Highly efficient model predictive controller design method

Abstract

A kind of high-efficiency model predictive controller design method, by constructing non-isothermal continuous stirred tank reactor model and establishing prototype predictive controller based on the model, system feedback output input data is generated by prototype predictive controller in offline stage and is trained approximate controller;In each control cycle in online stage, the field process variable reactor concentration information and reactor temperature signal of continuous stirred tank reactor are collected and input to the trained approximate controller, and then the control signal obtained is output to the actual continuous stirred tank reactor system actuator.This application can ensure the error boundary under a certain confidence by increasing the learning data set, avoid complex online optimization solving process, speed up the controller running speed, and ensure the stability and optimization performance of the system when running approximate predictive controller.

Description

Technical Field

[0001] This invention relates to a technology in the field of model predictive control, specifically an optimized implementation method for a high-efficiency model predictive controller. Background Technology

[0002] Model predictive control (MPC) solves a finite-time-domain multi-constraint optimization problem at each sampling time, and has the ability to handle explicit constraints. A major challenge in its online application is that solving optimization problems with a large number of variables or requiring high sampling rates requires a significant amount of computation time.

[0003] In existing technologies, data-driven and machine learning methods are widely used to improve the processing speed of model predictive control. Using data-driven methods such as machine learning, a pre-designed controller is learned offline to obtain an approximate controller. In online applications, the measured signals of the acquired field process variables are input into the approximate controller, and the output is applied to the system. This avoids the complex online optimization process and speeds up the controller's operation. However, since approximate controllers inevitably contain errors, these errors can affect the controller's stability performance. Summary of the Invention

[0004] This invention addresses the shortcomings of existing technologies in rapidly solving predictive control problems under multiple constraints and nonlinear constraints, and in increasing the control frequency. It proposes a highly efficient model predictive controller design method. This method employs a data-driven approach, considers the error of the approximate controller, adjusts the predictive controller parameters, and ensures the error boundary is maintained at a certain confidence level by increasing the learning dataset. This avoids complex online optimization processes, accelerates the controller's operation, and ensures the system's stability and optimization performance when running the approximate predictive controller.

[0005] This invention is achieved through the following technical solution:

[0006] This invention relates to a method for optimizing a model predictive controller. The method involves constructing a non-isothermal continuous stirred reactor model and establishing a prototype predictive controller based on this model. In the offline phase, the prototype predictive controller generates system feedback input / output data, which is then used to train an approximate controller. During each control cycle in the online phase, the on-site process variables of the continuous stirred reactor, including reactor concentration and temperature signals, are collected and input to the trained approximate controller. The resulting control signal is then output to the actuator of the actual continuous stirred reactor system.

[0007] The prototype prediction controller is specifically: Where: t k Where N is the current sampling time, N is the total predictive control step size, and L is the predictive control objective function. Let τ be the predicted state of the system. Let u be the state derivative. c For system output, f is the state-to-state derivative mapping function, g is the input-to-state derivative mapping function, and x(t) is the system output. k (t) represents the initial time. k State, X n U c These are the input and output constraint sets, X. f This is the set of terminal state constraints.

[0008] The approximate controller is a backpropagation neural network (BPNN) containing one input layer, two hidden layers and one output layer, wherein the first hidden layer contains seven neurons and the second hidden layer contains five neurons.

[0009] The present invention relates to a system for implementing the above method, comprising: a prototype predictive controller, an approximate controller, and a system actuator, wherein: the prototype predictive controller performs model predictive control solution based on system input state information to obtain an output result; the approximate controller trains a neural network based on the input and output information of the prototype predictive controller to obtain an approximate controller representing the input to the output; and the system actuator acts on the actual system based on the output information of the approximate controller to obtain the system state under control. Technical effect

[0010] This invention considers the error of the approximate controller and adjusts the parameters of the prototype controller. When designing the approximate controller, an additional learning dataset is added to ensure that the error boundary is maintained at a certain confidence level. Compared with existing technologies, this invention guarantees the controller's execution speed and ensures that, after the controller's output is applied to the system, the system's output meets the requirements for safe production output range at a certain confidence level, and converges to the set value at a certain confidence level. Attached Figure Description

[0011] Figure 1 This is a flowchart of the present invention;

[0012] Figure 2 This is a diagram of the neural network approximation controller structure in the embodiment;

[0013] Figure 3 This is a comparison chart of the neural network approximation controller output and the actual MPC solution output in the embodiment;

[0014] Figure 4 To account for input disturbances in the embodiments, the material concentration and temperature change curves are shown under the action of the actual controller and the approximate controller, respectively.

[0015] Figure 5To account for input disturbances in the embodiments, the curves showing the changes in inlet reactant concentration and external heat exchange under the action of the actual controller and the approximate controller are presented respectively.

[0016] Figure 6 To account for input disturbances in the embodiments, the state trajectory curves are shown under the action of the actual controller and the approximate controller, respectively. Detailed Implementation

[0017] like Figure 1 As shown, this embodiment illustrates an optimized implementation method for a model predictive controller, comprising:

[0018] S1. Construct a non-isothermal continuous stirred reactor model, specifically as follows: Where: F is the inlet gas velocity, T0 is the inlet gas temperature, and V is the reactor volume. ΔH is the heat of reaction, k0 is the pre-exponential factor, E is the activation energy, R is the gas constant, σ is the liquid density, and C is the liquid density. p For heat capacity, C A C represents the concentration in the reactor, T represents the temperature in the reactor; As T s These represent the concentration and temperature at steady state, respectively; the state variable is x = [C]. A -C As TT s ] T Q represents the heat input / output to the external environment. s For the heat under steady state, C C0 C represents the inlet reactant concentration. A0s The steady-state intake air concentration; the input is u = [u1, u2]. T =[C A0 -C A0s QQ s ] T .

[0019] S2. Construct a prototype predictive controller based on a non-isothermal continuous stirred reactor model, specifically as follows: Where: t k Where N is the current sampling time, N is the total predictive control step size, and L is the predictive control objective function. Let τ be the predicted state of the system. Let u be the state derivative. c For system output, f is the state-to-state derivative mapping function, g is the input-to-state derivative mapping function, and x(t) is the system output. k (t) represents the initial time. k State, X n U c These are the input and output constraint sets, X.f This is the set of terminal state constraints.

[0020] S3. During the offline phase, the prototype predictive controller generates input and output data, i.e., the system feedback input and output data. The prototype predictive controller adjusts its parameters based on the error of the approximate controller. Specifically, the state constraints of the prototype controller are: X min X max Define upper and lower bounds for state constraints. For a dimension of n x The state space; based on the variance σ of the approximate controller, adjust the state constraints of the original controller as follows: Where γ is the coefficient from control quantity to state. For terminal set Where K is the coefficient from the control quantity to the terminal state. This is a set operation.

[0021] S4. Define the error boundary of the approximate controller and ensure that the error boundary and controller performance can be guaranteed under a certain confidence level by increasing the training dataset. Specifically:

[0022] 4.1 Using an approximate controller to learn the prototype controller, an approximate strategy κ is obtained. map (x).

[0023] 4.2 The input-output set of the approximate controller is X p :={x(t), u(t):u=κ map Let x(t) be the approximate controller output, u(t) be the approximate controller output, and let the feasibility of the approximate controller be represented by an indicator function. Calculate the feasible probability of the approximate controller Where: r is the total number of samples, p is the current sample, and U c For the output constraint set, when the feasible probability is lower than the set value μ l If the sample size r is increased, return to step 3.

[0024] S5. During each control cycle in the online phase, the measurement signals of the on-site process variables of the continuous stirred reactor, namely reactor concentration and reactor temperature, are collected, input to the approximate controller, and the resulting control signal is output to the actuator of the actual continuous stirred reactor system.

[0025] Through specific practical experiments, the proposed method was tested using a well-mixed, non-isothermal continuous stirred tank reactor. In this embodiment, 1000 random initial points were selected within the stability region. Under identical conditions, the online processing time of the actual MPC controller and the approximate controller were compared. The LMPC took an average of 0.8325 s, while the approximate controller took an average of 0.0042 s. The computational efficiency was improved by approximately 230 times. Figure 3 As shown, the output of the approximate controller can fit the output of the actual MPC controller quite well. Figure 4 , Figure 5 As shown, the performance of the approximate controller considering model error is not significantly different from that of the actual MPC controller. Figure 6 As shown, the approximate controller considering model error and the actual MPC controller can control the system state to a stable state.

[0026] Compared with existing technologies, this method adjusts the constraint parameters of the prototype controller and dynamically increases the learning dataset when designing the approximate controller. This ensures that while maintaining the controller's execution speed, the system's output, once applied to the system, meets the requirements of safe production output range under a certain confidence level and converges to the set value under a certain confidence level.

[0027] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A method for optimizing a model predictive controller, characterized in that, A non-isothermal continuous stirred reactor model is constructed, and a prototype predictive controller is established based on this model. In the offline phase, the prototype predictive controller generates system feedback input / output data, which is then used to train an approximate controller. During each control cycle in the online phase, the on-site process variables of the continuous stirred reactor, namely reactor concentration and reactor temperature, are collected and input to the trained approximate controller. The resulting control signal is then output to the actuator of the actual continuous stirred reactor system, specifically including: S1. Construct a non-isothermal continuous stirred reactor model, specifically as follows: , ,in: The intake airflow velocity, Intake air temperature, Reactor volume, Heat of reaction Pre-exponential factors, activation energy, Gas constant, Liquid density, Heat capacity, The concentration in the reactor is [missing information]. The temperature of the reactor; , These represent the concentration and temperature at steady state, respectively; the state variables are: Heat input / output to the outside, Heat under steady state The concentration of reactants at the inlet. The steady-state intake concentration; the input is... ; S2. Construct a prototype predictive controller based on a non-isothermal continuous stirred reactor model, specifically as follows: ,in: At the current sampling time, The total predictive control step size, To predict the control objective function, For a moment The system's predicted state, For the state derivative, For system output, It is a state-to-state derivative mapping function. The input-to-state derivative mapping function, Initial time state, These are the input and output constraint sets, For the terminal state constraint set; S3. During the offline phase, the prototype predictive controller generates input and output data, i.e., the system feeds back the input and output data. Specifically, the state constraints of the prototype controller are: , , Define upper and lower bounds for state constraints. For dimension The state space; based on the variance of the approximate controller. Adjust the original controller state constraints to ,in: The coefficients for the control quantity to the state; for the terminal set ,in The coefficient for controlling the input to the terminal state. This is a set operation; S4. Define the error boundary of the approximate controller and ensure that the error boundary and controller performance can be guaranteed under a certain confidence level by increasing the training dataset. Specifically: 4.1 Using an approximate controller to learn the prototype controller and obtain an approximate strategy ; 4.2 The input-output set of the approximate controller is as follows The feasibility of an approximate controller can be represented by an indicator function. Calculate the feasible probability of the approximate controller ,in: The sample size is used when the feasible probability is lower than a set value. Then increase the number of samples. Return to step 3; S5. During each control cycle in the online phase, the measurement signals of the on-site process variables of the continuous stirred reactor, namely reactor concentration and reactor temperature, are collected, input to the approximate controller, and the resulting control signal is output to the actuator of the actual continuous stirred reactor system.

2. The method for optimizing the model predictive controller according to claim 1, characterized in that, The approximate controller is a backpropagation neural network comprising an input layer, two hidden layers, and an output layer, wherein the first hidden layer contains seven neurons and the second hidden layer contains five neurons.

3. A system for implementing the model predictive controller optimization method according to claim 1 or 2, characterized in that, include: The system comprises a prototype predictive controller, an approximate controller, and a system actuator. The prototype predictive controller performs model predictive control based on the system input state information to obtain the output result. The approximate controller trains a neural network based on the input and output information of the prototype predictive controller to obtain an approximate controller that represents the input to the output. The system actuator acts on the actual system based on the output information of the approximate controller to obtain the system state under control.

Images