11544 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 70, NO.

11, NOVEMBER 2023

LSTM-MPC: A Deep Learning Based Predictive

Control Method for Multimode Process Control
Keke Huang , Member, IEEE, Ke Wei, Fanbiao Li, Chunhua Yang , Fellow, IEEE, and Weihua Gui

Abstract—Modern industrial processes often operate un- industrial processes are becoming large-scale, complicated, and
der different modes, which brings challenges to model pre- highly coupled, various operation modes would appear due to
dictive control (MPC). Recently, most MPC related methods different system parameters or structures [5], [6]. Take a zinc
would establish prediction models independently for dif-
ferent modes, which results in their control effect highly roasting process as an example, according to field workers’
relying on switching strategies. Inspired by the powerful experience, different production indexes like production load
representation capabilities of deep learning, this article pro- and market demands may lead to various operation modes and
posed a deep learning based MPC method. Specifically, bring multimode control problem [7], [8]. For each mode, indus-
the LSTM network is applied to predict behaviors of con- trial process has various operating parameters and if the control
trolled system, which can automatically match different op-
eration modes without switching strategy. Then combined strategy fails to match the corresponding mode, it will reduce
with MPC framework, an adaptive gradient descent method the stability of the system, and even cause serious production
is introduced to handle optimization problem and its con- accidents and losses. Therefore, how to improve MPC method
straints. In addition, stability and feasibility analysis have to solve multimode control problem is becoming a key research
been conducted from the aspect of theory to ensure prac- focus [9].
tical application of the proposed method. Experiments on
a numerical simulation process and an industrial process Currently, some improved MPC methods aiming to deal with
platform show the strength and reliability of the proposed multimode process are called multimodel MPC or switching
method, which reduces the overshoot by about 10% com- MPC [10], [11]. Their goal is to establish multiple models for
pared to common learning-based MPC methods and im- each operation mode and design model switching strategy. Li
proves the control accuracy effectively. et al. combined Takagi–Sugeno fuzzy models with multimodel
Index Terms—Deep learning, long short-term memory framework to achieve stable control for different operation
network, model predictive control (MPC), multimode pro- modes [12]. Tang et al. proposed a multimodel neural network
cess. based MPC for a nonlinear pH neutralization process [13]. As for
model switching strategy, Wan et al. designed a soft-switching
I. INTRODUCTION scheme for autonomous underwater vehicle to get better control
performance [14]. Although these methods have achieved some
ODEL predictive control (MPC) is an efficient control
M method that is widely applied in industrial processes.
Generally, the MPC includes three fundamental elements: the
success, they still have two drawbacks. First, since these methods
would establish multiple models for each operation mode, it
may cost more resources if the number of operation modes
predictive model of the dynamics of controlled system, the ref- increases. Second, their control effect highly depends on model
erence trajectory, and the optimal controller obtained by rolling switching strategy. In real industrial processes, it is difficult to
optimization [1], [2], [3], [4]. Although MPC has been proven distinguish every mode accurately due to complex environment,
efficient in industrial plants, it still has some limitations. Since which further degrades the control performance and robustness.
Since prediction model in MPC largely determines the final
Manuscript received 16 August 2022; revised 26 October 2022; ac- control effect [15], one way to overcome the abovementioned
cepted 4 December 2022. Date of publication 20 December 2022; date limitations is to study designing more accurate predictive models
of current version 8 May 2023. This work was supported in part by the for multimode process while reducing the dependence of control
National Natural Science Foundation of China under Grant 62073340
and Grant 61860206014, in part by the Major Key Project of Peng effect on switching strategies.
Cheng Laboratory under Grant PCL2021A09, in part by the National Deep learning, as a typical automatic feature extraction
Key R&D Program of China under Grant 2022YFB3304900 and Grant method, has attracted more and more attention recently. Gen-
2019YFB1705300, and in part by the Science and Technology Inno-
vation Program of Hunan Province under Grant 2022JJ10083, Grant erally, when modeling with Big Data containing complex fea-
2021RC3018, and Grant 2021RC4054. (Corresponding author: Chun- tures, deep learning performs better than shallow learning meth-
hua Yang.) ods [16]. Based on the powerful representation capability of
The authors are with the School of Automation, Central South
University, Changsha 410083, China (e-mail: [email protected]; deep learning, some pioneering works combine deep learning
[email protected]; [email protected]; [email protected]; gwh@ with MPC framework and come up with DeepMPC [17], [18].
csu.edu.cn). Lucia et al. used deep neural networks (DNN) to extract the
Color versions of one or more figures in this article are available at
https://doi.org/10.1109/TIE.2022.3229323. changing characteristics of resonant power converters and suc-
Digital Object Identifier 10.1109/TIE.2022.3229323 cessfully deployed the control method to industrial sites with

0278-0046 © 2022 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 08,2023 at 13:24:20 UTC from IEEE Xplore. Restrictions apply.
HUANG et al.: LSTM-MPC: A DEEP LEARNING BASED PREDICTIVE CONTROL METHOD 11545

Fig. 1. Closed-loop MPC system.

Fig. 2. Model mismatch in multimode process control.

FPGA [19]. For the unknown dynamic characteristics of some
nonlinear systems, Pan et al. chose recurrent neural network The rest of this article is organized as follows. A brief in-
(RNN) to analyze the sequential characteristics, which improves troduction of MPC and multimode control problem is formu-
the stability and control accuracy of systems [20]. These meth- lated in Section II. The detail of the proposed LSTM-MPC
ods demonstrate a promising prospect of deep learning in MPC method is presented in Section III. The stability and feasibil-
and bring a new solution for multimode control problem: it is ity of the proposed method are analyzed in Section IV. The
possible to establish a more accurate prediction model suitable analyzing and discussion of the experimental results are shown
with all operation modes so that no additional switching strate- in Section V. Section VI gives the concluding remarks of this
gies are required. article.
Motivated by the powerful representation ability of deep
learning and the demand of multimode control, a Long Short- II. PRELIMINARIES
Term Memory based MPC (LSTM-MPC) is proposed in this
article. Here, the LSTM network is selected as predictive model A. Model Predictive Control
for multimode process due to the following reasons. First, LSTM MPC is a multivariable control method, which applies the
network has a unique memory cell that can extract sequential mathematical or data-driven model to predict the future behav-
features for each mode [21], which was in accord with the ior of controlled system and computes a sequence of optimal
feedback control feature of MPC, thus it lays foundation for control inputs that satisfies given constraints. In general, MPC
accurate multimode prediction. Second, the gate mechanism in includes three fundamental elements: the predictive model, the
LSTM network makes it automatically match each operation reference trajectory, and the rolling optimal controller. A typical
mode [22], which means no additional model switching strategy structure of closed-loop MPC system is shown in Fig. 1, where
needs to be designed. In summary, the detail of the LSTM-MPC r(t) is the reference trajectory, u(t) stands for manipulated
method are as follows. First, the LSTM network is introduced variable, y(t) represents controlled variable, ŷ(t) is output of
to establish a predictive model that can predict behavior of predictive model, d is external disturbance, and e represents
controlled process in each operation modes accurately. Then, the error.
predictive model is combined with MPC framework to formulate
optimization problem and an adaptive gradient descent (GD) is
B. Multimode Control Problem
proposed to settle the problem. Finally, the proposed method
is proven stable and feasible by theoretical analysis. Several Since industrial process becomes coupled and has a large time
experiments demonstrate that the proposed method can improve delay as its scale increases gradually, different operation modes
the control effect. In summary, the main contributions of this would appear due to complex environment. Generally, each
article include threefolds: operation mode has a corresponding optimal control strategy.
1) A novel multimode control method is proposed. Com- If the control strategy fails to match its corresponding operation
pared to conventional multimodel methods, the proposed mode of industrial process, it will lead to model mismatch, which
method only requires establishing a single prediction can cause great fluctuation in industrial process as shown in
model, which reduces the computational cost to meet the Fig. 2.
real-time requirement of industrial process in practice. Currently, some improved MPC methods aiming to deal with
2) The proposed prediction model based on LSTM network multimode process are called multimodel MPC or switching
can predict controlled system’s behavior in each operation MPC. These methods would establish multiple models for each
mode, which means no additional model switching strat- operation mode and design model switching strategy. Their
egy needs to be designed since the proposed prediction framework is shown in Fig. 3. Switching MPC for multimode
model can automatically match each operation mode. control is usually composed of three parts: bank of predictive
3) In order to obtain optimal controller, an adaptive GD models, bank of controllers, and decision unit. In detail, its
method is proposed to settle optimization problem and working process is demonstrated as follows: First, a series of
improve control effect. Meanwhile, theoretical analysis predictive models need to be established to track the behaviors of
is conducted to guarantee the stability and feasibility of controlled system under different operation modes. Then, design
the proposed method. a switching strategy in decision unit to distinguish different

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 08,2023 at 13:24:20 UTC from IEEE Xplore. Restrictions apply.
11546 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 70, NO. 11, NOVEMBER 2023

Fig. 4. Architecture of LSTM network.

Generally, the above assumptions are satisfied in most prac-

tical industrial process. Due to complex environment, industrial
process is often coupled and has a large time delay, which
Fig. 3. Framework of switching control.
makes Assumption 1 reasonable [23]. As for Assumption 2, it
can be explained by analyzing how control action works in real
operation modes. Based on these works, the optimal controller industrial processes. The execution unit of control action is usu-
can be obtained by solving problem to ensure the stability of ally a valve located at different nodes. Through changing valve
entire process for each operation mode. opening, the controlled variables can achieve given reference.
Although switching MPC control methods have achieved For most valves, their opening range is fixed once they have
success to some extent, they still have some limitations. First, been produced, which makes Assumption 2 acceptable. Based on
they need to build predictive models for each mode, respec- these assumptions, this article tries to find a new way to deal with
tively. If the number of operation modes increases, it will multimode control problem. Specifically, the proposed method
cost more computational resources. Second, these methods is called LSTM-MPC, which applies LSTM network to predict
highly rely on the effect of switching strategy. In real in- the behaviors of controlled system in different operation modes
dustrial process, it is difficult to distinguish every opera- and introduces adaptive GD method to obtain optimal control
tion mode accurately due to complex environment. Therefore, strategies.
these limitations motivated us to improve multimode control
methods.
B. LSTM Network Based Predictive Model
Compared to shallow machine learning or traditional DNN,
III. PROPOSED METHOD
long short-term memory (LSTM) network has superior predic-
A. Problem Description tion capability since it introduces gating mechanism and mem-
In this article, the controlled system can be formulated as ory blocks [22], [24], which can not only store historical state
follows: of system, but also adaptively controls the flow of information
between each block. The architecture of LSTM network is shown
y(t + 1) = p(U p (t), Y p (t)) (1) in Fig. 4.
In each time step, LSTM network has three input: input
where U p (t) = [u(t), u(t − 1), . . ., u(t − lu )], Y p (t) = [y(t), sequence xt , previous cell state C(t − 1), previous output of
y(t − 1), . . . , y(t − ly )], ly and lu represent lags of system’s LSTM memory block h(t − 1); and two output: current cell state
output and input, and p(·) is unknown dynamics function of the C(t) and current output of LSTM memory block h(t). During
system. Generally, due to complex environments and various the training stage, LSTM network can gradually learn the map-
operation modes, it is hard to find an accurate p(·) to represent ping between the historical input sequence xt = [xt 1 , xt 2 , . . .]
the behavior of controlled system. Therefore, the goal of the pro- and the predicted output sequence yt = [yt 1 , yt 2 , . . .]. Specifi-
posed method is to accurately predict the behavior of controlled cally, the input gate determines how much incoming information
system and obtain optimal control strategies in different modes. will be stored in the memory block
In order to study multimode control better, two assumptions are
introduced as follows: it = δ (Wi · [ht−1 , xt ] + bi ) (2)
Assumption 1: The controlled system is a slow change pro-
cess, which means the operation conditions will gradually switch Ĉt = tanh (Wc · [ht−1 , xt ] + bc ) . (3)
from one to another.
Assumption 2: The range of manipulated variable u(t) re- Then, the forget gate determines, which information will be left
mains constant in each mode, that is, ui ∈ [Umin,i , Umax,i ], i = out
1, 2, . . ., k, where Umin,i and Umax,i are constant, i represents
the ith mode. ft = δ (Wf · [ht−1 , xt ] + bf ) . (4)

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 08,2023 at 13:24:20 UTC from IEEE Xplore. Restrictions apply.
HUANG et al.: LSTM-MPC: A DEEP LEARNING BASED PREDICTIVE CONTROL METHOD 11547

Through these two gates, the state of memory cell can be updated optimization problem can be formulated as follows:
as follows:
min J(t) = min (R(t) − Y (t))T a(R(t) − Y (t))
U (t) U (t)
Ct = ft  Ct−1 + it  Ĉt . (5)
+ΔU (t)T bΔU (t)
Finally, the output gate is used to filter the information to obtain subject to
ŷ(t) = g(U p (t), Y p (t)) (10)
the network output yt
|Δu(t)| ≤ Δumax
ot = δ (Wo · [ht−1 , xt ] + bo ) (6) umin ≤ u(t) ≤ umax
ŷmin ≤ ŷ(t) ≤ ŷmax
ht = ot  tanh (Ct ) (7)
where a and b are weight parameters, the historical input
yt = φ (Wy ht + by ) (8) and output are U p (t) = [u(t), u(t − 1), . . ., u(t − lu )],
and Y p = [y(t), y(t − 1), . . ., y(t − ly )], R(t) = [r(t + 1),
where Ct is the cell state, Ĉt denotes the updated new state, ht r(t + 2), . . ., r(t + Tp )] is reference output, Ŷ (t) =
represents the hidden layer state, Wi , Wf , Wc , Wo , Wy denote [ŷ(t + 1), ŷ(t + 2), . . ., ŷ(t + Tp )] is predictive output,
weight matrices, and bi , bf , bc , bo , by stands for bias vectors. U (t) = [u(t), u(t + 1), . . . , u(t + Tc − 1)] is optimal control
These parameters can be confirmed by backpropagation mecha- input, ΔU (t) = [Δu(t), Δu(t + 1), . . . , Δu(t + Tc − 1)] is
nism during LSTM network training. Besides, δ(·) and tanh(·) incremental control moves, and g(·) represents predictive model
represent sigmoid function and Hyperbolic tangent function, based on LSTM network whose expression can be represented
respectively,  is the elementwise product of the vectors and below according to (6)–(8):
φ is the network output activation function.
Motivated by the strong predictive ability of LSTM, the g(U p (t), Y p (t)) = g(xt )
 
prediction model in proposed control method was estab- 1 eCt −e−Ct
lished through the following steps. First, design the loss func- = Wy  + by . (11)
1+e−(Wo [ht−1 ,xt ]+bo ) eCt +e−Ct
tion. Since LSTM network applies multistep states to pre-
dict system output, its loss function can be formulated as Since the proposed predictive method is based on deep learn-
follows: ing, its mathematical expression cannot be represented directly,
 2 which brings challenge to settle optimization problem (10).
 
L = R(t) − Ŷ (t) (9) In order to reduce the number of iterations and find optimal
2
result quickly, GD method is chosen to solve the optimization
where R(t) = [r(t + 1), r(t + 2), . . ., r(t + Tp )] is the refer- problem [15]:
ence output, Ŷ (t) = [ŷ(t + 1), ŷ(t + 2), . . ., ŷ(t + Tp )] is pre-
Uk+1 (t) = Uk (t) + ΔUk (t) (12)
dictive output, and Tp represents prediction horizon. Then, de-  
termine the form of network input. According to the controlled ∂J(t)
ΔUk (t) = η1 − (13)
system in this article formulated in (1), the input sequence ∂Uk (t)
comes from two parts: previous system states and previous ma-
where η1 > 0 is learning rate and k is the number of iteration.
nipulated variables, that is, xt = [U p (t), Y p (t)] = [u(t), u(t −
According to (10), the derivative of the objective function J(t)
1), . . .u(t − lu ), y(t), y(t − 1), . . .y(t − ly )]. Finally, update
can be rewritten as
network weights with training samples through backpropagation  T
through time mechanism [25]. ∂J(t) ∂ Y (t) 
Compared to traditional predictive methods, the main advan- = −a R(t) − Y (t) + bΔUk (t).
∂Uk (t) ∂Uk (t)
tage of the LSTM-based predictive model lies in two folds.
First, LSTM network has a unique memory cell that can extract (14)
sequential features for time series, which was in accord with Thus, (13) can be formulated as follows:
the feedback control feature of MPC, thus it lays foundation for ⎛  T ⎞
accurate multimode prediction. Second, unlike the traditional 
∂ Y (t) 
ΔUk (t) = η1 ⎝a R(t) − Y (t) − bΔUk (t)⎠
predictive methods like ARIMA, the order of system has less ∂Uk (t)
impact on LSTM network prediction effect, which enhances
 T
generalization ability of proposed predictive model. Therefore,
1 ∂ Y (t) 
LSTM network is chosen as prediction model for multimode = η1 a R(t) − Y (t) . (15)
process in this article [15], [18]. 1 + η1 b ∂Uk (t)
It can be found that the key to deal with optimization problem is
C. Online Optimal Control to calculate the derivative of the prediction model’s output, that
 (t)
∂Y
The goal of MPC is to find a suitable control signal u(t) is, Jacobian matrix ∂U k (t)
. Since the activation functions used
through online optimization so that the output of the system in LSTM network are sigmoid function and hyperbolic tangent
tracks the reference trajectory r(t) as closely as possible. In this function, which are continuous and derivable, it is feasible to
article, given prediction and control horizons Tp and Tc , the calculate Jacobian matrix of the proposed prediction model.

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 08,2023 at 13:24:20 UTC from IEEE Xplore. Restrictions apply.
11548 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 70, NO. 11, NOVEMBER 2023

Besides, numerical differentiation is applied as an alternative

Algorithm 1: Detail of LSTM-MPC method.
to ensure each element in Jacobian matrix is available.
For GD method, constraints can be managed through the 1: Input: reference trajectory R(t), historical input
projected gradient method where the optimization variables are U p (t) = [u(t), u(t − 1), . . ., u(t − lu )], historical output
projected onto the allowed hyperspace. Since the constraints in Y p (t) = [y(t), y(t − 1), . . ., y(t − ly )], optimization
the proposed optimization problem are box constraints (constant termination condition threshold Jset
or linear constraints), (12)–(13) can be modified as follows [26]: 2: Training predictive model based on LSTM network
3: Online optimization
Uk+1 (t) = P1 (Uk (t) + ΔUk (t)) (16) 4: If the value of objective function J > Jset
   5: Predict system’s output Ŷ (t) through LSTM network
∂J(t)
ΔUk (t) = P2 η1 − (17) 6: Calculate Jacobian matrix ∂U ∂ Ŷ (t)
∂Uk (t) k (t)
7: Update control signal through (12)–(15)
where P1 k (uk (t)), P2 k (Δuk (t)) is the projection function for 8: Update learning rate through η k+1 1 = η k 1 e−αk
one element of vectors Uk (t), ΔUk (k), which can be defined as 9: Calculate the value of the objective function to
P1k (uk (t)) = min(umax , max(umin , uk (t))) determine whether the termination condition is
reached
⎧
⎨umax , if uk (t) > umax
⎪ 10: Output: Optimal control sequence
U (t) = [u(t), u(t + 1), . . ., u(t + Tc − 1)]
= uk (t), if umin ≤ uk (t) ≤ umax (18)
⎪
⎩
umin , if uk (t) < umin
P2k (Δuk (t)) = min(Δumax , max(−Δumax , Δuk (t))) calculating through rolling optimization will keep controlled
system stable, achieving setting goal. However, since the pro-
⎧
⎨Δumax , if Δuk (t) > Δumax
⎪ posed method combines deep learning with MPC framework,
it brings new challenges for convergence and stability analysis:
= Δuk (t), if−Δumax ≤ Δuk (t) ≤ Δumax
⎪
⎩ deep learning models are usually black box models and their
−Δumax , if Δuk (t) < −Δumax mathematical expression cannot be represented directly. There-
(19) fore, the convergence and stability analysis are two important
where umax is the upper bound constraint on uk (t), umin is issues for the proposed method and this section will focus on
the lower bound constraint on uk (t), and Δumax is the bound dealing with them.
of incremental control input Δuk (t). When it comes to system
state constraints, like how to deal with control input constraints, A. Convergence of LSTM-Based Predictive Model
the projected method will be added to prediction model to ensure For the proposed LSTM-based predictive model, its con-
the constraints are satisfied. In order to speed up the optimization vergence will be studied in this section. In the convergence
process, we propose an improved adaptive GD method, which analysis, the main goal is to prove the training error of the
use natural logarithm to decay learning rate so that we can proposed method decreases gradually. Here, the training error
accelerate the convergence speed of the proposed method can be represented as e(t) = ŷ(t) − y(t) and ŷ(t), y(t) can be
η k+1 1 = η k 1 e−αk (20) formulated as follows [27]:

where α is the decay rate of each iteration and k is the number g1 (y(t), u(t)) = W ∗ θ(y(t), u(t)) = ẏ(t) + y(t) (21)
of iterations.
˙ + ŷ(t)
g2 (y(t), u(t)) = W θ(y(t), u(t)) = ŷ(t) (22)
Through solving optimization problem, the optimal control
input sequences U (t) can be obtained. Then, the first element
where g1 (·) and g2 (·) are nonlinear function, u(t) is manipu-
of U (t) will be applied as control signals in system to ensure
lated variable, W ∗ stands for ideal network weight matrix, W
the output of the system can track reference trajectory in dif-
represents real network weight matrix during training process,
ferent modes accurately. The proposed method is summarized
and θ(·) denotes activation function. During theoretical proving,
in Algorithm 1, where Jset represents optimization termination
several assumptions are introduced as follows:
threshold, which is designed according to the characteristics of
1) The training samples are the bounded sequences.
controlled system and set as Jset = 10−4 in this article.
2) The weight matrix W is bounded.
3) There exists optimal constant weight matrix W ∗ .
IV. CONVERGENCE AND STABILITY ANALYSIS
Lemma 1: if assumptions 1)–3) are valid, then we have
When it comes to practical industrial processes, stability is
an important feature to determine whether a control method W ∗ − W  < α (23)
performs better or not [27]. What makes MPC widely used has g1 (y(t), u(t)) − g2 (y(t), u(t)) = e(t) ≤ ef (24)
two reasons: First, its predictive model is convergence during
training stage so that the dynamic features of controlled system where ef > 0, α > 0,  ·  represents calculating Euclidean
can be fully extracted. Second, it can ensure that control law distance.

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 08,2023 at 13:24:20 UTC from IEEE Xplore. Restrictions apply.
HUANG et al.: LSTM-MPC: A DEEP LEARNING BASED PREDICTIVE CONTROL METHOD 11549

Proof: Since W ∗ is constant and W is bounded, there exists function as follows:

a positive real number α for W ∗ − W  < α. Meanwhile, in 1 T
order to prevent gradient exploding problem, the output of Vc (t) = E (t)E(t). (28)
2
network is usually limited in a certain range, which can be
Its derivative is then given as
represented as (24).
Theorem 1: Given the LSTM-based predictive model, the . 1 .
V̇c (t) = 2 E (t) E T (t) = E (t)E T (t). (29)
training error e(t) is uniformly ultimately bounded (UUB) ac- 2
cording to Lemma 1 and convergence theorem. Using the derivative chain rule, it has
Proof: Lyapunov function is an effective method to analyze
∂E(t) ∂(r(t) − y(t)) ∂u(t)
control stability [28]. In this section, a Lyapunov function can =
be designed as follows: ∂t ∂u(t) ∂t
y (t)
∂
1 T = − Δu(t). (30)
V (e(t)) = e (t)e(t). (25) ∂u(t)
2
According to (15)–(30), (29) can be rewritten as
Its derivative is then formulated as
 T
y (t) η1 a
∂ y (t)
∂
V̇ (e(t)) = eT (t)ė(t) V̇c (t) = − (r(t)− y (t))T
y (t)) (r(t)−
∂u(t) 1+η1 b ∂u(t)
= eT (t) (g1 (y(t), u(t)) − g2 (y(t), u(t)) − e(t))  T
η1 a y (t)
∂ y (t)
∂
=− y (t))T (r(t)−
(r(t)− y (t)).
≤ −e(t)2 +e(t) W ∗ θ(y(t), u(t)) 1+η1 b ∂u(t) ∂u(t)
(31)
−W θ(y(t), u(t))
Since V̇c (t) < 0, according to Lyapunov stability theory, the
= − e(t)2 + e(t) θ(y(t), u(t)) W ∗ − W  proposed control method is stable.
≤ − e(t)2 + e(t) W ∗ − W 
C. Feasibility Analysis of LSTM-MPC
≤ − e(t) (e(t) − α) (26)
Another important issue for the proposed method is to ensure
˙ − ẏ(t) is error dynamics. Above analysis
where ė(t) = ŷ(t) its control solution can always satisfy constraints of the opti-
implies that [15], [27]: mization problem, which means it is feasible. Since the proposed
1) If e(t) ≥ α, V̇ (e(t)) ≤ 0, and it has e(t) ≤ ef , rep- method is a learning model-based control method and applies
resenting the training error would always maintain in a iterative method to find optimal solution, its recursive feasibility
certain range. has been proven by Rosolia and Borrelli [29] and some ideas
2) If e(t) < α, V̇ (e(t)) > 0, the training error will in- can be introduced from it. The sampled safe set at iteration j
crease to e(t) = α or e(t) = α at some time. When is defined as SS j , which means the collection of all controlled
e(t) = α or e(t) > α, V̇ (e(t)) will be nonpositive, system output states that satisfy constraints. The iteration cost
which means e(t) ≤ ef and similar to case 1). at time t of the jth iteration can be defined as
Therefore, the training error will eventually satisfy e(t) ≤ ∞

j
ef , which makes it UUB and the LSTM-based predictive model Ct→∞ (ŷj (t)) = J (uj (k), ŷj (k)) (32)
convergent. k=t

where J(·) is the stage cost, which is similar to the objective

B. Stability Analysis of LSTM-MPC function, ŷj (k) and uj (k) are the system state and input at the
jth iteration. Then, function Qj (·) is defined as follows in order
Before analyzing the stability of the proposed method, it is
to prove the recursive feasibility of the proposed method
necessary to analyze whether the existing hyperparameters in  i
the optimization problem will affect the stability [15]. For the minj Ct→∞ (x), if x ∈ SS j
j
control weight parameters a and b, they are usually positive and Q (x) = (i,t)∈F (x) (33)
+∞, / SS j
if x ∈
have no impact on stability analysis [18].
Theorem 2: Given control weighting parameters a > 0 and where F j (x) in above equation is formulated as
b > 0, learning rate η1 > 0 for a control law sequence u(t), the
tracking control process in LSTM-MPC is stable. F j (x) = {(i, t) : i ∈ [0, j],t ≥ 0, with x = ŷi (t)
(34)
Proof: First, the tracking control error can be defined as f or ŷi (t) ∈ SS j .
follows: For each x ∈ SS j , function Qj (·) can be rewritten as [29]
∞

E(t) = r(t) − y(t) (27) ∗
∀x ∈ SS j , Qj (x) = Cti∗ →∞ (x) = J (ui∗ (k), ŷi∗ (k))
where r(t) is reference trajectory, y(t) is actual output of system, k=t∗

and has y(t)=ŷ(t) + e(t). In this section, consider a Lyapunov (35)

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 08,2023 at 13:24:20 UTC from IEEE Xplore. Restrictions apply.
11550 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 70, NO. 11, NOVEMBER 2023

where the indices pair (i∗ , t∗ ) is the minimizer, whose expression TABLE I
MODE DISTRIBUTION OF HAMMERSTEIN SYSTEM
is formulated as
(i∗ , t∗ ) = argmin Ct→∞
i
(x), f or x ∈ SS j . (36)
(i,t)∈F j (x)

Assumption 4: At initial state, we assume there exists feasible

system output and optimal control law sequence Ŷ (1) and U (1). overshoot, and average change of manipulated variable which
Assumption 4 is not restrictive in practice for numbers of are defined as follows:

applications. For instance, in the industrial process one can  n 2
1
always obtain a feasible state and input sequence from field RMSE =  (ŷi − ri )
workers’ experiment as initial state. n i=1
Theorem 3: Given the sampled safe set SS j at iteration j and
let Assumption 4 hold, the optimal control law and system states ymax − y(∞)
overshoot = × 100%
obtained by proposed method is always feasible. y(∞)
Proof: Assume that at time t of the jth iteration, the proposed n
1
method obtains feasible optimal control law sequence Uj (t), Δu = |Δu(i)| (41)
which can be formulated as follows: n i=1

Uj (t) = [uj (t|t), . . ., uj (t + Tc − 1|t)]. (37) where ri represents reference trajectory, ŷi stands for model
prediction output, ymax is the maximum output, y(∞) denotes
Similarly, the optimal control law sequence at time t + 1 of jth steady state of output, and Δu(i) is the change of manipulated
iteration is like variable.
Uj (t + 1) = [uj (t + 1|t + 1), . . ., uj (t + Tc |t + 1)]. (38)
A. Numerical Simulation
The goal is to find a suitable way to represent Uj (t + 1) with
the elements in Uj (t) to ensure recursive feasibility according In this section, Hammerstein system [32] is selected to study
to theory of set invariance [30]. We assume Tc = N and it can performance of each control method. To construct a multimode
be found that the length of optimal control law sequence is system, original expression has been modified as
⎧
equal to N . Under the MPC framework, only the first element ⎨x(k) = au(k) − bu(k)2 + 0.5u(k)3
in Uj (t) would be applied to the controlled system and others y(k + 1) = 0.6y(k) − 0.1y(k − 1) + 1.2x(k) (42)
are thought to be feasible optimal control law for later time ⎩
−0.1x(k − 1) + v(k)
step. Therefore, the first N − 1 elements in Uj (t + 1) can be
transferred as follows to ensure feasibility: where v(k) stands for Gaussian white noise output with mean 0
and the standard derivation 0.01. In this article, we mainly focus
uj (t + i|t + 1) = uj (t + i|t), i ∈ [1, N − 1]. (39) on different operation modes caused by different system param-
Since ŷj (t + N |t) ∈ SS j , according to (33), we can obtain eters. Therefore, three modes of the systems are established as
∗
shown in Table I by changing system parameters.
Qj (ŷj (t + N |t)) = Cti∗ →∞ (ŷj (t + N |t)) During the experiment, the range of system input is u ∈ (0, 1).
∞
 Grid search is used to fine-tune LSTM network parameters
= J (ui∗ (k), ŷi∗ (k)) . (40) for better prediction accuracy and the parameters used in the
k=t∗ experiment are as follows: the number of hidden nodes in the
network is 16, the time step is 2, the batch size is 128, and
Through the definition of Qj (·) and F j (x), it has ŷj (t + N |t) =
the number of iterations is 200. As for the proposed multimode
ŷi∗ (t∗ ) and the last element in Uj (t + 1) can be rewritten as
MPC, its hyperparameters are a = 1, b = 1, prediction horizon
uj (t + Tc |t + 1) = ui∗ (t∗ ) [29]. Therefore, it can be found that
is Tp = 2, control horizon is Tc = 1.
Uj (t + 1) can be represented with elements in Uj (t), which
1) Prediction Effect: Since prediction model is one of the
means if the proposed method is feasible at time t, then the
essential parts of MPC, we first conduct comparison experiment
proposed method is also feasible at time t + 1.
on it. In this part, LSTM network and DNN are selected to com-
Through above analysis, the solution obtained by the proposed
pare since LMPC mainly focuses on how to combine Lyapunov
method is recursively feasible.
constraints with MPC framework and its prediction model is the
same as DNN-MPC. The experiment result can be found in Fig. 5
V. CASE STUDY and it indicates LSTM network has better prediction effect since
In this part, a numerical simulation process and an industrial it can extract each operation mode features and automatically
process simulation platform are designed to verify whether the identify the difference of mode to achieve accurate dynamic
proposed method has practical application value. To demonstrate tracking with controlled system.
strengths of proposed method, DNN-MPC [19], and Lyapunov- 2) Optimization Effect: Another important procedure of
based MPC (LMPC) [31] are selected as compared methods. The MPC is rolling optimization, which would try to find the best
performance is measured using root mean-square error (RMSE), control strategy at each iteration. Thus, for each comparison

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 08,2023 at 13:24:20 UTC from IEEE Xplore. Restrictions apply.
HUANG et al.: LSTM-MPC: A DEEP LEARNING BASED PREDICTIVE CONTROL METHOD 11551

Fig. 7. Control results of numerical simulation. (a) LSTM-MPC.

(b) DNN-MPC. (c) LMPC.

TABLE II
CONTROL RESULTS OF NUMERICAL SIMULATION PROCESS

Fig. 5. Prediction results of numerical simulation process. Here,

(a) LSTM network. (b) DNN.

Fig. 8. Spending time of each compared method.

can not stabilize the system output near the set value. Although
DNN-MPC can realize stable control, it has a larger steady-state
error due to its insufficient capacity to predict controlled system
dynamic and find optimal control law for multimode process
shown in Figs. 5 and 6, which cannot meet the demand of precise
control. The proposed method applies LSTM-based predictive
Fig. 6. Value of objective function during optimization procedure of model to accurately predict the system output under different
each compared method. (a) LSTM-MPC. (b) DNN-MPC. (c) LMPC. modes, which lays a foundation for precise control. Meanwhile,
an improved GD method is used to solve the control optimization
method, the value of objective function has been collected to problem, which meets the real-time demand.
4) Computational Efficiency: Another aspect to evalu-
study the optimization effect as shown in Fig. 6. It can be found
that compared to other two methods, the proposed method has ate the effect of the proposed method is computational effi-
two advantages: First, when the optimization procedure is done, ciency. The computational efficiency of the proposed method
the value of objective function of proposed method is much can be analyzed from two aspects: computational complex-
smaller than other compared methods, leading to a better control ity and spending time. For the proposed method, if the con-
effect. Second, when operation mode switching, the proposed trol input and system output dimensions are nu and ny ,
the computational complexity of the proposed GD method is
method has less fluctuation and can eliminate the impact of  
different modes immediately, ensuring the optimization result O Tp 2 + Tc2 (nu + ny )3 [27]. Therefore, when applying
is always in the best state. the proposed method to a real industrial system, some parameters
3) Control Effect: We choose DNN-MPC and LMPC as like prediction horizon Tp can be adjusted to reduce the overall
comparison methods to show the superiority of the proposed computational complexity while ensuring the performance of
method for multimode process and the setting value for three the method.
modes is 0.6. The control results of all methods are shown in Fig. In order to demonstrate the comparison of time cost for each
7 and Table II. It can be found that affected by noise and mode method, another comparison experiment on spending time has
changes, LMPC control method causes system fluctuation and been conducted. From Fig. 8, it can be found the time cost of the

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 08,2023 at 13:24:20 UTC from IEEE Xplore. Restrictions apply.
11552 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 70, NO. 11, NOVEMBER 2023

TABLE III
MODE DISTRIBUTION OF CSTR PROCESS

Fig. 9. Structure of industrial process simulation platform.

Fig. 10. Control results of CSTR process. (a) LSTM-MPC. (b) DNN-
MPC. (c) LMPC.
proposed method is close to the DNN-MPC method since their
prediction models are based on neural network and GD method
is suitable for both methods. As for Lyapunov MPC (LMPC), TABLE IV
it combines theory of Lyapunov stability and MPC framework, CONTROL RESULTS OF NUMERICAL SIMULATION PROCESS
which means it has stricter constraint and would spend more
time solving optimization problem. Accordingly, the time cost
of the proposed method is at a suitable range.

B. Industrial Process Simulation Platform

In order to test the practical application of the proposed
method for real industrial processes, an experiment on an in-
dustrial process simulation platform is presented. The structure
of the platform is shown in Fig. 9. The digital plant is designed
to simulate a real industrial process. Since most industrial pro-
cess apply PLC device to conduct control strategy, we choose
Siemens PLC device to simulate this process. The proposed
method is deployed in advanced control center, which would near the setting goal while the proposed method has a smaller
calculate and send an optimal result to PLC device, ensuring overshoot, which means it can ensure smooth switching between
the simulated process stable. All running data would be stored each mode to reduce the negative effect of impulse response and
in industrial data center, which lays foundation for updating keep the system running stably.
control strategy in time. In order to demonstrate the trend of
controlled variables, operator station is established to monitor
the state of the simulation process in time. VI. CONCLUSION
After building up the simulation platform, CSTR process [33] In order to deal with multimode control problem, this article
is chosen to conduct multimode control experiment. For CSTR proposed a deep learning based MPC method: LSTM-MPC.
process, its goal is to maintain reaction temperature T and outlet First, a LSTM network-based prediction model was established,
concentration CA stable by adjusting the temperature of cooling which can automatically match different operation modes with-
water Tc and the velocity of reactant q. Thus, the manipulated out additional switching strategies and accurately predict sys-
variables are Tc ∈ (415, 425) and q ∈ (95, 115). In this article, tem’s behavior. Then combined with the MPC framework, an
three modes are defined in Table III. adaptive GD method was put forward to handle optimization
Control results are shown in Fig. 10 and Table IV. It can problem and its constraints. Finally, through theoretical analysis,
be found that LMPC control method takes more time to track the proposed method was proven stable and feasible. Several
the reference trajectory when the operation modes are changed, experiments demonstrated the proposed method can achieve
which cannot meet the demands of control. The proposed accuracy and robustness control performance for multimode
method and DNN-MPC can both control the system output process.

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 08,2023 at 13:24:20 UTC from IEEE Xplore. Restrictions apply.
HUANG et al.: LSTM-MPC: A DEEP LEARNING BASED PREDICTIVE CONTROL METHOD 11553

REFERENCES [24] Y. Bengio, P. Simard, and P. Frasconi, “Learning long-term dependencies

with gradient descent is difficult,” IEEE Trans. Neural Netw., vol. 5, no. 2,
[1] C. Shen, Y. Shi, and B. Buckham, “Trajectory tracking control of an pp. 157–166, Mar. 1994.
autonomous underwater vehicle using Lyapunov-based model predictive [25] S. S. Miriyala and K. Mitra, “Deep learning based system identification
control,” IEEE Trans. Ind. Electron., vol. 65, no. 7, pp. 5796–5805, of industrial integrated grinding circuits,” Powder Technol., vol. 360,
Jul. 2018. pp. 921–936, 2020.
[2] M. Liu, Y. Shi, and X. Liu, “Distributed MPC of aggregated heteroge- [26] J. Viljoen, C. Muller, and I. Craig, “Hybrid nonlinear model predictive
neous thermostatically controlled loads in smart grid,” IEEE Trans. Ind. control of a cooling water network,” Control Eng. Pract., vol. 97, 2020,
Electron., vol. 63, no. 2, pp. 1120–1129, Feb. 2016. Art. no. 104319.
[3] H. Han and J. Qiao, “Nonlinear model-predictive control for industrial [27] H. Han, X. Wu, and J. Qiao, “Real-time model predictive control using a
processes: An application to wastewater treatment process,” IEEE Trans. self-organizing neural network,” IEEE Trans. Neural Netw. Learn. Syst.,
Ind. Electron., vol. 61, no. 4, pp. 1970–1982, Apr. 2014. vol. 24, no. 9, pp. 1425–1436, Sep. 2013.
[4] H. Li and Y. Shi, “Network-based predictive control for constrained [28] Z. Wu, A. Tran, D. Rincon, and P. D. Christofides, “Machine learning-
nonlinear systems with two-channel packet dropouts,” IEEE Trans. Ind. based predictive control of nonlinear processes. Part I: Theory,” AIChE J.,
Electron., vol. 61, no. 3, pp. 1574–1582, Mar. 2014. vol. 65, no. 11, 2019, Art. no. e16729.
[5] Z. Yang, Y. Li, and J. E. Seem, “Multi-model predictive control for wind [29] U. Rosolia and F. Borrelli, “Learning model predictive control for iterative
turbine operation under meandering wake of upstream turbines,” Control tasks. a data-driven control framework,” IEEE Trans. Autom. Control,
Eng. Pract., vol. 45, pp. 37–45, 2015. vol. 63, no. 7, pp. 1883–1896, Jul. 2018.
[6] A. Ferramosca, A. H. González, and D. Limon, “Offset-free multi-model [30] F. Blanchini, “Set invariance in control,” Automatica, vol. 35, no. 11,
economic model predictive control for changing economic criterion,” J. pp. 1747–1767, 1999.
Process Control, vol. 54, pp. 1–13, 2017. [31] H. Oyama and H. Durand, “Interactions between control and process
[7] K. Huang et al., “Adaptive multimode process monitoring based on mode- design under economic model predictive control,” J. Process Control,
matching and similarity-preserving dictionary learning,” IEEE Trans. Cy- vol. 92, pp. 1–18, 2020.
bern., 2022, doi: 10.1109/TCYB.2022.3178878. [32] Z. Hou, S. Liu, and T. Tian, “Lazy-learning-based data-driven model-free
[8] K. Huang, Z. Tao, C. Wang, T. Guo, C. Yang, and W. Gui, “Cloud-edge adaptive predictive control for a class of discrete-time nonlinear systems,”
collaborative method for industrial process monitoring based on error- IEEE Trans. Neural Netw. Learn. Syst., vol. 28, no. 8, pp. 1914–1928,
triggered dictionary learning,” IEEE Trans. Ind. Informat., vol. 18, no. 12, Aug. 2017.
pp. 8957–8966, Dec. 2022. [33] G. Li, S. J. Qin, Y. Ji, and D. Zhou, “Reconstruction based fault prog-
[9] A. R. Ghanbari and A. A. Raie, “Proposing a multimode switching control nosis for continuous processes,” Control Eng. Pract., vol. 18, no. 10,
method for a half-bridge converter to improve its efficiency over the entire pp. 1211–1219, 2010.
load variations,” IEEE Trans. Ind. Electron., vol. 67, no. 3, pp. 1910–1920,
Mar. 2020.
[10] Y. Xi, F. Wang, and G. Wu, “Nonlinear multi-model predictive control,”
IFAC Proc. Vol., vol. 29, no. 1, pp. 2359–2364, 1996.
[11] F. D. Palma and L. Magni, “A multi-model structure for model predictive
control,” Annu. Rev. Control, vol. 28, no. 1, pp. 47–52, 2004.
[12] N. Li, S. Li, and Y. Xi, “Multi-model predictive control based on the Keke Huang (Member, IEEE) received the B.A.
Takagi–Sugeno fuzzy models: a case study,” Inf. Sci., vol. 165, no. 3, degree in automatic control from Northeastern
pp. 247–263, 2004. University, Shenyang, China, in 2012, and the
[13] W. Tang and M. N. Karim, “Multi-model MPC for nonlinear systems: Case Ph.D. degree in control science and engineer-
study of a complex PH neutralization process,” in Proc. 21st Eur. Symp. ing from Tsinghua University, Beijing, China, in
Comput. Aided Process Eng., 2011, vol. 29, pp. 622–627. 2017.
[14] J. Wan, B. He, Y. Shen, W. Liu, X. Ding, and S. Gao, “Heading multi-mode He was an Associate Professor with Central
control based on soft-switching for autonomous underwater vehicle,” South University, Changsha, China, from April
Ocean Eng., vol. 164, pp. 672–682, 2018. 2017 to September 2021. He is currently a Full
[15] G. Wang, Q.-S. Jia, J. Qiao, J. Bi, and M. Zhou, “Deep learning-based Professor with the School of Automation, Cen-
model predictive control for continuous stirred-tank reactor system,” IEEE tral South University, Changsha. His research
Trans. Neural Netw. Learn. Syst., vol. 32, no. 8, pp. 3643–3652, Aug. 2021. interests include industrial Big Data, process monitoring and control, and
[16] G. Wang, Q. Jia, J. Qiao, J. Bi, and C. Liu, “A sparse deep belief network sciences.
network with efficient fuzzy learning framework,” Neural Netw., vol. 121,
pp. 430–440, 2020.
[17] X. Yin and X. Zhao, “Deep neural learning based distributed predictive
control for offshore wind farm using high-fidelity LES data,” IEEE Trans.
Ind. Electron., vol. 68, no. 4, pp. 3251–3261, Apr. 2021.
[18] H. Han, L. Zhang, Y. Hou, and J. Qiao, “Nonlinear model predictive control
based on a self-organizing recurrent neural network,” IEEE Trans. Neural
Netw. Learn. Syst., vol. 27, no. 2, pp. 402–415, Feb. 2016.
[19] S. Lucia, D. Navarro, B. Karg, H. Sarnago, and O. Lucia, “Deep learning-
based model predictive control for resonant power converters,” IEEE
Trans. Ind. Inform., vol. 17, no. 1, pp. 409–420, Jan. 2021.
[20] Y. Pan and J. Wang, “Model predictive control of unknown nonlinear
dynamical systems based on recurrent neural networks,” IEEE Trans. Ind.
Electron., vol. 59, no. 8, pp. 3089–3101, Aug. 2012.
Ke Wei received B.A. degree in automation from
[21] X. Huang, Y. Ye, C. Wang, X. Yang, and L. Xiong, “A multi-mode
Central South University, Changsha, China, in
traffic flow prediction method with clustering based attention convolution
2020, where he is currently working toward the
LSTM,” Appl. Intell., vol. 52, pp. 14773–14786, 2021.
M.A. degree in control science and engineering.
[22] Y. Liu, C. Yang, K. Huang, and W. Gui, “Non-ferrous metals price
His research interests include industrial big
forecasting based on variational mode decomposition and LSTM network,”
data and process monitoring.
Knowl.-Based Syst., vol. 188, 2020, Art. no. 105006.
[23] K. Huang, K. Wei, Y. Li, C. Yang, and W. Gui, “Static and dynamic joint
analysis for operation condition division of industrial process with incre-
mental learning,” IEEE Internet Things J., vol. 9, no. 22, pp. 22081–22094,
Nov. 2022.

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 08,2023 at 13:24:20 UTC from IEEE Xplore. Restrictions apply.
11554 IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS, VOL. 70, NO. 11, NOVEMBER 2023

Fanbiao Li received the B.Sc. degree in applied Weihua Gui received the B.Eng. degree in elec-
mathematics from Mudanjiang Normal Univer- trical engineering and the M.S. degree in au-
sity, Mudanjiang, China, in 2008, the M.Sc. de- tomatic control engineering from Central South
gree in operational research and cybernetics University, Changsha, China, in 1976 and 1981,
from Heilongjiang University, Harbin, China, in respectively.
2012, and the Ph.D. degree in control theory Since 2013, he has been an Academician
and control engineering from the Harbin Insti- of the Chinese Academy of Engineering. He is
tute of Technology, Harbin, in 2015. currently with the School of Automation, Central
In 2016, he joined as an Associate Profes- South University. His current research interests
sor with Central South University, China. His include modeling and optimal control of com-
research interests include stochastic systems, plex industrial processes, fault diagnoses, and
sliding mode control, and fault diagnosis and identification. distributed robust control.

Chunhua Yang (Fellow, IEEE) received the

M.S. degree in automatic control engineering
and the Ph.D. degree in control science and en-
gineering from Central South University, Chang-
sha, China, in 1988 and 2002, respectively.
Since 1999, she has been a Full Professor
with the School of Information Science and En-
gineering, Central South University. She is cur-
rently the HoD of the School of Automation. Her
current research interests include modeling and
optimal control of complex industrial processes,
and intelligent control systems.

Authorized licensed use limited to: University of Science & Technology of China. Downloaded on July 08,2023 at 13:24:20 UTC from IEEE Xplore. Restrictions apply.