Batch reaction control method and control device

Abstract

The present invention provides a control method for batch reaction, comprising: an inner layer adopts an MPC controller to control the output value of a single batch of a batch reactor to track the set trajectory of the corresponding batch along the time axis direction, and the MPC controller includes multiple sub-model MPC controllers; an outer layer adopts an iterative learning controller to calculate the set trajectory in the MPC controller along the batch axis direction and output the set trajectory of the current batch, wherein, for the current batch, the error between the output value of the previous batch of the batch reactor and the set trajectory of the previous batch is used as the input of the iterative learning controller; the output of the iterative learning controller is used as the set trajectory of the current batch of the MPC controller for tracking control, thereby forming a double-layer control to continuously improve the product production quality of the batch reactor. A corresponding control device is also provided. The control method can realize accurate control of the batch reaction process, thereby improving product quality.

Description

Technical Field

[0001] The invention relates to the field of nuclear industry batch process control, and in particular to a batch reaction control method and control equipment. Background Art

[0002] Batch processes are usually used to produce high value-added products on a small scale to meet rapidly changing market demands. Batch processes are currently widely used in the production and preparation of a variety of products such as food, polymers, and pharmaceuticals, and they occupy an important position in chemical production processes and people’s daily lives.

[0003] Most batch processes have significant nonlinear time-varying characteristics, and each batch of raw materials cannot be guaranteed to be completely consistent. In addition, product quality is highly dependent on the precise control of process variables, which places strict requirements on the tracking performance of batch process controllers. Existing nuclear industry batch reactors are usually based on artificial experience models and adjust the reaction parameters of each batch according to the information of each batch of raw materials. The reaction process control also adopts the traditional PID (Proportion Integral Differential) control method, which consumes a lot of manpower and relies heavily on expert experience to intervene in the batch reaction process. There are problems such as poor control effect and low quality of the products produced. Summary of the invention

[0004] The technical problem to be solved by the present invention is to provide a batch reaction control method and control equipment in view of the above-mentioned deficiencies in the prior art, which can realize precise control of the batch reaction process, thereby improving product quality.

[0005] The present invention provides a control method for batch reaction, comprising: an inner layer adopting an MPC (Model Predictive Control) controller to control the output value of a single batch of a batch reactor to track the set trajectory of the corresponding batch along the time axis direction, wherein the MPC controller includes multiple sub-model MPC controllers; an outer layer adopting an iterative learning controller to calculate the set trajectory in the MPC controller along the batch axis direction and output the set trajectory of the current batch, wherein for the current batch, the error between the output value of the previous batch of the batch reactor and the set trajectory of the previous batch is used as the input of the iterative learning controller; the output of the iterative learning controller is used as the set trajectory of the current batch of the MPC controller for tracking control, thereby forming a double-layer control to continuously improve the product production quality of the batch reactor.

[0006] Preferably, the inner layer uses an MPC controller to control the output value of a single batch of the batch reactor to track the set trajectory of the corresponding batch along the time axis, specifically including: dividing the entire batch reaction process into multiple different stages, and obtaining sub-models corresponding to each different stage; using the sub-models corresponding to the different stages as prediction models to design each sub-model MPC controller; selecting a weighting function to globally model the entire batch reaction process; switching each sub-model MPC controller according to the global modeling result of the entire batch reaction process, so as to realize that the output value of a single batch of the batch reactor controlled by the MPC controller tracks the set trajectory of the corresponding batch along the time axis.

[0007] Preferably, the entire batch reaction process is divided into a plurality of different stages, and each sub-model corresponding to each different stage is obtained, specifically including: according to the operation trajectory of the scheduling variable and the nonlinear analysis of the process; the entire batch reaction process is divided into a plurality of different stages in chronological order, and identification experiments are performed on each stage to obtain test data corresponding to each stage; the test data corresponding to each stage is identified by using a prediction error or a subspace identification algorithm to obtain each sub-model G corresponding to each different stage. i ,i=1,...,S.

[0008] Preferably, the weighting function is selected to globally model the entire batch reaction process, specifically including: selecting a weighting function according to the test data corresponding to each stage; and interpolating each sub-model using a nonlinear function of the scheduling variable to globally model the entire batch reaction process as follows:

[0009]

[0010] Among them, the nonlinear function of the scheduling variable is the stage switching logic.

[0011] Preferably, the design of each sub-model MPC controller using the sub-models corresponding to different stages as prediction models specifically includes: each sub-model MPC controller uses the corresponding sub-model as the prediction model; the parameters of each sub-model MPC controller are designed according to the control objectives of different stages, and the parameters of the sub-model MPC controller include a controlled variable weighted matrix, a control increment weighted matrix, and input and output constraints.

[0012] Preferably, the MPC controller parameters of each sub-model are designed according to the control objectives of different stages, specifically including:

[0013] For each sub-model G i The MPC controller is specifically a dynamic matrix control DMC controller. Specifically, for asymptotically stable systems, a finite set of unit step response sampling data {a1, a2, …, a N} describes the dynamic characteristics of the system. The unit step vector of the system is a=[a1 a2…a N ] T is called the model vector of DMC, N is called the modeling domain, and a suitable N is selected so that i The value of (i>N) and the static final value a of the step response N The difference can be ignored. According to the superposition principle of linear systems, the future output value of the system is predicted using the object unit step response model and the given input control increment. At time k, the predicted output value of the system under the control amount Δu(k) is as follows:

[0014]

[0015] In the formula

[0016]

[0017] in, It represents the output prediction value at time k, assuming that the control action remains unchanged for the next N time periods. The symbol “~” in the formula represents prediction, and “k+i / k, i=1, 2, …, N” represents the prediction at time k for time k+i. Similarly, under the action of M continuous control increments Δu(k), Δu(k+1), …, Δu(k+M-1), the output prediction value of the controlled object at the next P time periods is as follows:

[0018]

[0019] In the formula

[0020]

[0021] in It represents the output prediction value at the kth moment, assuming that the control action remains unchanged for the next P moments; the matrix A is a dynamic matrix, whose elements are the step response coefficients describing the dynamic characteristics of the system; P is the length of the rolling optimization time domain; M is the length of the control time domain, P and M should satisfy M≤P≤N, and the vector The first subscript of is the length of the predicted future output, and the second subscript is the number of changes in the control quantity. At different sampling times, considering the input and output constraints, the DMC optimization performance indicators are as follows:

[0022]

[0023] Stu min ≤u≤u max

[0024] y min ≤y≤y max

[0025] Among them, w P (k) = [w(k+1)…w(k+p)] T

[0026] Q=diag(q1…q P )

[0027] R=diag(r1…r M )

[0028] where w P (k) is the expected value vector, Q and R are the error weight matrix and the control weight matrix respectively, Q and R are diagonal matrices composed of weight coefficients, calculate the constrained DMC optimization performance index, and obtain Δu(k), ..., Δu(k+M-1). DMC takes the immediate control increment Δu(k) to form the actual control u(k)=u(k-1)+Δu(k) acting on the object. At the next moment, Δu(k+1) is obtained through the same optimization problem, and u(k+1) is obtained to act on the object. When the predicted value of the system output deviates from the actual value, there is a prediction error as follows:

[0029]

[0030] Real-time information of prediction errors is used for feedback correction.

[0031] Preferably, the iterative learning controller is specifically:

[0032] u k =u k-1 +Le k-1 , where e k =y d -y k

[0033] Among them, k is the number of iterations, e k is the deviation between the actual output and the expected trajectory during the k-th iteration; the matrix L is the gain matrix of the iterative learning algorithm, and the gain matrix L of the iterative learning algorithm is designed using the optimal iterative learning method, that is,

[0034]

[0035] sty k =G·u k +y0

[0036] e k =y d -y k

[0037] Δu k =u k -u k-1

[0038] By solving the above optimization proposition, we get the following learning law u k =u k-1 +Le k-1 , where the gain matrix L = (G T QG+R) -1 G T Q.

[0039] Furthermore, the present invention also provides a batch reaction control device, comprising an inner layer control device, an outer layer control device and a batch reactor.

[0040] A batch reactor is used to complete a batch reaction process. An inner control device is connected to the batch reactor and is used to control the output value of a single batch of the batch reactor to track the set trajectory of the corresponding batch along the time axis direction using an MPC controller, wherein the MPC controller includes multiple sub-model MPC controllers. An outer control device is connected to the batch reactor and the inner control device and is used to use an iterative learning controller to calculate and output the set trajectory of the current batch along the batch axis direction for the set trajectory in the MPC controller of the inner control device, wherein, for the current batch, the error between the output value of the previous batch of the batch reactor and the set trajectory of the previous batch is used as the input of the iterative learning controller, and the output of the iterative learning controller is used as the set trajectory of the current batch of the MPC controller of the inner control device for tracking control.

[0041] Preferably, the inner control device includes an acquisition unit, a design unit, a modeling unit and a switching unit. The acquisition unit is used to divide the entire batch reaction process into a plurality of different stages and acquire the sub-models corresponding to the different stages. The design unit is connected to the acquisition unit and is used to design each sub-model MPC controller using the sub-models corresponding to the different stages as prediction models. The modeling unit is connected to the acquisition unit and is used to select a weighting function to globally model the entire batch reaction process. The switching unit is connected to the modeling unit and the design unit and is used to switch each sub-model MPC controller according to the global modeling result of the entire batch reaction process, so as to realize that the output value of a single batch of the MPC controller controlling the batch reactor tracks the set trajectory of the corresponding batch along the time axis.

[0042] Preferably, the acquisition unit includes a partitioning component, a test component and an acquisition component. The partitioning component is used to divide the entire batch reaction process into a plurality of different stages in chronological order according to the operation trajectory of the scheduling variable and the nonlinear analysis of the process. The test component is connected to the partitioning component and is used to perform identification experimental tests on each stage to obtain test data corresponding to each stage. The acquisition component is connected to the test component and is used to identify the test data corresponding to each stage using a prediction error or a subspace identification algorithm to obtain each submodel G corresponding to each different stage. i ,i=1,...,S.

[0043] The control method and control device of the batch reaction of the present invention controls the output of the batch reactor according to each sub-model MPC controller corresponding to different stages of the entire batch reaction process, thereby achieving rapid and stable control along the time axis, and utilizing the repetitive characteristics of the batch reaction to perform iterative learning control, so that the closed-loop control quality of the batch reactor is continuously improved along the batch axis, ultimately achieving rapid and high-precision control of the batch reactor, thereby improving product quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 A schematic flow chart of a batch reaction control method provided in Example 1 of the present invention;

[0045] Figure 2 A schematic diagram of a control system corresponding to a batch reaction control method provided in Example 1 of the present invention;

[0046] Figure 3 A schematic diagram of identifying sub-models and weighting functions provided in Example 1 of the present invention;

[0047] Figure 4 The MPC structure diagram of the multi-model provided in Example 1 of the present invention;

[0048] Figure 5 This is a comparison chart of the MPC+iterative learning control and PID control effects of batch reactors. DETAILED DESCRIPTION

[0049] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is further described in detail below with reference to the accompanying drawings and embodiments.

[0050] Embodiment 1:

[0051] At present, MPC has been successfully applied in many process industries. The advantage of MPC is that it can explicitly handle constraints and handle multivariable objects well. Since the existing nuclear industry batch reactors have nonlinear time-varying characteristics, and each batch of raw materials cannot be guaranteed to be completely consistent, the inventor believes that the entire batch reaction process should be decomposed into multiple different stages, and a corresponding MPC controller should be designed for each different stage, so as to achieve fast and smooth control for a single batch along the time axis. The more decomposed stages there are, the higher the control accuracy. In addition, iterative learning control is performed using the repetitive characteristics of batch reactions, so that the closed-loop control quality of the batch reactor continues to improve along the batch axis, further improving the control accuracy. Therefore, based on this principle, a control method for batch reactions is proposed.

[0052] like Figure 1 As shown, this embodiment provides a batch reaction control method, comprising:

[0053] Step 101 : The inner layer uses an MPC controller to control the output value of a single batch of a batch reactor to track a set trajectory of the corresponding batch along the time axis, wherein the MPC controller includes a plurality of sub-model MPC controllers.

[0054] In this embodiment, Figure 2 As shown, the inner layer uses an MPC controller to act on the batch reactor. In a single batch, the output value of the batch reactor is controlled to track the set trajectory of the corresponding batch along the time axis. Among them, the entire batch process is divided into multiple different stages, and the sub-model MPC controllers corresponding to different stages each use a sub-model as a prediction model, and the sub-model is a linear model, that is, the prediction model in the MPC controller includes multiple different sub-models. The controlled variables of the batch reactor include the temperature, pressure, etc. of the reactor. By designing the prediction model of the MPC controller as each sub-model corresponding to the different stages of the entire batch reaction process, the nonlinear time-varying problem of the batch reaction can be solved, thereby achieving precise control of a single batch.

[0055] Step 102, the outer layer uses an iterative learning controller to calculate the set trajectory in the MPC controller along the batch axis direction and output the set trajectory of the current batch, wherein for the current batch, the error between the output value of the previous batch of the batch reactor and the set trajectory of the previous batch is used as the input of the iterative learning controller.

[0056] Step 103 , the output of the iterative learning controller is used as the set trajectory of the current batch of the MPC controller for tracking control, thereby forming a double-layer control to continuously improve the product production quality of the batch reactor.

[0057] In this embodiment, Figure 2As shown, due to the repeatability of batch reactions, the outer layer uses iterative learning control (ILC) controller to calculate the set trajectory of the adjusted controlled variable according to the MPC control effect of the previous batch, so that the output value of the batch reactor (i.e., the controlled variable) continues to improve along the batch axis, further improving the control accuracy of the batch reaction and the product production quality.

[0058] Optionally, step 101: the inner layer uses an MPC controller to control the output value of a single batch of the batch reactor to track the set trajectory of the corresponding batch along the time axis, specifically including:

[0059] Step 1011, divide the entire batch reaction process into multiple different stages, and obtain sub-models corresponding to each different stage.

[0060] Step 1012: Use the sub-models corresponding to different stages as prediction models to design MPC controllers for each sub-model.

[0061] Step 1013, selecting a weighting function to globally model the entire batch reaction process.

[0062] Step 1014, switching each sub-model MPC controller according to the global modeling result of the entire batch reaction process, so as to achieve that the output value of a single batch of the batch reactor controlled by the MPC controller tracks the set trajectory of the corresponding batch along the time axis.

[0063] Optionally, step 1011: dividing the entire batch reaction process into a plurality of different stages, and obtaining each sub-model corresponding to each different stage, specifically includes: dividing the entire batch reaction process into a plurality of different stages in chronological order according to the operation trajectory of the scheduling variable and the nonlinear analysis of the process; performing identification experimental tests on each stage to obtain test data corresponding to each stage; using a prediction error or a subspace identification algorithm to identify the test data corresponding to each stage, and obtaining each sub-model G corresponding to each different stage. i ,i=1,...,S.

[0064] In this embodiment, according to the operation trajectory of the scheduling variables (such as the reaction time of the batch reactor, events, etc.) and the nonlinear analysis of the process, several different typical stages (such as Figure 2Then, an identification experiment is designed for the system dynamic characteristics of different typical stages. The design of the identification experiment specifically includes: determining the sampling period, using the generalized binary noise (GBN) signal as the identification excitation signal, and selecting 1 / 10 of the object transition time as the average switching time of the GBN signal. Figure 3 As shown in Figure 1, an identification software (such as FRONT, etc.) is used to identify the experimental data. The software automatically selects the optimal model order and calculates the identification model to obtain linear models at different stages, that is, each sub-model G i ,i=1,...,S. It should be noted that the entire batch reaction process is divided into a number of different typical stages, so as to design corresponding MPC controllers for different stages in the future, solve the nonlinear time-varying problem of the batch reaction process, and thus realize accurate control of the entire batch reaction process. The more typical stages the entire batch reaction process is divided into, the higher the control accuracy achieved.

[0065] Optionally, step 1013: selecting a weighting function to globally model the entire batch reaction process, specifically includes: selecting a weighting function according to the test data corresponding to each stage; using a nonlinear function of the scheduling variable to interpolate each sub-model to globally model the entire batch reaction process as follows:

[0066]

[0067] Among them, the nonlinear function of the scheduling variable is the stage switching logic.

[0068] In this embodiment, Figure 2 As shown, according to the change trajectory of the controlled variables of the batch reactor (such as reactor pressure, temperature, etc.), a single batch process is divided into several different typical stages, and the batch data is divided into multiple sub-data sets corresponding to each typical stage. Time / event is used as the scheduling variable, and the commonly used stage switching logic is selected as the nonlinear function of the scheduling variable to obtain the above-mentioned global modeling results. Specifically, according to the test data of the typical stage and the transition data between each typical stage, the nonlinear function of the scheduling variable (such as cubic spline function, polynomial function, etc.) is used to interpolate each local linear model (i.e., each sub-model) to perform large-scale global modeling. The parameters of the interpolation function are obtained by minimizing the sum of squares of the error between the predicted output value and the actual value. Since the constructed global model result includes the situation of switching each sub-model, it is closer to the actual process of the batch reaction process, and the constructed global model is more accurate, which makes the final batch reaction control effect better. As shown Figure 2The setting curve shown has different typical stages along the time axis, such as heating stage, heat preservation stage, heating stage, heat preservation stage, and cooling stage. According to the different typical stages divided according to the setting trajectory (i.e., setting curve) in the batch, the MPC controller of each sub-model is switched according to the global modeling results to complete the entire control process.

[0069] Optionally, step 1012: using the sub-models corresponding to different stages as prediction models to design each sub-model MPC controller, specifically including: each sub-model MPC controller uses the corresponding sub-model as the prediction model; designing the parameters of each sub-model MPC controller according to the control objectives of different stages, the parameters of the sub-model MPC controller include the controlled variable weighted matrix, the control increment weighted matrix, and the input and output constraints.

[0070] Specifically, Figure 4 The structure of the MPC controller including multiple sub-models (i.e., multi-model MPC controller) shown in FIG. designs the MPC controller parameters of each sub-model according to the control objectives at different stages, specifically including:

[0071] For each sub-model G i The MPC controller is specifically a DMC (dynamic matrix control) controller. Specifically, for asymptotically stable systems, a finite set of unit step response sampling data {a1, a2, …, a N} describes the dynamic characteristics of the system. The unit step vector of the system is a=[a1 a2…a N ] T is called the model vector of DMC, N is called the modeling domain, and a suitable N is selected so that i The value of (i>N) and the static final value a of the step response N The difference can be ignored.

[0072] According to the superposition principle of linear systems, the future output value of the system is predicted using the object unit step response model and the given input control increment. At time k, the predicted output value of the system under the control amount Δu(k) is as follows:

[0073]

[0074] In the formula

[0075]

[0076] in, It represents the output predicted value at time k for the next N time periods assuming that the control action remains unchanged. The symbol “~” in the formula represents prediction, and “k+i / k, i=1, 2,…, N” represents the prediction made at time k for time k+i.

[0077] Similarly, under the action of M continuous control increments Δu(k), Δu(k+1),…,Δu(k+M-1), the output prediction value of the controlled object at the next P moments is as follows:

[0078]

[0079] In the formula

[0080]

[0081] in It represents the output prediction value at the kth moment, assuming that the control action remains unchanged for the next P moments; the matrix A is a dynamic matrix, whose elements are the step response coefficients describing the dynamic characteristics of the system; P is the length of the rolling optimization time domain; M is the length of the control time domain, P and M should satisfy M≤P≤N, and the vector The first subscript of is the length of the predicted future output, and the second subscript is the number of changes in the control quantity.

[0082] At different sampling times, considering input and output constraints, the DMC optimization performance indicators are as follows:

[0083]

[0084] Stu min ≤u≤u max

[0085] y min ≤y≤y max

[0086] Among them, w P (k) = [w(k+1)…w(k+p)] T

[0087] Q=diag(q1…q P )

[0088] R=diag(r1…r M )

[0089] where w P (k) is the expected value vector, Q and R are the error weight matrix and control weight matrix respectively, and Q and R are diagonal matrices composed of weight coefficients.

[0090] Calculate the constrained DMC optimization performance index and obtain Δu(k),…,Δu(k+M-1). DMC takes the immediate control increment Δu(k) to form the actual control u(k)=u(k-1)+Δu(k) acting on the object. At the next moment, the same optimization problem is used to solve Δu(k+1) and obtain u(k+1) acting on the object.

[0091] Due to model errors, external interference, weak nonlinearity and other uncertain factors in the actual process, the predicted value of the system output will deviate from the actual value, that is, there is a prediction error as follows:

[0092]

[0093] Real-time information of prediction errors is used for feedback correction.

[0094] In this embodiment, advanced control software (such as Aspen DMCPlus, Zhejiang University FRONT-Suite, etc.) is used to design the MPC controller. First, the basic parameters of the MPC are set, the control cycle is selected, and the upper and lower bounds of the input and output constraints {u min ,u max ,y min ,y max}; Then according to the sub-models G of the controlled object i , i=1,...,S characteristics, the error weighting coefficient matrix Q of each linear DMC i And the control increment weighting coefficient matrix R i They are designed separately to adjust the closed-loop response speed. The error weighting coefficient matrix Q i Reflects the importance of different output variable errors; control increment weighting coefficient matrix R i The role of is to appropriately suppress the drastic changes in the control increment, and it is added to the performance index as a soft constraint.

[0095] Optionally, the iterative learning controller is specifically:

[0096] u k =u k-1 +Le k-1 , where e k =y d -y k

[0097] Among them, k is the number of iterations, e k is the deviation between the actual output and the expected trajectory during the k-th iteration; the matrix L is the gain matrix of the iterative learning algorithm,

[0098] The gain matrix L of the iterative learning algorithm is designed using the optimal iterative learning method, which significantly improves the convergence speed, that is,

[0099]

[0100] sty k =G·u k +y0

[0101] e k =y d -y k

[0102] Δu k =u k -u k-1

[0103] By solving the above optimization proposition, we get the following learning law u k =u k-1 +Le k-1 , where the gain matrix L = (G T QG+R) -1 G T Q.

[0104] like Figure 5 As shown, the control effect comparison between traditional PID and multi-model MPC+iterative learning is shown, among which the control effect of the multi-model MPC+iterative learning of this embodiment is closer to the set value. It can be seen that the two-layer control structure of MPC+iterative learning of this embodiment can make the tracking error smaller, that is, the control effect is better.

[0105] The control method of batch reaction in this embodiment uses multi-model MPC control in the inner layer, that is, in a single batch, the entire batch is divided into multiple different stages along the time axis according to the output tracking trajectory, and the MPC controller corresponding to each stage uses a sub-model as a prediction model. Then, the corresponding MPC controller parameters are designed according to the control objectives and adjustment means of each stage, and the controller is switched according to the stage changes within the batch, so as to achieve fast and smooth control of a single batch along the time axis. The outer layer uses an iterative learning controller to calculate the adjusted controlled variable setting trajectory according to the product process index requirements input by the user and the MPC control effect of the previous batch, so as to achieve continuous improvement of product quality along the batch axis and improve product quality.

[0106] Embodiment 2:

[0107] This embodiment provides a batch reaction control device, including an inner layer control device, an outer layer control device and a batch reactor.

[0108] Batch reactor, used to complete batch reaction process.

[0109] The inner layer control device is connected to the batch reactor and is used to control the output value of a single batch of the batch reactor to track the set trajectory of the corresponding batch along the time axis using an MPC controller, wherein the MPC controller includes multiple sub-model MPC controllers.

[0110] The outer control device is connected to the batch reactor and the inner control device, and is used to use an iterative learning controller to calculate the set trajectory in the MPC controller of the inner control device along the batch axis direction and output the set trajectory of the current batch, wherein, for the current batch, the error between the output value of the previous batch of the batch reactor and the set trajectory of the previous batch is used as the input of the iterative learning controller, and the output of the iterative learning controller is used as the set trajectory of the current batch of the MPC controller of the inner control device for tracking control.

[0111] Optionally, the inner control device includes an acquisition unit, a design unit, a modeling unit and a switching unit.

[0112] The acquisition unit is used to divide the entire batch reaction process into a plurality of different stages and acquire the sub-models corresponding to the different stages. The design unit is connected to the acquisition unit and is used to design the MPC controllers of each sub-model by using the sub-models corresponding to the different stages as prediction models. The modeling unit is connected to the acquisition unit and is used to select a weighting function to globally model the entire batch reaction process. The switching unit is connected to the modeling unit and the design unit and is used to switch the MPC controllers of each sub-model according to the global modeling results of the entire batch reaction process, so as to realize that the output value of a single batch of the MPC controller controlling the batch reactor tracks the set trajectory of the corresponding batch along the time axis.

[0113] Optionally, the acquisition unit includes a partitioning component, a testing component and an acquisition component.

[0114] The partitioning component is used to divide the entire batch reaction process into multiple different stages in chronological order according to the operation trajectory of the scheduling variables and the nonlinear analysis of the process. The test component is connected to the partitioning component and is used to perform identification experiments on each stage to obtain the test data corresponding to each stage. The acquisition component is connected to the test component and is used to identify the test data corresponding to each stage using the prediction error or subspace identification algorithm to obtain the sub-models G corresponding to each different stage. i ,i=1,...,S.

[0115] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present invention, but the present invention is not limited thereto. For those of ordinary skill in the art, various modifications and improvements can be made without departing from the spirit and essence of the present invention, and these modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A batch reaction control method, characterized in that: include: The inner layer uses an MPC controller to control the output value of a single batch of the batch reactor to track the set trajectory of the corresponding batch along the time axis, wherein the MPC controller includes multiple sub-model MPC controllers; The outer layer uses an iterative learning controller to calculate the set trajectory in the MPC controller along the batch axis direction and output the set trajectory of the current batch, wherein for the current batch, the error between the output value of the previous batch of the batch reactor and the set trajectory of the previous batch is used as the input of the iterative learning controller; The output of the iterative learning controller is used as the set trajectory of the current batch of the MPC controller for tracking control, thus forming a double-layer control to continuously improve the product production quality of the batch reactor. The inner layer uses an MPC controller to control the output value of a single batch of the batch reactor to track the set trajectory of the corresponding batch along the time axis, specifically including: The entire batch reaction process is divided into multiple different stages, and the sub-models corresponding to each different stage are obtained; The sub-models corresponding to different stages are used as prediction models to design the MPC controllers for each sub-model; Selecting a weighting function to globally model the entire batch reaction process; The MPC controllers of each sub-model are switched according to the global modeling results of the entire batch reaction process, so that the output value of a single batch of the MPC controller controlling the batch reactor tracks the set trajectory of the corresponding batch along the time axis. Global modeling meets: Among them, the nonlinear function of the scheduling variable is the stage switching logic.

2. The batch reaction control method according to claim 1, characterized in that: The entire batch reaction process is divided into a plurality of different stages, and each sub-model corresponding to each different stage is obtained, specifically including: According to the operation trajectory of the scheduling variables and the nonlinear analysis of the process, the entire batch reaction process is divided into multiple different stages in chronological order; Carry out identification experiments and tests on each stage to obtain the corresponding test data of each stage; The prediction error or subspace identification algorithm is used to identify the test data corresponding to each stage, and the sub-models G corresponding to different stages are obtained. i ,i=1,...,S.

3. The batch reaction control method according to claim 2, characterized in that: The weighting function is selected to globally model the entire batch reaction process, specifically including: Select a weighting function based on the test data corresponding to each stage; The sub-models are interpolated with nonlinear functions of the scheduling variables to globally model the entire batch reaction process.

4. The batch reaction control method according to claim 1, characterized in that: The method of using the sub-models corresponding to different stages as prediction models to design the MPC controllers of each sub-model specifically includes: Each sub-model MPC controller uses the corresponding sub-model as the prediction model; The parameters of each sub-model MPC controller are designed according to the control objectives of different stages. The parameters of the sub-model MPC controller include the controlled variable weighted matrix, the control increment weighted matrix, and the input and output constraints.

5. The batch reaction control method according to claim 4, characterized in that: The MPC controller parameters of each sub-model are designed according to the control objectives of different stages, specifically including: For each sub-model G i The MPC controller is specifically a dynamic matrix control DMC controller. Specifically, for asymptotically stable systems, a finite set of unit step response sampling data {a1, a2, …, a N } describes the dynamic characteristics of the system. The unit step vector of the system is a=[a1 a2 … a N ] T is called the model vector of DMC, N is called the modeling domain, and a suitable N is selected so that i The value of (i>N) and the static final value a of the step response N The difference can be ignored. According to the superposition principle of linear systems, the future output value of the system is predicted using the object unit step response model and the given input control increment. At time k, the predicted output value of the system under the control amount Δu(k) is as follows: In the formula in, It represents the output prediction value at time k for the next N time periods assuming that the control action remains unchanged. The symbol "~" in the formula represents prediction, and "k+i / k,i=1,2,…,N" represents the prediction made at time k for time k+i. Similarly, under the action of M continuous control increments Δu(k), Δu(k+1), …, Δu(k+M-1), the output prediction value of the controlled object at the next P moments is as follows: In the formula in It represents the output prediction value at the kth moment, assuming that the control action remains unchanged for the next P moments; the matrix A is a dynamic matrix, whose elements are the step response coefficients describing the dynamic characteristics of the system; P is the length of the rolling optimization time domain; M is the length of the control time domain, P and M should satisfy M≤P≤N, and the vector The first subscript is the length of the predicted future output, and the second subscript is the number of control changes. At different sampling times, considering input and output constraints, the DMC optimization performance indicators are as follows: s.t.u min ≤u≤u max and min ≤y≤y max where, w P (k)=[w(k + 1)…w(k + p)] T Q=diag(q1…q P ) R=diag(r1…r M ) where w P (k) is the expected value vector, Q and R are the error weight matrix and the control weight matrix respectively, Q and R are diagonal matrices consisting of weight coefficients, Calculate the constrained DMC optimization performance index and obtain Δu(k), ..., Δu(k+M-1). DMC takes the immediate control increment Δu(k) to form the actual control u(k)=u(k-1)+Δu(k) acting on the object. At the next moment, solve Δu(k+1) through the same optimization problem and obtain u(k+1) acting on the object. When the predicted value output by the system deviates from the actual value, there is a prediction error as follows: Real-time information of prediction errors is used for feedback correction.

6. The batch reaction control method according to claim 1, characterized in that: The iterative learning controller is specifically: u k = u k-1 + Le k-1 , where e k = y d - y k Among them, k is the number of iterations, e k is the deviation between the actual output and the expected trajectory during the k-th iteration; the matrix L is the gain matrix of the iterative learning algorithm, The gain matrix L of the iterative learning algorithm is designed using the optimal iterative learning method, that is, s.t.y k =G·u k +y0 and k =and d -and k Δu k =u k -u k-1 Among them, G is the steady-state gain of the controlled object, Q and R are the error weight matrix and the control weight matrix respectively. By solving the above optimization proposition, the following learning law is obtained: k =u k-1 +Le k-1 , where the gain matrix L = (G T QG+R) -1 G T Q.

7. A batch reaction control device, characterized in that: It includes an inner control device, an outer control device and a batch reactor, Batch reactor, used to complete the batch reaction process, The inner layer control device is connected to the batch reactor and is used to control the output value of a single batch of the batch reactor to track the set trajectory of the corresponding batch along the time axis direction using an MPC controller, wherein the MPC controller includes a plurality of sub-model MPC controllers. The outer layer control device is connected to the batch reactor and the inner layer control device, and is used to calculate the set trajectory in the MPC controller of the inner layer control device along the batch axis direction and output the set trajectory of the current batch using an iterative learning controller, wherein, for the current batch, the error between the output value of the previous batch of the batch reactor and the set trajectory of the previous batch is used as an input of the iterative learning controller, and the output of the iterative learning controller is used as the set trajectory of the current batch of the MPC controller of the inner layer control device for tracking control, The inner control device includes an acquisition unit, a design unit, a modeling unit and a switching unit. The acquisition unit is used to divide the entire batch reaction process into multiple different stages and obtain the sub-models corresponding to each different stage. The design unit is connected to the acquisition unit and is used to design the MPC controllers of each sub-model by taking the sub-models corresponding to different stages as prediction models. A modeling unit, connected to the acquisition unit, is used to select a weighting function to globally model the entire batch reaction process. The switching unit is connected to the modeling unit and the design unit, and is used to switch each sub-model MPC controller according to the global modeling result of the entire batch reaction process, so as to realize that the output value of a single batch of the MPC controller controlling the batch reactor tracks the set trajectory of the corresponding batch along the time axis direction. Global modeling meets: Among them, the nonlinear function of the scheduling variable is the stage switching logic.

8. The batch reaction control device according to claim 7, characterized in that: The acquisition unit includes a partition component, a test component and an acquisition component. The partitioning component is used to divide the entire batch reaction process into multiple different stages in chronological order according to the operation trajectory of the scheduling variables and the nonlinear analysis of the process. The test component is connected to the partition component to perform identification experiments on each stage and obtain the test data corresponding to each stage. The acquisition component is connected to the test component and is used to identify the test data corresponding to each stage using the prediction error or subspace identification algorithm to obtain the sub-models G corresponding to different stages. i ,i=1,...,S.

Images