1}^{K}
\Big|
\Delta x_i^{(j)}(t_{n+s})-\Delta \hat{x}i^{(j)}(t{n+s})
\Big|^2.
\end{aligned}
\label{eq:loss_clean}
\end{equation}
Here, enforces one-step accuracy; explicitly suppresses long-horizon drift under recursion;
regularizes the learned dynamics by enforcing reciprocal consistency between forward and backward rollouts;
and trains the backward model for the consistency regularization.
In implementation, these terms are combined as
\begin{equation}
L_{\mathrm{total}}=\lambda_1 L_{\mathrm{1step}}+\lambda_2 L_{\mathrm{roll}}+\lambda_3 L_{\mathrm{msrp}}+\lambda_4 L_{\mathrm{bwd}},
\end{equation}
where are tuned on a validation set.
\Delta x_i(t_n)
+
\mathcal{N}i!\Big(
\Delta x_i(t_n),,\Delta x{Z_i}(t_n),,
\Gamma_{i,n},,\delta_n;,\Theta_i^*
\Big),
\label{eq:pred_clean}
\end{equation}
and multi-step prediction is obtained by recursive rollout of \eqref{eq:pred_clean}.
This learned predictor is the internal model used by the MPC optimizer in the next section.
Finally, network parameters are optimized using Adam:
\begin{equation}
\Theta_{i,t+1} = \Theta_{i,t} - \alpha \frac{\hat{g}{i,t}}{\sqrt{\hat{r}{i,t}} + \varepsilon},
\end{equation}
where is the learning rate (we use to avoid conflict with other symbols),
and are bias-corrected moment estimates, and is a small constant for numerical stability.
Figure~\ref{fig:rnn_logic} illustrates the overall structure.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.85]{picture/x6.pdf}
\caption{Logic diagram of the residual neural network.}
\label{fig:rnn_logic}
\end{figure}
\subsection{Explainability of the residual network}
\Delta x_i(t_n)
+
\underbrace{\int_{0}^{\delta_n}
f_i!\Big(\Delta x_i(t_n+\tau),,\Delta x_{Z_i}(t_n+\tau),,u_i(t_n+\tau),,d_i(t_n+\tau)\Big),d\tau}{\triangleq\ \varphi{i,n}},
\label{eq:increment_integral}
\end{equation}
where is the one-interval state increment generated by the mechanistic dynamics.
Our learned model \eqref{eq:learned_dyn_clean} adopts the same increment form as \eqref{eq:increment_integral}:
\begin{equation}
\Delta x_i(t_{n+1})
\approx
\Delta x_i(t_n)
+
\mathcal{N}i!\Big(\Delta x_i(t_n),,\Delta x{Z_i}(t_n),,\Gamma_{i,n},,\delta_n;\Theta_i\Big).
\end{equation}
Here, plays the role of a data-driven approximation of the integral increment ,
i.e., it approximates the accumulated effect of the mechanistic dynamics over .
This is consistent with the well-known interpretation that a residual network behaves like a one-step time integrator:
the identity path propagates the current state, while the residual branch represents the increment over the time lag.
Inside each interval, we do not optimize point-wise but parameterize the increment trajectory by
.
Hence, the mechanistic increment in \eqref{eq:increment_integral} depends on the \emph{whole} within-interval trajectory.
Feeding into is therefore a compact way to represent how different candidate
gap/speed trajectories change the integral effect and thus the next thickness--tension state.
When varies and is not very small, directly learning may be harder.
Motivated by the generalized residual idea,
we decompose the increment predictor into two parts in \eqref{eq:aux_clean}:
\begin{equation}
\mathcal{N}_i(\cdot)=\psi_i(\cdot)+\rho_i(\cdot).
\end{equation}
Conceptually, captures low-frequency and scale effects strongly related to ,
while captures the remaining nonlinear coupling corrections.
This provides a mechanism-consistent interpretation: a baseline increment plus a residual correction
that compensates unmodeled nonlinearities and inter-stand coupling.
%========================
\section{Nash Equilibrium-Based RNE-DMPC}
%========================
The five-stand tandem cold rolling system is strongly coupled through inter-stand tension propagation.
As a result, changes in control actions (roll gap and stand speed) at one stand can affect both upstream and downstream stands,
making centralized online optimization over all stands' decision variables computationally demanding.
To mitigate this issue, we decompose the global predictive-control problem into local subproblems associated with individual stands.
Each local controller optimizes its own decision variables while accounting for coupling via limited information exchange with neighboring controllers.
Motivated by game-theoretic coordination \citep{rawlings2008coordinating}, we formulate distributed coordination as a Nash-equilibrium-seeking iteration.
Based on the trained residual neural network surrogate model, we construct a Nash-equilibrium-based distributed MPC method (RNE-DMPC)
for coordinated thickness--tension regulation and tracking. The overall control structure is shown in Figure~\ref{4}.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/x2.pdf}
\caption{Schematic diagram of the control architecture for a tandem cold rolling mill.}\label{4}
\end{figure*}
At sampling time , stand chooses the polynomial-parameter sequence
, where .
Let
denote the joint strategy profile, and let denote the collection of all strategies except stand .
Given the current measured/estimated deviation state and the strategies
,
the multi-step prediction used by stand is written explicitly as
\begin{equation}
\begin{aligned}
\Delta \hat{x}i(t{n+s+1};\mathbf{\Gamma}i,\mathbf{\Gamma}{Z_i})
&=
\Delta \hat{x}i(t{n+s};\mathbf{\Gamma}i,\mathbf{\Gamma}{Z_i})
+
\mathcal{N}i!\Big(
\Delta \hat{x}i(t{n+s};\cdot),,
\Delta \hat{x}{Z_i}(t_{n+s};\mathbf{\Gamma}{Z_i}),\
&\qquad
\Gamma{i,n+s},,
\delta_{n+s};\Theta_i^*
\Big),
\end{aligned}
\label{eq:rollout_mpc_game}
\end{equation}
for , with initialization .
Here the neighbor stack is generated from neighbors' strategies via the same learned predictors.
\sum_{m=0}^{n}\Gamma_{i,n+s,m}\frac{\delta_{n+s}^{m}}{m+1}.
\label{eq:du_avg_clean}
\end{equation}
\begin{remark}
Because inter-stand tension is jointly affected by the adjacent stands and ,
the predicted evolution of depends on neighbors' future actions,
hence the MPC problems are not independent but form a coupled dynamic game.
\end{remark}
\mathrm{col}{\Gamma_{i,n},\Gamma_{i,n+1},\ldots,\Gamma_{i,n+N_c-1}}
\in \mathbb{R}^{pN_c}.
\end{equation}
In deviation coordinates, the regulation/tracking objective is , i.e.
\begin{equation}
\Delta x_{i,\mathrm{ref}}(t_{n+s})\equiv 0\in\mathbb{R}^{d},\qquad d=3.
\end{equation}
Recall .
Define the row selectors
\begin{equation}
C^- \triangleq [0\ \ 1\ \ 0]\in\mathbb{R}^{1\times 3},\qquad
C^+ \triangleq [0\ \ 0\ \ 1]\in\mathbb{R}^{1\times 3},
\end{equation}
so that (upstream interface) and (downstream interface).
C^-\Delta \hat{x}{i+1}(t{n+s};\mathbf{\Gamma}),\qquad i=1,\ldots,4.
\label{eq:shared_tension_mismatch}
\end{equation}
\sum_{s=1}^{N_p}
\left|
\Delta \hat{x}i(t{n+s};\mathbf{\Gamma}i,\mathbf{\Gamma}{Z_i})
\right|{Q_i}^{2}
+
\sum{s=0}^{N_c-1}
\left|\Gamma_{i,n+s}\right|_{R_i}^{2}
+
J_i^{\mathrm{cpl}}(\mathbf{\Gamma}i;\mathbf{\Gamma}{-i})
\label{eq:Ji_game}
\end{equation}
where weights thickness and tension deviations, and penalizes actuation magnitudes.
C^-\Delta \hat{x}{i+1}(t{n+s};\mathbf{\Gamma})
\big|^2
\Big),
\end{aligned}
\label{eq:coupling_cost}
\end{equation}
with and indicator .
This term makes the coupling conflict explicit: unilateral actions that locally reduce thickness error may worsen shared-tension
compatibility and thus increase , and also affect neighbors' objectives.
We enforce the absolute-input bounds and within-interval increment bounds.
Absolute input bounds (roll gap and speed):
\begin{equation}
u_{i,\min}\le u_i(t_{n+s})\le u_{i,\max},
\qquad s=0,\ldots,N_p-1,
\label{eq:u_abs_game}
\end{equation}
where .
Within-interval increment-trajectory bounds:
\begin{equation}
\Delta u_{i,\min}\le \Delta u_{i,n+s}(\tau;\Gamma_{i,n+s})\le \Delta u_{i,\max},
\qquad \forall\tau\in[0,\delta_{n+s}],\ s=0,\ldots,N_p-1,
\label{eq:du_traj_game}
\end{equation}
where is given by the polynomial parameterization.
\sum_{m=0}^{n}\Gamma_{i,n+s,m}\frac{\delta_{n+s}^{m}}{m+1},
\label{eq:du_avg_game}
\end{equation}
and propagate the absolute input along the horizon by
\begin{equation}
u_i(t_n)=u_i(t_{n-1})+\Delta u_i(t_n),\qquad
u_i(t_{n+s})=u_i(t_{n+s-1})+\Delta u_i(t_{n+s}),\ s=1,\ldots,N_p-1,
\label{eq:u_prop_game}
\end{equation}
where is the applied (measured) input from the previous sampling instant.
Compact feasible set:
\begin{equation}
\Omega_i \triangleq
\Big{\mathbf{\Gamma}_i\ \Big|\
\eqref{eq:rollout_mpc_game}\ \text{holds and}
\eqref{eq:u_abs_game},\eqref{eq:du_traj_game},\eqref{eq:u_prop_game}\ \text{are satisfied}
\Big}.
\label{eq:Omega_i_game}
\end{equation}
\arg\min_{\mathbf{\Gamma}_i\in\Omega_i}\
J_i(\mathbf{\Gamma}i;\mathbf{\Gamma}{-i}).
\label{eq:local_BR}
\end{equation}
Because the learned surrogate is differentiable, \eqref{eq:local_BR} can be solved by standard gradient-based NLP solvers.
At each sampling time , the distributed MPC coordination induces a finite-horizon dynamic game:
players are stands ; strategy sets are ; and payoff (cost) functions are
defined in \eqref{eq:Ji_game}--\eqref{eq:coupling_cost}.
A joint strategy profile
is a Nash equilibrium if
\begin{equation}
\forall i\in{1,\ldots,5},\qquad
\mathbf{\Gamma}i^*\in
\arg\min{\mathbf{\Gamma}_i\in\Omega_i}
J_i(\mathbf{\Gamma}i;\mathbf{\Gamma}{-i}^*).
\label{eq:NE_def}
\end{equation}
This definition explicitly characterizes the strategic coupling:
each player's optimal decision depends on neighbors' decisions through the shared-tension dynamics and the coupling term.
(1-\omega)\mathbf{\Gamma}_i^{(l-1)}
+
\omega,\mathbf{\Gamma}_i^{\mathrm{BR},(l)},
\qquad \omega\in(0,1].
\label{eq:relaxed_BR}
\end{equation}
The relaxation factor mitigates oscillations caused by strong coupling and improves practical convergence.
The Nash equilibrium is computed through distributed best-response iterations, summarized in Table ~\ref{tab:nash_iter_en}.
\max_i
\frac{\left|
\mathbf{\Gamma}_i^{(l)}-\mathbf{\Gamma}_i^{(l-1)}
\right|_2}{
\left|
\mathbf{\Gamma}_i^{(l-1)}
\right|_2+\epsilon},
\end{equation}
with small.
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.12}
\setlength{\tabcolsep}{3.5pt}
\caption{Relaxed distributed Nash best-response iteration for RNE-DMPC (five-stand).}
\label{tab:nash_iter_en}
\begin{tabularx}{\linewidth}{>{\centering\arraybackslash}p{0.11\linewidth} X}
\toprule
\textbf{Step} & \textbf{Description} \
\midrule
A &
Initialize and warm-start (e.g., from the previous sampling time). \
B &
Communicate (or the induced predicted trajectories) among neighbors; form . \
C &
Given , solve the best-response NLP \eqref{eq:local_BR} to obtain . \
D &
Update the relaxed strategy using \eqref{eq:relaxed_BR} and compute the induced predictions
via \eqref{eq:rollout_mpc_game}. \
E &
Broadcast and the predicted interface quantities needed in \eqref{eq:coupling_cost}
(e.g., ) to neighbors. \
F &
Compute ; if , stop and set ; otherwise and repeat. \
\bottomrule
\end{tabularx}
\end{table}
\sum_{m=0}^{n}\Gamma_{i,n,m}^*\frac{\delta_n^{m}}{m+1},
\end{equation}
followed by .
The overall control flow chart is shown in Fig~\ref{liu}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{picture/x5.pdf}
\caption{The overall system control flow chart.}\label{liu}
\end{figure}
\begin{remark}
Equations \eqref{eq:Ji_game}--\eqref{eq:coupling_cost} and the NE definition \eqref{eq:NE_def} explicitly show that each player's objective
depends on others' strategies through the coupled prediction dynamics and the shared-tension interface mismatch penalty,
thus making the multi-stand strategic interaction and coordination mechanism transparent.
\end{remark}
\section{Simulations}
To verify the effectiveness of the proposed residual-neural-network and Nash-equilibrium based distributed MPC (RNE-DMPC) for coordinated thickness--tension regulation in a five-stand cold tandem rolling system, a distributed control architecture composed of multiple local controllers and a communication/coordination module is considered. Each stand is treated as a subsystem, and information exchange through the communication system enforces coordination among subsystems, enabling distributed thickness--tension control for the interconnected plant.
In this study, the simulation data were generated using the mathematical model reported by Hu et al.~\citep{hu2021nash}.
\sqrt{\frac{1}{N}\sum_{i=1}^{N}\left(x_i-\hat{x}_i\right)^2 }.
\end{equation}
where is the number of samples, and are the actual and predicted values, respectively.
\subsection{Surrogate-model prediction comparison}
In this paper, all of the spatial dynamics network and temporal evolution networks are constructed
based on the ResNet architecture and trained using the loss function in \eqref{eq:loss_clean}. Specifically, one-block ResNet method is chosen, with 5 hidden layers in each block, and each layer contains 160 neurons. And the optimizer is chosen as Adam with betas equal (0.9,0.99), initial learning rate is 0.001 and will decay to 0.00001 gradually. All networks are trained in the PyTorch
environment for 500 epochs. Meanwhile, the multi-order parameterization in the simulation part uses the second-order approach.
To justify the use of the proposed residual predictor in MPC, we first compare its prediction accuracy against representative baseline models, including
a plain feedforward network (MLP) with comparable model size trained by one-step regression\citep{Hornik1989MLPUniversal}, and a sequence model (LSTM) trained for trajectory prediction\citep{Hochreiter1997LSTM}.
All models are trained and tested under the same dataset split.
Figure~\ref{fig:pred_compare} shows an illustrative trajectory overlay on a representative test case.
It can be observed that the proposed residual network produces a closer match to the ground truth and exhibits less drift in multi-step rollouts.
The aggregated results can be summarized in Table~\ref{tab:pred_rmse_states}, where the proposed residual network achieves the lowest prediction errors among the compared models.
\begin{figure*}[htbp]
\centering
\subfloat[State ]{%
\includegraphics[width=0.32\linewidth]{picture/pred1.pdf}%
\label{F1_pred}%
}\hfill
\subfloat[State ]{%
\includegraphics[width=0.32\linewidth]{picture/pred2.pdf}%
\label{F2_pred}%
}\hfill
\subfloat[State ]{%
\includegraphics[width=0.32\linewidth]{picture/pred3.pdf}%
\label{F3_pred}%
}
\caption{Prediction trajectory overlay.}
\label{fig:pred_compare}
\end{figure*}
% in preamble:
% \usepackage{booktabs}
% \usepackage{multirow}
% \usepackage{booktabs} % 需要 booktabs
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.15}
\setlength{\tabcolsep}{8pt}
\caption{RMSE comparison of multi-step prediction for each state.}
\label{tab:pred_rmse_states}
\begin{tabular}{cccc}
\toprule
Methods & \multicolumn{3}{c}{RMSEs of each state} \
\cmidrule(lr){2-4}
& State & State & State \
\midrule
Proposed method & 0.0311 & 0.0798 & 0.1010 \
LSTM & 0.0376 & 0.0678 & 0.1312 \
MLP & 0.0418 & 0.0975 & 0.1368 \
\bottomrule
\end{tabular}
\end{table}
\subsection{Closed-loop tracking performance}
This subsection evaluates the closed-loop tracking performance of RNE-DMPC under (i) multiple sequential reference changes and (ii) single step changes in thickness and tension references.
In Figure~\ref{simulation}, three reference targets were modified sequentially during the simulation for the five-stand thickness--tension coupled system, and the resulting closed-loop responses were recorded.
It can be observed that the tension and thickness outputs of all stands were driven toward the updated references promptly, while overall stability was maintained.
At each switching instant, short-lived deviations of the tracking errors were induced due to the strong inter-stand coupling; however, these deviations were rapidly attenuated and kept at small levels afterwards.
Such behavior indicates that effective dynamic tracking and coordination were achieved under multivariable coupling and repeated setpoint switching.
Moreover, the contribution of the distributed architecture was reflected by the information exchange enabled by the communication mechanism, through which consistency across subsystems was enforced and coordinated regulation was maintained throughout the successive reference updates.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/simulation.pdf}
\caption{Tension--thickness tracking responses under multiple reference changes: (a) shows the tension tracking response; (b) shows the thickness tracking response.}\label{simulation}
\end{figure*}
In Figure~\ref{thickness}, a single step change was imposed on the thickness reference, and both the thickness trajectories and the tracking errors were evaluated.
Following the step, the thickness outputs were guided toward the new reference levels with a fast transition, and the tracking errors were suppressed quickly and maintained within a small range, implying strong transient regulation and high steady-state accuracy in the thickness channel.
In particular, after the responses reached steady state, the absolute tracking error of the outlet thickness (target value ) was , corresponding to a relative error of 2.28%.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/thickness.pdf}
\caption{Thickness responses and tracking errors under a step change of the thickness reference: (a) shows the thickness response; (b) shows the tracking error.}\label{thickness}
\end{figure*}
In Figure~\ref{tension}, a single step change was applied to the tension reference, and the corresponding tension responses and error evolutions were examined.
After the reference update, the tension outputs were pulled toward the new setpoints rapidly, and the errors were driven to small values without being accompanied by sustained oscillations, indicating satisfactory damping and closed-loop stability.
After steady state, the absolute tracking error of the tension (target value ) was , corresponding to a relative error of 1.91%.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/tension.pdf}
\caption{Tension responses and tracking errors under a step change of the tension reference: (a) shows the tension response; (b) shows the tracking error.}\label{tension}
\end{figure*}
\subsection{Impact of Nash coordination}
To highlight the role of Nash-equilibrium coordination in strongly coupled distributed control, we compare the proposed RNE-DMPC against a distributed MPC baseline without Nash coordination\citep{camponogara2002distributed}.
Specifically, the baseline performs only a single local update per sampling instant while keeping neighbors' strategies fixed from the previous update, i.e., no online best-response iterations are carried out.
Figure~\ref{fig:five_subplots1} and Figure~\ref{fig:five_subplots2} present the comparison of tracking errors for thickness and tension, where the Nash-coordinated scheme yields consistently smaller errors and faster convergence.
For a compact quantitative summary, we report the mean steady-state absolute tracking errors averaged over all stands, and the peak absolute error.The average steady-state absolute tracking error is denoted as for thickness and for tension.
The steady-state window is defined from the first time the tracking error enters and remains within the band until the end of the simulation.
Specifically, is computed by time-averaging over the steady-state window and then taking the mean across the five stands (), while is computed similarly and then averaged across the four inter-stand tensions ().
The results are summarized in Table~\ref{tab:nash_metrics}, confirming that Nash coordination improves the tracking accuracy in the coupled multi-stand system.
\begin{figure*}[htbp]
\centering
% -------- Row 1: three subfigures --------
\subfloat[Stand 1]{%
\includegraphics[width=0.32\linewidth]{picture/1.1.pdf}%
\label{fig:five_a}%
}\hfill
\subfloat[Stand 2]{%
\includegraphics[width=0.32\linewidth]{picture/1.2.pdf}%
\label{fig:five_b}%
}\hfill
\subfloat[Stand 3]{%
\includegraphics[width=0.32\linewidth]{picture/1.3.pdf}%
\label{fig:five_c}%
}\[2mm]
% -------- Row 2: two subfigures (same size as row 1, centered) --------
\makebox[\linewidth][c]{%
\subfloat[Stand 4]{%
\includegraphics[width=0.32\linewidth]{picture/1.4.pdf}%
\label{fig:five_d}%
}\hspace{0.04\linewidth}%
\subfloat[Stand 5]{%
\includegraphics[width=0.32\linewidth]{picture/1.5.pdf}%
\label{fig:five_e}%
}%
}
\caption{Thickness tracking error comparison with and without Nash coordination.}
\label{fig:five_subplots1}
\end{figure*}
\begin{figure*}[htbp]
\centering
% -------- Row 1: three subfigures --------
\subfloat[Stand 1]{%
\includegraphics[width=0.32\linewidth]{picture/2.1.pdf}%
\label{fig:five_a}%
}\hfill
\subfloat[Stand 2]{%
\includegraphics[width=0.32\linewidth]{picture/2.2.pdf}%
\label{fig:five_b}%
}\hfill
\subfloat[Stand 3]{%
\includegraphics[width=0.32\linewidth]{picture/2.3.pdf}%
\label{fig:five_c}%
}\[2mm]
% -------- Row 2: two subfigures (same size as row 1, centered) --------
\makebox[\linewidth][c]{%
\subfloat[Stand 4]{%
\includegraphics[width=0.32\linewidth]{picture/2.4.pdf}%
\label{fig:five_d}%
}\hspace{0.04\linewidth}%
\subfloat[Stand 5]{%
\includegraphics[width=0.32\linewidth]{picture/2.5.pdf}%
\label{fig:five_e}%
}%
}
\caption{Tension tracking error comparison with and without Nash coordination.}
\label{fig:five_subplots2}
\end{figure*}
% Preamble:
% \usepackage{booktabs}
% Preamble:
% \usepackage{booktabs}
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.15}
\setlength{\tabcolsep}{10pt}
\caption{Steady-state tracking accuracy comparison.}
\label{tab:nash_metrics}
\begin{tabular}{lcccc}
\toprule
\textbf{Method} &
\multicolumn{2}{c}{\textbf{Thickness error}} &
\multicolumn{2}{c}{\textbf{Tension error}} \
\cmidrule(lr){2-3}\cmidrule(lr){4-5}
& & (mm)
& & (kN) \
\midrule
RNE-DMPC & 2.9% & 0.0411 & 3.5% & 0.0714 \
Distributed MPC & 3.2% & 0.0503 & 4.1% & 0.0706 \
\bottomrule
\end{tabular}
\end{table}
\subsection{Computation rate comparison}
Through thickness and tension simulations, the computation time of each control cycle was recorded.
Figure~\ref{rate} compares the computation time of the proposed RNE-DMPC with a Nash-equilibrium-based distributed predictive control baseline~\citep{hu2021nash}.
Statistical results show that the per-cycle computation time of both Nash-type distributed controllers remains small, which is suitable for online receding-horizon implementation.
More importantly, the proposed RNE-DMPC consistently yields a noticeably lower per-cycle computation time than the Nash baseline in Ref.~\citep{hu2021nash}, indicating faster equilibrium seeking and coordination in online operation.
Overall, the results indicate that RNE-DMPC offers excellent computational performance and can effectively reduce system load and implementation cost.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/rate.pdf}
\caption{Comparison of calculation rates for different control strategies: (a) step disturbance in outlet thickness; (b) step disturbance in tension.}\label{rate}
\end{figure*}
\section{Conclusion}
This paper proposes a distributed model predictive control strategy based on residual neural networks and Nash equilibrium (RNE-DMPC) to achieve coordinated and high-precision control of thickness and tension in the cold rolling process. By using a residual neural network to model the input--output behavior of the system in an increment form, the interpretability of the learned dynamics is improved compared with a purely black-box predictor. Further, by combining the distributed model predictive control method based on Nash equilibrium, the constraint optimization and coordinated control of multi-variable coupled systems have been achieved, thereby alleviating the problem of slow convergence speed and improving the calculation speed. Numerical simulation results show that the proposed strategy exhibits excellent comprehensive performance in thickness and tension control, and has practical significance for improving the quality and production efficiency of cold-rolled steel strips. Future work can further consider shortening the system stability time in actual noisy environments and enhancing the control robustness of the system.
\section*{Disclosure statement}
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
\section*{Funding}
An unnumbered section, e.g.\ \verb"\section*{Funding}", may be used for grant details, etc.\ if required and included \emph{in the non-anonymous version} before any Notes or References.
\section*{Notes on contributors}
\noindent
\begin{minipage}[t]{0.28\textwidth}
\vspace{0pt}
\includegraphics[width=\linewidth]{picture/guochenhao.jpg} % ←务必写清扩展名
\end{minipage}\hspace{0.03\textwidth}
\begin{minipage}[t]{0.69\textwidth}
\vspace{0pt}
\textit{Chenhao Guo} is pursuing the B.E. degree at Northeastern University, Shenyang, China. His research interests include the control of multi-agent systems, safety control, and adaptive control.
\end{minipage}
\par\medskip
\noindent
\begin{minipage}[t]{0.28\textwidth}
\vspace{0pt}
\includegraphics[width=\linewidth]{picture/niejiahao.jpg} % ←务必写清扩展名
\end{minipage}\hspace{0.03\textwidth}
\begin{minipage}[t]{0.69\textwidth}
\vspace{0pt}
\textit{Jiahao Nie} is currently pursuing the B.S. degree at the College of Information Science and Engineering, Northeastern University, Shenyang, China. His research interests include intelligent sensing and gas detection.
\end{minipage}
\par\medskip
\noindent
\begin{minipage}[t]{0.28\textwidth}
\vspace{0pt}
\includegraphics[width=\linewidth]{} % ←务必写清扩展名
\end{minipage}\hspace{0.03\textwidth}
\begin{minipage}[t]{0.69\textwidth}
\vspace{0pt}
\textit{Chenhao Guo} is pursuing a bachelor’s degree at North-eastern University,
Shenyang, China. His research interests include the control of multi-agent systems,
safety control, and adaptive control.
\end{minipage}
\par\medskip
\bibliographystyle{apacite}
\bibliography{interactapasample}
\end{document}
这是我的最新版,帮我再检查检查 看看有没有措辞和格式上的问题
% interactapasample.tex
% v1.05 - August 2017
\documentclass[]{interact}
% 让你仍然可以用 \citep / \citet 这类 natbib 命令
\usepackage[natbibapa,nodoi]{apacite}
\setlength\bibhang{12pt}
\renewcommand\bibliographytypesize{\fontsize{10}{12}\selectfont}
\usepackage{epstopdf}% To incorporate .eps illustrations using PDFLaTeX, etc.
\usepackage[caption=false]{subfig}% Support for small, sub' figures and tables %\usepackage[nolists,tablesfirst]{endfloat}% To separate' figures and tables from text if required
%\usepackage[doublespacing]{setspace}% To produce a `double spaced' document if required
%\setlength\parindent{24pt}% To increase paragraph indentation when line spacing is doubled
\usepackage{booktabs}
\usepackage[numbers,longnamesfirst,sort,authoryear,round]{natbib}
\bibpunct[, ]{(}{)}{;}{a}{,}{,}% Citation support using natbib.sty
\usepackage{tabularx}
\usepackage{array}
\usepackage{amsmath,amssymb}
\usepackage{graphicx}
\usepackage[caption=false,font=footnotesize]{subfig}
\usepackage{wrapfig}
\usepackage{multirow}
\usepackage[final]{graphicx} % 关键:即使 class 里是 draft,也强制显示图片
\bibpunct[, ]{(}{)}{;}{a}{,}{,}% Citation support using natbib.sty
\renewcommand\bibfont{\fontsize{10}{12}\selectfont}% To set the list of references in 10 point font using natbib.sty
%\usepackage[natbibapa,nodoi]{apacite}% Citation support using apacite.sty. Commands using natbib.sty MUST be deactivated first!
%\setlength\bibhang{12pt}% To set the indentation in the list of references using apacite.sty. Commands using natbib.sty MUST be deactivated first!
%\renewcommand\bibliographytypesize{\fontsize{10}{12}\selectfont}% To set the list of references in 10 point font using apacite.sty. Commands using natbib.sty MUST be deactivated first!
\theoremstyle{plain}% Theorem-like structures provided by amsthm.sty
\newtheorem{theorem}{Theorem}[section]
\newtheorem{lemma}[theorem]{Lemma}
\newtheorem{corollary}[theorem]{Corollary}
\newtheorem{proposition}[theorem]{Proposition}
\theoremstyle{definition}
\newtheorem{definition}[theorem]{Definition}
\newtheorem{example}[theorem]{Example}
\theoremstyle{remark}
\newtheorem{remark}{Remark}
\newtheorem{notation}{Notation}
\begin{document}
\articletype{ARTICLE TEMPLATE}% Specify the article type or omit as appropriate
\title{Distributed MPC Control for Tandem Cold Rolling Based on Residual Network and Nash Equilibrium}
\author{
\name{Chenhao Guo\textsuperscript{a}, Jiahao Nie\textsuperscript{a}, Xiaojian Li\textsuperscript{a}\textsuperscript{b}\thanks{CONTACT Xiaojian Li. Email: [email protected]}}
\affil{\textsuperscript{a}College of Information Science and Engineering, Northeastern University, Shenyang, China; \textsuperscript{b}State Key Laboratory of Synthetical Automation for Process
Industries, Northeastern University, Shenyang, China}
}
\maketitle
\begin{abstract}
With the rapid development of multi-stand tandem cold rolling technology, high-precision and high-efficiency thickness–tension regulation has become increasingly important for product quality and process stability. However, the multi-stand cold rolling system has a complex structure. Not only is it difficult to establish the mechanism model, but there are also coupling relationships among the stands. This coupling relationships have led to the problem of optimization conflicts among multiple stands. Given that the complexity and coupling of the system, this study proposes a distributed model predictive control strategy based on residual neural network (ResNet) and Nash equilibrium (RNE-DMPC). A ResNet is trained to capture the input-output behaviors of the system, avoiding reliance on an explicit mechanism model. For online constrained multivariable optimization, a Nash-equilibrium-based distributed MPC is developed to coordinate multi-objective tension–thickness regulation across stands. The performance of the proposed method is verified by a series numerical simulations.
\end{abstract}
\begin{keywords}
Residual neural network; Nash equilibrium; Distributed model predictive control; Tandem cold rolling; Tension control; Thickness control
\end{keywords}
\section{Introduction}
Cold-rolled strip steel is a crucial intermediate product in the strip-steel manufacturing chain and significantly affects downstream process performance and final product quality \citep{mancini2012effects}. It is commonly produced by tandem cold rolling mills, where the strip passes sequentially through multiple stands and is progressively reduced in thickness by rolling forces. In such mills, regulating the exit thickness and inter-stand tensions is essential for product quality and process stability. Figure~\ref{AutoReg} shows a typical layout of a five-stand tandem cold rolling mill.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/x3.pdf}
\caption{Equipment layout of a five-stand tandem cold rolling mill.}\label{AutoReg}
\end{figure*}
From a control perspective, thickness--tension regulation is a large-scale multivariable problem featuring strong coupling, nonlinearities, constraints, time-varying parameters, and disturbances. These characteristics complicate reliable prediction of rolling dynamics and make high-performance closed-loop control difficult to sustain over changing operating conditions.
PID control remains prevalent in practice due to its simplicity, yet fixed-parameter tuning often struggles in strongly coupled multivariable rolling processes \citep{gupta2025grid,liu2025data,hasilci2025superior}. Model predictive control (MPC) provides a systematic framework to handle multivariable coupling and constraints and has been widely investigated in rolling applications \citep{yaghini2024linear,luis2019trajectory,han2021hierarchical,song2025model,sun2017dynamic}. Depending on the system scale and information interaction, MPC implementations can be centralized, decentralized, or distributed \citep{song2025model}.
In particular, distributed MPC decomposes the large-scale optimization into local problems solved in parallel with limited information exchange, improving scalability and reducing computational burden for interconnected systems \citep{scattolini2009architectures}. Multi-stand tandem cold rolling is naturally an interconnected large-scale system, making distributed optimization particularly suitable. \citep{zhao2025multi,ji2022coordinate,wang2025distributed}.
However, most MPC implementations still rely on mechanistic prediction models, which typically require extensive parameter identification and calibration. In tandem cold rolling, the rolling dynamics vary with steel grade, friction, rolling speed, and thermal conditions. Such operating-condition changes often induce model mismatch, leading to degraded prediction accuracy and consequently undermining closed-loop MPC performance.
With the increasing availability of industrial data, data-driven methods have attracted growing interest in rolling processes. Recent progress in machine learning has enabled effective mapping from measured process variables to key quality-related outputs, providing an alternative to repeatedly re-identifying mechanistic parameters under varying regimes \citep{zong2025machine}.
In particular, neural networks have been widely explored due to their strong nonlinear approximation capability and flexible modeling of coupled multivariable dynamics \citep{xu2024novel,cheng2015neural,xu2022control,han2019data,feng2025model}. At present, Niu et al. have developed a distributed nonlinear model predictive control based on RBF neural network for tandem cold rolling process \citep{niu2025rbf}; Huang et al. proposed the LSTM-AGD multi-mode adaptive MPC framework, which reduces overshoot and improves accuracy in a steady state \citep{huang2022lstm}; Xia et al. used BFGS to optimize the ANN topology and data-driven high-precision prediction of key parameters in cold continuous rolling \citep{xia2023using}.
Nevertheless, when neural predictors are embedded into receding-horizon optimization, two practical issues become critical. First, multi-step prediction is typically obtained by recursive rollout, where small one-step errors may accumulate and cause drift over the horizon, especially under strong inter-stand coupling. Second, purely black-box models may lack a structure consistent with the incremental evolution of rolling dynamics, which can reduce robustness and reliability when operating conditions change. These challenges motivate the development of a control-dependent residual neural surrogate with improved long-horizon robustness for distributed MPC in multi-stand cold rolling.
Therefore, to address these issues, this paper integrates the ResNet into MPC for prediction. This data-based ResNet can model dynamic evolution processes,aiming to improve prediction accuracy and interpretability for thickness and tension regulation.
In multi-stand rolling, inter-stand tension propagation introduces strong coupling among local control objectives: actions at one stand affect neighboring tensions and thickness responses, leading to potential strategy conflicts among local MPC controllers. Game-theoretic coordination provides a principled way to resolve such conflicts. Nash-equilibrium-based distributed predictive control has been studied for thickness--tension coordination in rolling mills \citep{hu2021nash,ye2021adaptive}. However, the effectiveness and convergence of online equilibrium seeking can be sensitive to prediction accuracy, especially when operating conditions vary and modeling errors increase.Although some studies have introduced Nash equilibrium into the distributed predictive control of thickness tension, this iterative process is highly dependent on the accuracy of the prediction model and the shared information. Once the on-site operating conditions change and lead to an increase in prediction errors, the equilibrium iterative effect increased prediction error may slow convergence or lead to termination at an approximate equilibrium, thereby degrading coordination quality and constraint satisfaction.
Therefore, this paper proposes a distributed model predictive control strategy based on residual neural networks and Nash equilibrium (RNE-DMPC).
The main contributions of this article are as follows:
\begin{enumerate}
\item A residual neural network model is developed to capture the input--output relationship, aiming to improve prediction accuracy while enhancing model interpretability.
\item For strongly coupled multi-stand thickness--tension coordination, we propose a Nash-equilibrium-based distributed MPC (RNE-DMPC). The trained residual network is embedded as the local predictor, and Nash coordination is achieved online via consistency penalty and a relaxed best-response iteration.
\end{enumerate}
The rest of this article is organized as follows: Section 2 describes the construction of the dataset. Section 3 builds the residual neural network. Section 4 discusses the specific content of RNE-DMPC. Section 5 presents the simulation results of the thickness and tension of the five-stand system. Section 6 presents the conclusion of the full text.
%========================
\section{Construction of Dataset}
%========================
Consider a five-stand tandem cold rolling mill indexed by .
Let denote the -th sampling instant and the corresponding sampling interval length.
Define the local within-interval time variable .
Let denote the exit thickness of stand (),
and let denote the inter-stand strip tension between stand and ().
The system is strongly coupled because the inter-stand tensions propagate along the mill line and are affected by neighboring stands' actions.
Let and be the desired references (setpoints) given by process requirements. We define deviation variables
\begin{equation}
\Delta h_i(t)\triangleq h_i(t)-h_i^{\mathrm{ref}}(t),\qquad
\Delta T_i(t)\triangleq T_i(t)-T_i^{\mathrm{ref}}(t).
\label{eq:dev_def}
\end{equation}
Throughout this paper, the symbol ``'' attached to states always means deviation from reference.
For each stand , we choose the local deviation state as
\begin{equation}
\Delta x_i(t)\triangleq
\begin{bmatrix}
\Delta h_i(t)\
\Delta T_{i-1}(t)\
\Delta T_i(t)
\end{bmatrix}\in\mathbb{R}^{d},\qquad d=3,
\label{eq:xi_def_clean}
\end{equation}
with the boundary convention and to keep a unified dimension for all stands.
For a five-stand tandem mill, the dominant coupling is between adjacent stands, hence we define
\begin{equation}
Z_1={2},\quad
Z_i={i-1,i+1}\ (i=2,3,4),\quad
Z_5={4}.
\label{eq:Zi_clean}
\end{equation}
Define the neighbor-state stack
\begin{equation}
\Delta x_{Z_i}(t_n)=\mathrm{col}{\Delta x_k(t_n),|,k\in Z_i}.
\label{eq:xZi_clean}
\end{equation}
\begin{bmatrix}
\Delta s_i(t_n)\
\Delta v_i(t_n)
\end{bmatrix}.
\label{eq:du_discrete_clean}
\end{equation}
Throughout this paper, the symbol ``'' attached to inputs means sample-to-sample increment.
Thus, and are conceptually different, and this is fixed by definition.
Let denote exogenous disturbances.
We denote the interval-level equivalent disturbance by .
denotes the identity matrix; denotes the zero matrix.
M_d,\Delta x_i(t_n)
+
N_d,\Delta u_i(t_n)
+
F_d,\Delta d_i(t_n),
\label{eq:linear_form_concept}
\end{equation}
where represent equivalent discrete-time matrices around operating conditions.
In a practical five-stand cold rolling mill, accurately deriving and identifying these matrices and disturbance models from first principles is difficult,
due to strong coupling, unmodeled nonlinearities, and time-varying operating regimes.
Therefore, this paper aims to learn a high-fidelity approximation of the interval evolution from data and then embed it into distributed MPC.
\begin{remark}
In fact, due to the existence of complex coupling relationships, it is difficult to directly and accurately establish \eqref{eq:linear_form_concept}
based on first principles. Therefore, in this paper, we learn an approximate mapping of \eqref{eq:true_mapping_clean} from data.
\end{remark}
Although decisions are updated at discrete instants , the hydraulic gap and drive systems evolve continuously inside each interval,
and abrupt within-interval changes may excite tension oscillations and deteriorate thickness stability.
Thus, parameterizing the within-interval increment trajectory by a low-order polynomial:
(i) yields a compact finite-dimensional decision representation;
(ii) enforces smooth profiles inside the interval;
(iii) enables enforcing increment constraints for all .
This is appropriate when is not excessively large relative to actuator bandwidth and the within-interval evolution is well approximated by a low-order basis.
\sum_{m=0}^{n}\Gamma_{i,nm}\tau^{m},
\qquad \tau\in[0,\delta_n],
\label{eq:du_poly_vec_clean}
\end{equation}
where are coefficient vectors (), and is the chosen polynomial order.
In simulations, one may select (quadratic) for a good accuracy--complexity trade-off.
Component-wise, \eqref{eq:du_poly_vec_clean} corresponds to
\begin{equation}
\begin{aligned}
\Delta s_{i,n}(\tau) &= \sum_{m=0}^{n}\gamma^{(s)}{i,nm}\tau^{m},\
\Delta v{i,n}(\tau) &= \sum_{m=0}^{n}\gamma^{(v)}{i,nm}\tau^{m}.
\end{aligned}
\label{eq:du_components_clean}
\end{equation}
Define the stacked parameter vector
\begin{equation}
\Gamma{i,n}\triangleq
\big[
(\Gamma_{i,n0})^\top,,
(\Gamma_{i,n1})^\top,,
\ldots,,
(\Gamma_{i,nn})^\top
\big]^\top
\in\mathbb{R}^{p},
\qquad
p=(n+1)n_u.
\label{eq:Gamma_clean}
\end{equation}
Here, is the baseline increment at , while () describe the higher-order variation rates.
Define the interval-averaged equivalent increments as
\begin{equation}
\begin{aligned}
\Delta u_i(t_n) &\triangleq \frac{1}{\delta_n}\int_0^{\delta_n}\Delta u_{i,n}(\tau),d\tau,\
\Delta d_i(t_n) &\triangleq \frac{1}{\delta_n}\int_0^{\delta_n}\Delta d_i(\tau),d\tau.
\end{aligned}
\label{eq:avg_def_clean}
\end{equation}
With \eqref{eq:du_poly_vec_clean}, the input average has a closed form:
\begin{equation}
\Delta u_i(t_n)=
\sum_{m=0}^{n}\Gamma_{i,nm}\frac{\delta_n^{m}}{m+1}.
\label{eq:avg_closed_clean}
\end{equation}
Let denote the sampling domain of deviation states and neighbor stacks ,
and let denote the sampling domain of polynomial parameters .
These domains specify the operating envelope used to generate supervised training data.
Given the above parameterization, one training sample is generated on each interval .
In addition to the local deviation state, the neighbor deviation states are included to represent inter-stand coupling.
The process is summarized in Table~\ref{tab:interval_sample_generation_en}.
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.15}
\caption{Procedure for generating one interval-level sample on (five-stand coupled mill).}
\label{tab:interval_sample_generation_en}
\begin{tabularx}{\linewidth}{>{\centering\arraybackslash}p{0.09\linewidth} X}
\toprule
\textbf{Step} & \textbf{Operation} \
\midrule
1 & \textbf{State sampling:} sample and from . \
2 & \textbf{Parameter sampling:} draw (coefficients for both and ). \
3 & \textbf{Control construction:} compute via \eqref{eq:du_poly_vec_clean}. \
4 & \textbf{State propagation:} integrate the \emph{five-stand coupled} mill model on (e.g., RK4) using the within-interval control trajectory, and record . \
\bottomrule
\end{tabularx}
\end{table}
Accordingly, an interval sample for subsystem can be represented as
\begin{equation}
\mathcal{D}{i,n}=\big{\Delta x_i(t_n),\ \Delta x{Z_i}(t_n),\ \Delta u_{i,n}(\tau),\ \Delta x_i(t_{n+1})\big}.
\label{eq:interval_sample_clean}
\end{equation}
Note that is fully determined by via \eqref{eq:du_poly_vec_clean},
therefore it is sufficient to store as the learning input.
For each subsystem , by repeating the above procedure across multiple intervals and randomized draws,
the local one-step training dataset is formed as
\begin{equation}
\begin{split}
S_i=\Big{&
\big(\Delta x_i^{(j)}(t_n),,\Delta x_{Z_i}^{(j)}(t_n),,\Delta x_i^{(j)}(t_{n+1});,
\Gamma_{i,n}^{(j)},,\delta_n^{(j)}\big)
\ \Big|\ j=1,\ldots,J
\Big}.
\end{split}
\label{eq:S_i_clean}
\end{equation}
Here is the number of one-step samples for subsystem .
The overall dataset for the five-stand mill is denoted by .
The point-cloud visualization of the training dataset is shown in Figure~\ref{2}.
\begin{figure*}[htbp]
\centering
\includegraphics[scale=0.5]{picture/Fig2.pdf}
\caption{Point cloud map of the training dataset.}\label{2}
\end{figure*}
The one-step set is sufficient for one-step regression, but it is not sufficient for training with multi-step rollout loss
and reciprocal-consistency regularization, because these objectives require ground-truth deviation-state trajectories over a horizon of consecutive intervals.
Therefore, without changing the single-interval sampling mechanism above, we additionally organize the offline-simulated samples
into -step trajectory segments.
Specifically, for each starting time we generate a segment of length by consecutively sampling
(and the corresponding inputs/disturbances),
and integrating the five-stand coupled mill model over for .
Hence, we obtain the deviation-state sequence as well as the neighbor stacks
.
Define a -step segment sample for subsystem as
\begin{equation}
\begin{aligned}
\mathcal{W}{i,n}=
\Big{&
\big(\Delta x_i(t{n+s}),,\Delta x_{Z_i}(t_{n+s}),,\Gamma_{i,n+s},,\delta_{n+s}\big){s=0}^{K-1}; \
&\big(\Delta x_i(t{n+s+1})\big){s=0}^{K-1}
\Big}.
\end{aligned}
\label{eq:segment_clean}
\end{equation}
By repeating the above segment generation, we form the multi-step training set
\begin{equation}
S_i^{(K)}=\Big{\mathcal{W}{i,n}^{(j)}\ \Big|\ j=1,\ldots,J_K\Big},
\label{eq:S_i_K_clean}
\end{equation}
where is the number of -step segment samples.
Note that can be viewed as the marginal one-step projection of (keeping only ),
thus the original dataset design is preserved, and only an additional \emph{segment organization} is introduced for multi-step training.
%========================
\section{Construction of Residual Neural Network}
%========================
\subsection{Residual Neural Network Structure Construction and Training Method}
Given the dataset, the neural network model is trained to learn a stand-wise, control-dependent one-step evolution law of deviation states:
\begin{equation}
\Delta x_i(t_{n+1})
\approx
\Delta x_i(t_n)+
\mathcal{N}i!\Big(\Delta x_i(t_n),,\Delta x{Z_i}(t_n),,\Gamma_{i,n},,\delta_n;,\Theta_i\Big),
\label{eq:learned_dyn_clean}
\end{equation}
where outputs the one-step deviation-state change and are trainable parameters.
\begin{remark}
If does not take control information as input (here and ),
the predictor becomes an autoregressive model that only reproduces trajectories under the training input patterns
and cannot answer the counterfactual question: ``what will happen if we choose a different roll gap and speed trajectory?''
Since MPC optimizes over candidate decisions, a control-dependent predictor \eqref{eq:learned_dyn_clean} is necessary
to evaluate the predicted thickness and tension behavior under different candidate actuator trajectories.
\end{remark}
Let (state dimension), be the number of neighbors of stand in \eqref{eq:Zi_clean}, and in \eqref{eq:Gamma_clean}.
Define the input vector
\begin{equation}
X_{i,\text{in}} \triangleq
\big[
\Delta x_i(t_n)^\top,,
\Delta x_{Z_i}(t_n)^\top,,
\Gamma_{i,n}^\top,,
\delta_n
\big]^\top
\in \mathbb{R}^{d(1+|Z_i|)+p+1}.
\label{eq:X_in_clean}
\end{equation}
The network mapping is
\begin{equation}
\mathcal{N}_i:\mathbb{R}^{d(1+|Z_i|)+p+1}\rightarrow\mathbb{R}^{d}.
\end{equation}
To improve training stability and long-horizon rollout robustness, we use a residual form.
Let be a selection matrix extracting the local state block:
\begin{equation}
\hat{I}i = [I_d,, 0{d\times(d|Z_i|+p+1)}].
\label{eq:Ihat_clean}
\end{equation}
Then the one-step predictor is written as
\begin{equation}
X_{i,\text{out}} = \hat{I}i X{i,\text{in}} + \mathcal{N}i(X{i,\text{in}}; \Theta_i),
\label{eq:res_predict_clean}
\end{equation}
where represents the predicted .
This structure implements a baseline-plus-correction interpretation:
the shortcut propagates the current deviation state , while the network learns the correction capturing
unmodeled nonlinearities and inter-stand coupling (via ) under varying operating conditions.
To improve robustness when varies, we introduce an auxiliary branch inside :
\begin{equation}
\mathcal{N}i(X{i,\text{in}};\Theta_i)\triangleq
\psi_i(X_{i,\text{in}};\Theta_{\psi_i}) + \rho_i(X_{i,\text{in}};\theta_i),
\label{eq:aux_clean}
\end{equation}
where is a lightweight feedforward branch that captures low-frequency/scale effects strongly related to ,
and captures the remaining nonlinear coupling corrections.
When , the model reduces to a standard residual network.
For the -th sample in \eqref{eq:S_i_clean}, define
\begin{equation}
X_{i,\text{in}}^{(j)} =
\big[
\Delta x_i^{(j)}(t_n),\ \Delta x_{Z_i}^{(j)}(t_n),\
\Gamma_{i,n}^{(j)},\ \delta_n^{(j)}
\big]^{\top},
\end{equation}
and the supervised residual target
\begin{equation}
\Delta r_i^{(j)}=\Delta x_i^{(j)}(t_{n+1})-\Delta x_i^{(j)}(t_n).
\label{eq:target_clean}
\end{equation}
To suppress accumulation drift induced by long-horizon recursion and to improve long-term predictive stability,
we train the forward predictor jointly with an auxiliary backward residual model
and impose a multi-step reciprocal-consistency regularization over a -step segment from .
Construct a backward residual network
\begin{equation}
\mathcal{B}i:\mathbb{R}^{d(1+|Z_i|)+p+1}\rightarrow\mathbb{R}^{d},
\end{equation}
parameterized by . For the backward step associated with interval , define
\begin{equation}
\begin{aligned}
X{i,\mathrm{in}}^{b}
&=
\big[
\Delta x_i(t_{n+1}),\ \Delta x_{Z_i}(t_{n+1}),
\Gamma_{i,n},\ \delta_n
\big]^{\top},\
X_{i,\mathrm{out}}^{b}
&=
\hat{I}i X{i,\mathrm{in}}^{b} + \mathcal{B}i(X{i,\mathrm{in}}^{b};\bar{\Theta}i),
\end{aligned}
\label{eq:back_clean}
\end{equation}
where represents the backward estimate of .
The supervised backward residual target is
\begin{equation}
\Delta r_i^{b}=\Delta x_i(t_n)-\Delta x_i(t{n+1}).
\end{equation}
Given a segment sample , initialize
\begin{equation}
\Delta \hat{x}i(t_n)=\Delta x_i(t_n),
\end{equation}
and recursively apply the forward predictor for steps:
\begin{equation}
\begin{aligned}
\Delta \hat{x}i(t{n+s+1})
&=
\Delta \hat{x}i(t{n+s})
+
\mathcal{N}i!\Big(
\Delta \hat{x}i(t{n+s}),,\Delta \hat{x}{Z_i}(t{n+s}),,
\Gamma_{i,n+s},,\delta_{n+s};,\Theta_i
\Big),\
&\qquad s=0,\ldots,K-1.
\end{aligned}
\label{eq:fwd_roll_clean}
\end{equation}
Set the terminal condition
\begin{equation}
\Delta \bar{x}i(t{n+K})=\Delta \hat{x}i(t{n+K}),
\end{equation}
and roll back using :
\begin{equation}
\begin{aligned}
\Delta \bar{x}i(t{n+s})
&=
\hat{I}i X{i,\mathrm{in}}^{b}(t_{n+s})
+
\mathcal{B}i!\Big(X{i,\mathrm{in}}^{b}(t_{n+s});,\bar{\Theta}i\Big),
\quad s=K-1,\ldots,0,
\end{aligned}
\label{eq:bwd_roll_clean}
\end{equation}
where
\begin{equation}
X{i,\mathrm{in}}^{b}(t_{n+s})=
\big[
\Delta \bar{x}i(t{n+s+1}),\ \Delta \hat{x}{Z_i}(t{n+s+1}),
\Gamma_{i,n+s},\ \delta_{n+s}
\big]^{\top}.
\end{equation}
\sum_{s=0}^{K}
\left|
\Delta \hat{x}i(t{n+s})-\Delta \bar{x}i(t{n+s})
\right|^2.
\end{equation}
We jointly minimize:
\begin{equation}
\begin{aligned}
L_{\mathrm{1step}}(\Theta_i)
&= \frac{1}{J_K}\sum_{j=1}^{J_K}\frac{1}{K}\sum_{s=0}^{K-1}
\Big|
\big(\Delta x_i^{(j)}(t_{n+s+1})-\Delta x_i^{(j)}(t_{n+s})\big)
-\mathcal{N}i!\left(
X{i,\mathrm{in}}^{(j)}(t_{n+s});\Theta_i
\right)
\Big|^2,\[2mm]
L_{\mathrm{bwd}}(\bar{\Theta}i)
&= \frac{1}{J_K}\sum{j=1}^{J_K}\frac{1}{K}\sum_{s=0}^{K-1}
\Big|
\big(\Delta x_i^{(j)}(t_{n+s})-\Delta x_i^{(j)}(t_{n+s+1})\big)
-\mathcal{B}i!\left(
X{i,\mathrm{in}}^{b,(j)}(t_{n+s});\bar{\Theta}i
\right)
\Big|^2,\[2mm]
L{\mathrm{msrp}}(\Theta_i,\bar{\Theta}i)
&= \frac{1}{J_K}\sum{j=1}^{J_K} E_i^{(j)}(t_n),\[2mm]
L_{\mathrm{roll}}(\Theta_i)
&= \frac{1}{J_K}\sum_{j=1}^{J_K}\sum_{s=1}^{K}
\Big|
\Delta x_i^{(j)}(t_{n+s})-\Delta \hat{x}i^{(j)}(t{n+s})
\Big|^2.
\end{aligned}
\label{eq:loss_clean}
\end{equation}
Here, enforces one-step accuracy; explicitly suppresses long-horizon drift under recursion;
regularizes the learned dynamics by enforcing reciprocal consistency between forward and backward rollouts;
and trains the backward model for the consistency regularization.
In implementation, these terms are combined as
\begin{equation}
L_{\mathrm{total}}=\lambda_1 L_{\mathrm{1step}}+\lambda_2 L_{\mathrm{roll}}+\lambda_3 L_{\mathrm{msrp}}+\lambda_4 L_{\mathrm{bwd}},
\end{equation}
where are tuned on a validation set.
\Delta x_i(t_n)
+
\mathcal{N}i!\Big(
\Delta x_i(t_n),,\Delta x{Z_i}(t_n),,
\Gamma_{i,n},,\delta_n;,\Theta_i^*
\Big),
\label{eq:pred_clean}
\end{equation}
and multi-step prediction is obtained by recursive rollout of \eqref{eq:pred_clean}.
This learned predictor is the internal model used by the MPC optimizer in the next section.
Finally, network parameters are optimized using Adam:
\begin{equation}
\Theta_{i,t+1} = \Theta_{i,t} - \alpha \frac{\hat{g}{i,t}}{\sqrt{\hat{r}{i,t}} + \varepsilon},
\end{equation}
where is the learning rate (we use to avoid conflict with other symbols),
and are bias-corrected moment estimates, and is a small constant for numerical stability.
Figure~\ref{fig:rnn_logic} illustrates the overall structure.
\begin{figure}[htbp]
\centering
\includegraphics[scale=0.85]{picture/x6.pdf}
\caption{Logic diagram of the residual neural network.}
\label{fig:rnn_logic}
\end{figure}
\subsection{Explainability of the residual network}
\Delta x_i(t_n)
+
\underbrace{\int_{0}^{\delta_n}
f_i!\Big(\Delta x_i(t_n+\tau),,\Delta x_{Z_i}(t_n+\tau),,u_i(t_n+\tau),,d_i(t_n+\tau)\Big),d\tau}{\triangleq\ \varphi{i,n}},
\label{eq:increment_integral}
\end{equation}
where is the one-interval state increment generated by the mechanistic dynamics.
Our learned model \eqref{eq:learned_dyn_clean} adopts the same increment form as \eqref{eq:increment_integral}:
\begin{equation}
\Delta x_i(t_{n+1})
\approx
\Delta x_i(t_n)
+
\mathcal{N}i!\Big(\Delta x_i(t_n),,\Delta x{Z_i}(t_n),,\Gamma_{i,n},,\delta_n;\Theta_i\Big).
\end{equation}
Here, plays the role of a data-driven approximation of the integral increment ,
i.e., it approximates the accumulated effect of the mechanistic dynamics over .
This is consistent with the well-known interpretation that a residual network behaves like a one-step time integrator:
the identity path propagates the current state, while the residual branch represents the increment over the time lag.
Inside each interval, we do not optimize point-wise but parameterize the increment trajectory by
.
Hence, the mechanistic increment in \eqref{eq:increment_integral} depends on the \emph{whole} within-interval trajectory.
Feeding into is therefore a compact way to represent how different candidate
gap/speed trajectories change the integral effect and thus the next thickness--tension state.
When varies and is not very small, directly learning may be harder.
Motivated by the generalized residual idea,
we decompose the increment predictor into two parts in \eqref{eq:aux_clean}:
\begin{equation}
\mathcal{N}_i(\cdot)=\psi_i(\cdot)+\rho_i(\cdot).
\end{equation}
Conceptually, captures low-frequency and scale effects strongly related to ,
while captures the remaining nonlinear coupling corrections.
This provides a mechanism-consistent interpretation: a baseline increment plus a residual correction
that compensates unmodeled nonlinearities and inter-stand coupling.
%========================
\section{Nash Equilibrium-Based RNE-DMPC}
%========================
The five-stand tandem cold rolling system is strongly coupled through inter-stand tension propagation.
As a result, changes in control actions (roll gap and stand speed) at one stand can affect both upstream and downstream stands,
making centralized online optimization over all stands' decision variables computationally demanding.
To mitigate this issue, we decompose the global predictive-control problem into local subproblems associated with individual stands.
Each local controller optimizes its own decision variables while accounting for coupling via limited information exchange with neighboring controllers.
Motivated by game-theoretic coordination \citep{rawlings2008coordinating}, we formulate distributed coordination as a Nash-equilibrium-seeking iteration.
Based on the trained residual neural network surrogate model, we construct a Nash-equilibrium-based distributed MPC method (RNE-DMPC)
for coordinated thickness--tension regulation and tracking. The overall control structure is shown in Figure~\ref{4}.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/x2.pdf}
\caption{Schematic diagram of the control architecture for a tandem cold rolling mill.}\label{4}
\end{figure*}
At sampling time , stand chooses the polynomial-parameter sequence
, where .
Let
denote the joint strategy profile, and let denote the collection of all strategies except stand .
Given the current measured/estimated deviation state and the strategies
,
the multi-step prediction used by stand is written explicitly as
\begin{equation}
\begin{aligned}
\Delta \hat{x}i(t{n+s+1};\mathbf{\Gamma}i,\mathbf{\Gamma}{Z_i})
&=
\Delta \hat{x}i(t{n+s};\mathbf{\Gamma}i,\mathbf{\Gamma}{Z_i})
+
\mathcal{N}i!\Big(
\Delta \hat{x}i(t{n+s};\cdot),,
\Delta \hat{x}{Z_i}(t_{n+s};\mathbf{\Gamma}{Z_i}),\
&\qquad
\Gamma{i,n+s},,
\delta_{n+s};\Theta_i^*
\Big),
\end{aligned}
\label{eq:rollout_mpc_game}
\end{equation}
for , with initialization .
Here the neighbor stack is generated from neighbors' strategies via the same learned predictors.
\sum_{m=0}^{n}\Gamma_{i,n+s,m}\frac{\delta_{n+s}^{m}}{m+1}.
\label{eq:du_avg_clean}
\end{equation}
\begin{remark}
Because inter-stand tension is jointly affected by the adjacent stands and ,
the predicted evolution of depends on neighbors' future actions,
hence the MPC problems are not independent but form a coupled dynamic game.
\end{remark}
\mathrm{col}{\Gamma_{i,n},\Gamma_{i,n+1},\ldots,\Gamma_{i,n+N_c-1}}
\in \mathbb{R}^{pN_c}.
\end{equation}
In deviation coordinates, the regulation/tracking objective is , i.e.
\begin{equation}
\Delta x_{i,\mathrm{ref}}(t_{n+s})\equiv 0\in\mathbb{R}^{d},\qquad d=3.
\end{equation}
Recall .
Define the row selectors
\begin{equation}
C^- \triangleq [0\ \ 1\ \ 0]\in\mathbb{R}^{1\times 3},\qquad
C^+ \triangleq [0\ \ 0\ \ 1]\in\mathbb{R}^{1\times 3},
\end{equation}
so that (upstream interface) and (downstream interface).
C^-\Delta \hat{x}{i+1}(t{n+s};\mathbf{\Gamma}),\qquad i=1,\ldots,4.
\label{eq:shared_tension_mismatch}
\end{equation}
\sum_{s=1}^{N_p}
\left|
\Delta \hat{x}i(t{n+s};\mathbf{\Gamma}i,\mathbf{\Gamma}{Z_i})
\right|{Q_i}^{2}
+
\sum{s=0}^{N_c-1}
\left|\Gamma_{i,n+s}\right|_{R_i}^{2}
+
J_i^{\mathrm{cpl}}(\mathbf{\Gamma}i;\mathbf{\Gamma}{-i})
\label{eq:Ji_game}
\end{equation}
where weights thickness and tension deviations, and penalizes actuation magnitudes.
C^-\Delta \hat{x}{i+1}(t{n+s};\mathbf{\Gamma})
\big|^2
\Big),
\end{aligned}
\label{eq:coupling_cost}
\end{equation}
with and indicator .
This term makes the coupling conflict explicit: unilateral actions that locally reduce thickness error may worsen shared-tension
compatibility and thus increase , and also affect neighbors' objectives.
We enforce the absolute-input bounds and within-interval increment bounds.
Absolute input bounds (roll gap and speed):
\begin{equation}
u_{i,\min}\le u_i(t_{n+s})\le u_{i,\max},
\qquad s=0,\ldots,N_p-1,
\label{eq:u_abs_game}
\end{equation}
where .
Within-interval increment-trajectory bounds:
\begin{equation}
\Delta u_{i,\min}\le \Delta u_{i,n+s}(\tau;\Gamma_{i,n+s})\le \Delta u_{i,\max},
\qquad \forall\tau\in[0,\delta_{n+s}],\ s=0,\ldots,N_p-1,
\label{eq:du_traj_game}
\end{equation}
where is given by the polynomial parameterization.
\sum_{m=0}^{n}\Gamma_{i,n+s,m}\frac{\delta_{n+s}^{m}}{m+1},
\label{eq:du_avg_game}
\end{equation}
and propagate the absolute input along the horizon by
\begin{equation}
u_i(t_n)=u_i(t_{n-1})+\Delta u_i(t_n),\qquad
u_i(t_{n+s})=u_i(t_{n+s-1})+\Delta u_i(t_{n+s}),\ s=1,\ldots,N_p-1,
\label{eq:u_prop_game}
\end{equation}
where is the applied (measured) input from the previous sampling instant.
Compact feasible set:
\begin{equation}
\Omega_i \triangleq
\Big{\mathbf{\Gamma}_i\ \Big|\
\eqref{eq:rollout_mpc_game}\ \text{holds and}
\eqref{eq:u_abs_game},\eqref{eq:du_traj_game},\eqref{eq:u_prop_game}\ \text{are satisfied}
\Big}.
\label{eq:Omega_i_game}
\end{equation}
\arg\min_{\mathbf{\Gamma}_i\in\Omega_i}\
J_i(\mathbf{\Gamma}i;\mathbf{\Gamma}{-i}).
\label{eq:local_BR}
\end{equation}
Because the learned surrogate is differentiable, \eqref{eq:local_BR} can be solved by standard gradient-based NLP solvers.
At each sampling time , the distributed MPC coordination induces a finite-horizon dynamic game:
players are stands ; strategy sets are ; and payoff (cost) functions are
defined in \eqref{eq:Ji_game}--\eqref{eq:coupling_cost}.
A joint strategy profile
is a Nash equilibrium if
\begin{equation}
\forall i\in{1,\ldots,5},\qquad
\mathbf{\Gamma}i^*\in
\arg\min{\mathbf{\Gamma}_i\in\Omega_i}
J_i(\mathbf{\Gamma}i;\mathbf{\Gamma}{-i}^*).
\label{eq:NE_def}
\end{equation}
This definition explicitly characterizes the strategic coupling:
each player's optimal decision depends on neighbors' decisions through the shared-tension dynamics and the coupling term.
(1-\omega)\mathbf{\Gamma}_i^{(l-1)}
+
\omega,\mathbf{\Gamma}_i^{\mathrm{BR},(l)},
\qquad \omega\in(0,1].
\label{eq:relaxed_BR}
\end{equation}
The relaxation factor mitigates oscillations caused by strong coupling and improves practical convergence.
The Nash equilibrium is computed through distributed best-response iterations, summarized in Table ~\ref{tab:nash_iter_en}.
\max_i
\frac{\left|
\mathbf{\Gamma}_i^{(l)}-\mathbf{\Gamma}_i^{(l-1)}
\right|_2}{
\left|
\mathbf{\Gamma}_i^{(l-1)}
\right|_2+\epsilon},
\end{equation}
with small.
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.12}
\setlength{\tabcolsep}{3.5pt}
\caption{Relaxed distributed Nash best-response iteration for RNE-DMPC (five-stand).}
\label{tab:nash_iter_en}
\begin{tabularx}{\linewidth}{>{\centering\arraybackslash}p{0.11\linewidth} X}
\toprule
\textbf{Step} & \textbf{Description} \
\midrule
A &
Initialize and warm-start (e.g., from the previous sampling time). \
B &
Communicate (or the induced predicted trajectories) among neighbors; form . \
C &
Given , solve the best-response NLP \eqref{eq:local_BR} to obtain . \
D &
Update the relaxed strategy using \eqref{eq:relaxed_BR} and compute the induced predictions
via \eqref{eq:rollout_mpc_game}. \
E &
Broadcast and the predicted interface quantities needed in \eqref{eq:coupling_cost}
(e.g., ) to neighbors. \
F &
Compute ; if , stop and set ; otherwise and repeat. \
\bottomrule
\end{tabularx}
\end{table}
\sum_{m=0}^{n}\Gamma_{i,n,m}^*\frac{\delta_n^{m}}{m+1},
\end{equation}
followed by .
The overall control flow chart is shown in Fig~\ref{liu}.
\begin{figure}
\centering
\includegraphics[width=\linewidth]{picture/x5.pdf}
\caption{The overall system control flow chart.}\label{liu}
\end{figure}
\begin{remark}
Equations \eqref{eq:Ji_game}--\eqref{eq:coupling_cost} and the NE definition \eqref{eq:NE_def} explicitly show that each player's objective
depends on others' strategies through the coupled prediction dynamics and the shared-tension interface mismatch penalty,
thus making the multi-stand strategic interaction and coordination mechanism transparent.
\end{remark}
\section{Simulations}
To verify the effectiveness of the proposed residual-neural-network and Nash-equilibrium based distributed MPC (RNE-DMPC) for coordinated thickness--tension regulation in a five-stand cold tandem rolling system, a distributed control architecture composed of multiple local controllers and a communication/coordination module is considered. Each stand is treated as a subsystem, and information exchange through the communication system enforces coordination among subsystems, enabling distributed thickness--tension control for the interconnected plant.
In this study, the simulation data were generated using the mathematical model reported by Hu et al.~\citep{hu2021nash}.
\sqrt{\frac{1}{N}\sum_{i=1}^{N}\left(x_i-\hat{x}_i\right)^2 }.
\end{equation}
where is the number of samples, and are the actual and predicted values, respectively.
\subsection{Surrogate-model prediction comparison}
In this paper, all of the spatial dynamics network and temporal evolution networks are constructed
based on the ResNet architecture and trained using the loss function in \eqref{eq:loss_clean}. Specifically, one-block ResNet method is chosen, with 5 hidden layers in each block, and each layer contains 160 neurons. And the optimizer is chosen as Adam with betas equal (0.9,0.99), initial learning rate is 0.001 and will decay to 0.00001 gradually. All networks are trained in the PyTorch
environment for 500 epochs. Meanwhile, the multi-order parameterization in the simulation part uses the second-order approach.
To justify the use of the proposed residual predictor in MPC, we first compare its prediction accuracy against representative baseline models, including
a plain feedforward network (MLP) with comparable model size trained by one-step regression\citep{Hornik1989MLPUniversal}, and a sequence model (LSTM) trained for trajectory prediction\citep{Hochreiter1997LSTM}.
All models are trained and tested under the same dataset split.
Figure~\ref{fig:pred_compare} shows an illustrative trajectory overlay on a representative test case.
It can be observed that the proposed residual network produces a closer match to the ground truth and exhibits less drift in multi-step rollouts.
The aggregated results can be summarized in Table~\ref{tab:pred_rmse_states}, where the proposed residual network achieves the lowest prediction errors among the compared models.
\begin{figure*}[htbp]
\centering
\subfloat[State ]{%
\includegraphics[width=0.32\linewidth]{picture/pred1.pdf}%
\label{F1_pred}%
}\hfill
\subfloat[State ]{%
\includegraphics[width=0.32\linewidth]{picture/pred2.pdf}%
\label{F2_pred}%
}\hfill
\subfloat[State ]{%
\includegraphics[width=0.32\linewidth]{picture/pred3.pdf}%
\label{F3_pred}%
}
\caption{Prediction trajectory overlay.}
\label{fig:pred_compare}
\end{figure*}
% in preamble:
% \usepackage{booktabs}
% \usepackage{multirow}
% \usepackage{booktabs} % 需要 booktabs
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.15}
\setlength{\tabcolsep}{8pt}
\caption{RMSE comparison of multi-step prediction for each state.}
\label{tab:pred_rmse_states}
\begin{tabular}{cccc}
\toprule
Methods & \multicolumn{3}{c}{RMSEs of each state} \
\cmidrule(lr){2-4}
& State & State & State \
\midrule
Proposed method & 0.0311 & 0.0798 & 0.1010 \
LSTM & 0.0376 & 0.0678 & 0.1312 \
MLP & 0.0418 & 0.0975 & 0.1368 \
\bottomrule
\end{tabular}
\end{table}
\subsection{Closed-loop tracking performance}
This subsection evaluates the closed-loop tracking performance of RNE-DMPC under (i) multiple sequential reference changes and (ii) single step changes in thickness and tension references.
In Figure~\ref{simulation}, three reference targets were modified sequentially during the simulation for the five-stand thickness--tension coupled system, and the resulting closed-loop responses were recorded.
It can be observed that the tension and thickness outputs of all stands were driven toward the updated references promptly, while overall stability was maintained.
At each switching instant, short-lived deviations of the tracking errors were induced due to the strong inter-stand coupling; however, these deviations were rapidly attenuated and kept at small levels afterwards.
Such behavior indicates that effective dynamic tracking and coordination were achieved under multivariable coupling and repeated setpoint switching.
Moreover, the contribution of the distributed architecture was reflected by the information exchange enabled by the communication mechanism, through which consistency across subsystems was enforced and coordinated regulation was maintained throughout the successive reference updates.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/simulation.pdf}
\caption{Tension--thickness tracking responses under multiple reference changes: (a) shows the tension tracking response; (b) shows the thickness tracking response.}\label{simulation}
\end{figure*}
In Figure~\ref{thickness}, a single step change was imposed on the thickness reference, and both the thickness trajectories and the tracking errors were evaluated.
Following the step, the thickness outputs were guided toward the new reference levels with a fast transition, and the tracking errors were suppressed quickly and maintained within a small range, implying strong transient regulation and high steady-state accuracy in the thickness channel.
In particular, after the responses reached steady state, the absolute tracking error of the outlet thickness (target value ) was , corresponding to a relative error of 2.28%.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/thickness.pdf}
\caption{Thickness responses and tracking errors under a step change of the thickness reference: (a) shows the thickness response; (b) shows the tracking error.}\label{thickness}
\end{figure*}
In Figure~\ref{tension}, a single step change was applied to the tension reference, and the corresponding tension responses and error evolutions were examined.
After the reference update, the tension outputs were pulled toward the new setpoints rapidly, and the errors were driven to small values without being accompanied by sustained oscillations, indicating satisfactory damping and closed-loop stability.
After steady state, the absolute tracking error of the tension (target value ) was , corresponding to a relative error of 1.91%.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/tension.pdf}
\caption{Tension responses and tracking errors under a step change of the tension reference: (a) shows the tension response; (b) shows the tracking error.}\label{tension}
\end{figure*}
\subsection{Impact of Nash coordination}
To highlight the role of Nash-equilibrium coordination in strongly coupled distributed control, we compare the proposed RNE-DMPC against a distributed MPC baseline without Nash coordination\citep{camponogara2002distributed}.
Specifically, the baseline performs only a single local update per sampling instant while keeping neighbors' strategies fixed from the previous update, i.e., no online best-response iterations are carried out.
Figure~\ref{fig:five_subplots1} and Figure~\ref{fig:five_subplots2} present the comparison of tracking errors for thickness and tension, where the Nash-coordinated scheme yields consistently smaller errors and faster convergence.
For a compact quantitative summary, we report the mean steady-state absolute tracking errors averaged over all stands, and the peak absolute error.The average steady-state absolute tracking error is denoted as for thickness and for tension.
The steady-state window is defined from the first time the tracking error enters and remains within the band until the end of the simulation.
Specifically, is computed by time-averaging over the steady-state window and then taking the mean across the five stands (), while is computed similarly and then averaged across the four inter-stand tensions ().
The results are summarized in Table~\ref{tab:nash_metrics}, confirming that Nash coordination improves the tracking accuracy in the coupled multi-stand system.
\begin{figure*}[htbp]
\centering
% -------- Row 1: three subfigures --------
\subfloat[Stand 1]{%
\includegraphics[width=0.32\linewidth]{picture/1.1.pdf}%
\label{fig:five_a}%
}\hfill
\subfloat[Stand 2]{%
\includegraphics[width=0.32\linewidth]{picture/1.2.pdf}%
\label{fig:five_b}%
}\hfill
\subfloat[Stand 3]{%
\includegraphics[width=0.32\linewidth]{picture/1.3.pdf}%
\label{fig:five_c}%
}\[2mm]
% -------- Row 2: two subfigures (same size as row 1, centered) --------
\makebox[\linewidth][c]{%
\subfloat[Stand 4]{%
\includegraphics[width=0.32\linewidth]{picture/1.4.pdf}%
\label{fig:five_d}%
}\hspace{0.04\linewidth}%
\subfloat[Stand 5]{%
\includegraphics[width=0.32\linewidth]{picture/1.5.pdf}%
\label{fig:five_e}%
}%
}
\caption{Thickness tracking error comparison with and without Nash coordination.}
\label{fig:five_subplots1}
\end{figure*}
\begin{figure*}[htbp]
\centering
% -------- Row 1: three subfigures --------
\subfloat[Stand 1]{%
\includegraphics[width=0.32\linewidth]{picture/2.1.pdf}%
\label{fig:five_a}%
}\hfill
\subfloat[Stand 2]{%
\includegraphics[width=0.32\linewidth]{picture/2.2.pdf}%
\label{fig:five_b}%
}\hfill
\subfloat[Stand 3]{%
\includegraphics[width=0.32\linewidth]{picture/2.3.pdf}%
\label{fig:five_c}%
}\[2mm]
% -------- Row 2: two subfigures (same size as row 1, centered) --------
\makebox[\linewidth][c]{%
\subfloat[Stand 4]{%
\includegraphics[width=0.32\linewidth]{picture/2.4.pdf}%
\label{fig:five_d}%
}\hspace{0.04\linewidth}%
\subfloat[Stand 5]{%
\includegraphics[width=0.32\linewidth]{picture/2.5.pdf}%
\label{fig:five_e}%
}%
}
\caption{Tension tracking error comparison with and without Nash coordination.}
\label{fig:five_subplots2}
\end{figure*}
% Preamble:
% \usepackage{booktabs}
% Preamble:
% \usepackage{booktabs}
\begin{table}[t]
\centering
\small
\renewcommand{\arraystretch}{1.15}
\setlength{\tabcolsep}{10pt}
\caption{Steady-state tracking accuracy comparison.}
\label{tab:nash_metrics}
\begin{tabular}{lcccc}
\toprule
\textbf{Method} &
\multicolumn{2}{c}{\textbf{Thickness error}} &
\multicolumn{2}{c}{\textbf{Tension error}} \
\cmidrule(lr){2-3}\cmidrule(lr){4-5}
& & (mm)
& & (kN) \
\midrule
RNE-DMPC & 2.9% & 0.0411 & 3.5% & 0.0714 \
Distributed MPC & 3.2% & 0.0503 & 4.1% & 0.0706 \
\bottomrule
\end{tabular}
\end{table}
\subsection{Computation rate comparison}
Through thickness and tension simulations, the computation time of each control cycle was recorded.
Figure~\ref{rate} compares the computation time of the proposed RNE-DMPC with a Nash-equilibrium-based distributed predictive control baseline~\citep{hu2021nash}.
Statistical results show that the per-cycle computation time of both Nash-type distributed controllers remains small, which is suitable for online receding-horizon implementation.
More importantly, the proposed RNE-DMPC consistently yields a noticeably lower per-cycle computation time than the Nash baseline in Ref.~\citep{hu2021nash}, indicating faster equilibrium seeking and coordination in online operation.
Overall, the results indicate that RNE-DMPC offers excellent computational performance and can effectively reduce system load and implementation cost.
\begin{figure*}[htbp]
\centering
\includegraphics[width=\linewidth]{picture/rate.pdf}
\caption{Comparison of calculation rates for different control strategies: (a) step disturbance in outlet thickness; (b) step disturbance in tension.}\label{rate}
\end{figure*}
\section{Conclusion}
This paper proposes a distributed model predictive control strategy based on residual neural networks and Nash equilibrium (RNE-DMPC) to achieve coordinated and high-precision control of thickness and tension in the cold rolling process. By using a residual neural network to model the input--output behavior of the system in an increment form, the interpretability of the learned dynamics is improved compared with a purely black-box predictor. Further, by combining the distributed model predictive control method based on Nash equilibrium, the constraint optimization and coordinated control of multi-variable coupled systems have been achieved, thereby alleviating the problem of slow convergence speed and improving the calculation speed. Numerical simulation results show that the proposed strategy exhibits excellent comprehensive performance in thickness and tension control, and has practical significance for improving the quality and production efficiency of cold-rolled steel strips. Future work can further consider shortening the system stability time in actual noisy environments and enhancing the control robustness of the system.
\section*{Disclosure statement}
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
\section*{Funding}
An unnumbered section, e.g.\ \verb"\section*{Funding}", may be used for grant details, etc.\ if required and included \emph{in the non-anonymous version} before any Notes or References.
\section*{Notes on contributors}
\noindent
\begin{minipage}[t]{0.28\textwidth}
\vspace{0pt}
\includegraphics[width=\linewidth]{picture/guochenhao.jpg} % ←务必写清扩展名
\end{minipage}\hspace{0.03\textwidth}
\begin{minipage}[t]{0.69\textwidth}
\vspace{0pt}
\textit{Chenhao Guo} is pursuing the B.E. degree at Northeastern University, Shenyang, China. His research interests include the control of multi-agent systems, safety control, and adaptive control.
\end{minipage}
\par\medskip
\noindent
\begin{minipage}[t]{0.28\textwidth}
\vspace{0pt}
\includegraphics[width=\linewidth]{picture/niejiahao.jpg} % ←务必写清扩展名
\end{minipage}\hspace{0.03\textwidth}
\begin{minipage}[t]{0.69\textwidth}
\vspace{0pt}
\textit{Jiahao Nie} is currently pursuing the B.S. degree at the College of Information Science and Engineering, Northeastern University, Shenyang, China. His research interests include intelligent sensing and gas detection.
\end{minipage}
\par\medskip
\noindent
\begin{minipage}[t]{0.28\textwidth}
\vspace{0pt}
\includegraphics[width=\linewidth]{} % ←务必写清扩展名
\end{minipage}\hspace{0.03\textwidth}
\begin{minipage}[t]{0.69\textwidth}
\vspace{0pt}
\textit{Chenhao Guo} is pursuing a bachelor’s degree at North-eastern University,
Shenyang, China. His research interests include the control of multi-agent systems,
safety control, and adaptive control.
\end{minipage}
\par\medskip
\bibliographystyle{apacite}
\bibliography{interactapasample}
\end{document}
这是我的最新版,帮我再检查检查 看看有没有措辞和格式上的问题
我通读了你这份最新版(从导言区到结论/作者简介),下面按“会导致编译问题/警告”和“措辞不够地道/需要润色”两大类,把你文中具体位置指出来,并给出可直接替换的改法。
apacite 和 natbib 同时用 + 重复设置 \bibpunct你现在同时加载了:
latex\usepackage[natbibapa,nodoi]{apacite} ... \usepackage[numbers,longnamesfirst,sort,authoryear,round]{natbib} \bibpunct... ... \bibpunct...
问题:
apacite(带 natbibapa)本身就让你能用 \citep/\citet,通常不需要再加载 natbib。natbib 的参数里还同时写了 numbers 和 authoryear(互相冲突)。sort,authoryear 这个逗号是中文全角逗号,,会直接导致参数解析异常(很常见的编译报错来源)。\bibpunct 重复两次。✅ 建议(最稳、最少改动):保留 apacite,删掉 natbib 与 bibpunct 两行
把下面这些删掉即可:
latex\usepackage[numbers,longnamesfirst,sort,authoryear,round]{natbib} \bibpunct[, ]{(}{)}{;}{a}{,}{,} ... \bibpunct[, ]{(}{)}{;}{a}{,}{,} \renewcommand\bibfont{...}
如果你确实想用 natbib 控制标点/格式,那就反过来:保留 natbib,别用 apacite,但你现在 \bibliographystyle{apacite} 又是 apacite 流派,所以更推荐“保留 apacite,删除 natbib”。
subfig、graphicx 重复加载你现在有:
latex\usepackage[caption=false]{subfig} ... \usepackage[caption=false,font=footnotesize]{subfig}
以及:
latex\usepackage{graphicx} ... \usepackage[final]{graphicx}
✅ 只保留一份即可。比如保留这两行就够了:
latex\usepackage{graphicx} \usepackage[caption=false,font=footnotesize]{subfig}
\theoremstyle 但没显式加载 amsthm \theoremstyle{plain} 等命令来自 amsthm。有的 class 会自带,但不一定。
✅ 建议加一行(保险):
latex\usepackage{amsthm}
你在“厚度误差对比”和“张力误差对比”两张图里都用了同一组 label:
latex\label{fig:five_a} \label{fig:five_b} \label{fig:five_c} ...
这会触发 Label multiply defined,并且后面引用会跳到错误位置。
✅ 改法:第二张图的 label 全部换一套,例如:
fig:thick_s1 ... fig:thick_s5fig:tension_s1 ... fig:tension_s5\label{2} 这种不推荐(虽不一定报错)你有:
latex\caption{Point cloud map ...}\label{2}
能用,但可读性差。
✅ 建议改成:
latex\label{fig:dataset_cloud}
你写了:
latex\includegraphics[width=\linewidth]{}
这会直接报错(找不到文件名)。
✅ 方案二选一:
picture/lixiaojian.jpg另外你第三段简介把 Chenhao Guo 又写了一遍,应该是 Xiaojian Li(而且 North-eastern 建议统一 Northeastern)。
你摘要目前有几处典型英文问题:
This coupling relationships have led ...These coupling relationships have led ... 或更地道 This strong coupling leads to ...establish the mechanism modelestablish an accurate mechanistic model / derive a physics-based modelGiven that the complexity and coupling of the system, ...Given the complexity and coupling of the system, ...verified by a series numerical simulationsverified by a series of numerical simulationsresidual neural network (ResNet),下一句又 A ResNet...a residual neural network (ResNet-style),后面统一用 residual network 或 residual neural network。(把你表里 RMSE 和你文中跟踪误差塞进去了)
latex\begin{abstract} High-precision thickness--tension regulation in multi-stand tandem cold rolling is challenging due to strong inter-stand coupling and the difficulty of maintaining accurate mechanistic models under varying operating conditions. To address this issue, this paper proposes a residual-neural-network and Nash-equilibrium integrated distributed model predictive control scheme (RNE-DMPC). A residual neural network (ResNet-style) is trained as a control-dependent surrogate predictor for the stand-wise deviation dynamics, avoiding reliance on an explicit mechanistic model. For online constrained coordination, a Nash-equilibrium-based distributed MPC is developed using shared-tension consistency penalties and a relaxed best-response iteration. Numerical simulations on a five-stand benchmark show improved multi-step prediction accuracy (e.g., RMSE of state $x_1$: 0.0311 vs. 0.0376 (LSTM) and 0.0418 (MLP)) and high-accuracy tracking with steady-state errors of 0.029~mm (2.28\%) in thickness and 0.021~kN (1.91\%) in tension. \end{abstract}
原句:
latex... making distributed optimization particularly suitable. \citep{zhao2025multi,ji2022coordinate,wang2025distributed}.
✅ 改成:
latex... making distributed optimization particularly suitable \citep{zhao2025multi,ji2022coordinate,wang2025distributed}.
原句:
latexprocesses,aiming to ...
✅ 改成:
latexprocesses, aiming to ...
原句:
latex... errors increase.Although some studies ...
✅ 改成:
latex... errors increase. Although some studies ...
原句:
... the equilibrium iterative effect increased prediction error may slow convergence or lead to termination ...
✅ 建议替换为更清晰的一句(保留你原意):
latexOnce operating conditions change, increased prediction errors may slow the equilibrium iteration or result in convergence to an approximate equilibrium, thereby degrading coordination quality and constraint satisfaction.
你现在 contribution 1 仍然写了:
latex... enhancing model interpretability.
如果你想更“审稿友好”,建议改成更保守、与你后面“increment form”一致的说法:
✅ 替换成:
latex... aiming to improve prediction accuracy and to improve interpretability via an increment-form modeling structure.
贡献 2 里这句:
latex... via consistency penalty ...
建议补冠词并具体一点(你正文是 shared-tension consistency penalty):
✅ 替换成:
latex... via a shared-tension consistency penalty and a relaxed best-response iteration.
--(thickness--tension)。建议全文统一用 --(最保险,避免字体/编码问题)。Distributed MPC for Tandem Cold Rolling Based on a Residual Neural Network and Nash Equilibrium(可选)如果你愿意,我可以把你这份 tex **按“最小改动可编译、期刊友好”**的原则,给你一个“导言区精简版(不冲突)”的完整 preamble 模板,你直接替换顶部即可。只要你告诉我:你最终想用 apacite 还是 natbib(二选一)。