DMO rectification temperature model prediction control method based on LSTM feature extraction and decreasing prediction step length
Through the DMO distillation temperature model prediction control method of LSTM feature extraction and decreasing prediction step size, the problems of multivariable and strong coupling in the distillation tower are solved, and more efficient control and energy consumption optimization are achieved, ensuring product quality and system Stablize.
Patent Information
- Application Number
- CN202411147610.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-21
- Publication Date
- 2024-11-29
AI Technical Summary
The existing distillation tower control technology is difficult to effectively deal with the problems of multivariable, strong coupling and control input and output constraints, which makes it difficult to achieve product quality and energy consumption optimization.
The DMO distillation temperature model prediction control method based on LSTM feature extraction and decreasing prediction step length is adopted. Data processing and mechanism modeling are performed through the LSTM time series model, combined with the H∞ type cost function and the decreasing step length model prediction control algorithm, the system is optimized Status and control inputs.
Improves model accuracy and control time accuracy of system state, reduces the time when the system enters the terminal set, and optimizes the energy consumption and product quality of the device.
Smart Images

Figure CN119046756A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to a DMO distillation temperature model prediction control method based on LSTM feature extraction and decreasing prediction step used in a distillation tower distillation process, and belongs to the technical field of control of chemical distillation processes. Background Art
[0002] The coal-to-ethylene glycol process is immature, complex, and prone to side reactions, resulting in excessive impurities in the ethylene glycol product and difficulty in separation. Distillation is one of the most important separation operations in the coal-to-ethylene glycol process, but the distillation process has complex control characteristics, such as large hysteresis, slow dynamic response, high model order, severe nonlinearity, multiple control variables and strong coupling. However, most current studies focus on traditional PID control in the distillation section, without considering the strong coupling and mutual influence of multiple variables in the distillation tower; therefore, the use of PID controllers alone can no longer meet the requirements of modern industrial production for product quality accuracy. Therefore, in order to solve the problems of multiple variables, strong coupling, and control input and output constraints in the production process, researchers have adopted advanced control technology with model predictive control as the core.
[0003] At present, the problems of existing predictive control technology are:
[0004] (1) Traditional data processing methods mainly rely on expert experience and historical data reference. Although this data processing method is relatively fast, its efficiency and accuracy are difficult to guarantee, and the accuracy of the model cannot be guaranteed.
[0005] (2) Fixed-step predictive control can enable the system state to reach the terminal set within a finite time. However, the specific arrival time cannot be predicted, and it cannot be guaranteed to arrive within the specified time.
[0006] From the above, we can see that the introduction of the LSTM time series model can process the data effectively and quickly, solving the problem that the traditional data method still has small deviations, making the established model more accurate.
[0007] Next, the decreasing step size predictive control method is used to guide the system state into the terminal constraint set through the variable step size control strategy, which solves the problem that the time to reach the terminal set is difficult to predict and cannot be guaranteed to arrive within the specified time using the traditional model predictive control method. This allows the system state to enter the terminal within a limited number of steps or even a specified number of steps. Therefore, this control method has important theoretical and practical significance. Summary of the invention
[0008] The purpose of the present invention is to provide a DMO distillation temperature model predictive control method based on LSTM feature extraction and decreasing prediction step size, so as to solve the problems of multivariable, strong coupling and control input and output constraints in the production process, and optimize the energy consumption, product yield, product quality and adaptability of the device to changes in device load.
[0009] To achieve the above object, the technical solution adopted by the present invention is: a DMO distillation temperature model predictive control method based on LSTM feature extraction and decreasing prediction step, which comprises the following steps:
[0010] S1: Collect chemical industry data and use LSTM time series model to extract and preprocess data features;
[0011] S2: Perform mechanism modeling on the data obtained in step S1 to determine the system model of predictive control;
[0012] S3: Using H ∞ Type cost function, which determines the weight of system state or control input by adjusting the parameter matrix of the cost function;
[0013] S4: Use the model predictive control algorithm with decreasing step size to achieve the minimization optimization of the cost function and reduce the control time of the overall system.
[0014] Furthermore, in step S1, LSTM time series model prediction is introduced, and the specific steps are as follows:
[0015] S1.1: Measure and compile historical data of the input indicators of the model built on site to form input variables, take temperature data at different times and in different environments as data sets and perform data preprocessing;
[0016] S1.2: The data is normalized using the maximum and minimum values and mapped to the range [0,1]. The normalization formula is:
[0017]
[0018] Where x * is the normalized data; x min and x max are the minimum and maximum values of the sample data set respectively; x is the original sample data;
[0019] S1.3: Divide the dataset into training set and test set;
[0020] S1.4: Determine the number of LSTM input nodes and output nodes according to the number of input variables and output variables, set the number of hidden layer nodes and learning rate, and establish an LSTM prediction model;
[0021] S1.5: Train the established LSTM temperature prediction model through the training set samples to obtain the optimal LSTM prediction model parameters that minimize the least squares error E between the predicted temperature and the actual temperature; where:
[0022]
[0023] Where n represents the sample size of the training set; X act (i) and X pred (i) is the actual value and predicted value of the temperature at the i-th moment;
[0024] S1.6: After obtaining the LSTM temperature prediction model with the optimal parameters, predict the test set samples to obtain the predicted temperature value to be predicted, and evaluate the accuracy of the prediction results according to the evaluation indicators.
[0025] Furthermore, after obtaining the prediction results, the prediction results are converted back to the scale of the original data through a denormalization operation to facilitate interpretation and use in subsequent steps.
[0026] Furthermore, in step S2, the Hankel matrix of the model is constructed, and the system equation is solved using the least squares method. The specific steps are as follows:
[0027] S2.1: First, the Hankel matrix is established based on the collected input and output data. is the state vector of the system, is the input vector of the system, is the output vector of the system, is the process noise vector k=1,2,…,M,…,N; the input matrix established is as follows:
[0028]
[0029] Where p represents the past and f represents the future, that is, U p represents the past input matrix, U f represents the future input matrix;
[0030] S2.2: Similarly, the output Hankel matrix and the perturbation Hankel matrix are defined as follows:
[0031]
[0032] Where p represents the past and f represents the future, that is, Y p represents the past output matrix, Y f represents the future output matrix; and E p represents the past noise matrix, E f represents the future noise matrix;
[0033] S2.3: Then, the state variables in the equation are iterated starting from time t=k:
[0034] x(k+1)=Ax(k)+Bu(k)+Ke(k)
[0035]
[0036] S2.4: Let k = k+1, k+2, ... k+N-1, and combine the equations in columns, that is:
[0037]
[0038] S2.5: Finally, the regression method and least squares method are used to solve the system matrix and obtain the state space model of the system.
[0039] Furthermore, in step S3, the objective function of the established model predictive controller is:
[0040]
[0041] in
[0042]
[0043] k is the sampling time, i is 0-N k The middle natural number, N k To predict the step length, Q and R are positive definite matrices, and the weights given to the system state and control input can be determined by adjusting Q and R; τ is a positive constant, the smaller its value, the better the closed-loop system control performance, and the appropriate τ value is calculated when the state terminal set is selected. The above weights can be freely allocated and determined according to actual conditions.
[0044] Furthermore, in step S3, the state of the system at time k is defined, x(k) is the state vector of the system, u(k) is the input vector of the system, e(k) is the process noise vector, and the input and interference sequences are as follows:
[0045] x(k)=[x(k),x(k+1∣k),…,x(k+N k ∣k)] T
[0046] u(k)=[u(k),u(k+1|k),…,u(k+N k -1|k)] T
[0047] e(k)=[e(k),e(k+1∣k),…,e(k+N k -1|k)] T
[0048] Then the system state can be described by the matrix:
[0049] x(k+1)=Ax(k)+Bu(k)+Ce(k).
[0050] They are the state matrix, input matrix and Kalman gain matrix of the system respectively.
[0051] Furthermore, in step S4, the specific steps of the process of the decreasing step size model predictive control algorithm are as follows:
[0052] S4.1: Select the system state input and output and the system parameter values;
[0053] S4.2: At time k, detect the system state x(k). If the system state has not ended, stop the calculation, otherwise set the prediction step size N k is k-1, and proceed to step 4.3;
[0054] S4.3: Solve the optimization problem, then apply the first element of the optimization control input sequence to the system and proceed to step 4.4;
[0055] S4.4: Wait for sampling time k+1 to arrive and repeat step S4.2.
[0056] Compared with the prior art, the present invention has the following technical effects:
[0057] The present invention adopts the LSTM time series model and prevents the gradient vanishing problem through a gating mechanism, thereby ensuring the stability of the training process and being able to effectively handle long-term dependency problems.
[0058] Based on variable step-size predictive control, the present invention regards the arrival time as a hard constraint and proposes a decreasing step-size model predictive control, which ensures that the system state can enter the terminal set within the specified time and also reduces the time it takes for the system to enter the terminal set.
[0059] The present invention uses H while guiding the system status. ∞ The model cost function optimizes system performance, ensuring that the system achieves its intended purpose while meeting control requirements and system stability. BRIEF DESCRIPTION OF THE DRAWINGS
[0060] Figure 1 It is a schematic diagram of the control scheme of the DMO distillation tower of the present invention.
[0061] Figure 2 It is a structural diagram of the basic unit composition of the LSTM network of the present invention.
[0062] Figure 3It is a flow chart of the LSTM temperature prediction method of the present invention.
[0063] Figure 4 It is the overall system design flow chart of the present invention.
[0064] Figure 5 It is a schematic diagram of the predictive control principle of the decreasing step size model of the present invention. DETAILED DESCRIPTION
[0065] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0066] A DMO distillation temperature model predictive control method based on LSTM feature extraction and decreasing prediction step size includes the following steps:
[0067] Step 1: Collect chemical industry data and use the LSTM time series model to extract and preprocess data features.
[0068] Step 1.1: Measure and count the historical data of the input indicators of the model built on site to form input variables. Take the temperature data at different times and in different environments as the data set and perform data preprocessing to make the experiment more convincing.
[0069] Step 1.2: Use the maximum and minimum normalization to process the sample data and map it to [0,1]. The normalization formula is:
[0070]
[0071] Where x * is the normalized data; x min and x max are the minimum and maximum values of the sample data set respectively; x is the original sample data;
[0072] Step 1.3: Divide the sample data set into training set and test set;
[0073] Step 1.4: Determine the number of LSTM input nodes and output nodes according to the number of input variables and output variables, set the appropriate number of hidden layer nodes and learning rate, and establish the LSTM prediction model;
[0074] Step 1.5: Train the established LSTM temperature prediction model through training samples to obtain the optimal LSTM prediction model parameters that minimize the least squares error E between the predicted temperature and the actual temperature.
[0075]
[0076] Where n represents the sample size of the training set; X act (i) and X pred(i) is the actual value and predicted value of the temperature at the i-th moment;
[0077] Step 1.6: After obtaining the LSTM temperature prediction model with the optimal parameters, predict the test set samples to obtain the predicted temperature value to be predicted, and evaluate the accuracy of the prediction results according to the evaluation indicators.
[0078] Of course, after obtaining the prediction results, it is necessary to convert the prediction results back to the scale of the original data through a denormalization operation to facilitate interpretation and use in the following steps.
[0079] Step 2: Based on the prediction results, the distillation tower is modeled to obtain its system model.
[0080] For the model to construct its Hankel matrix, the least squares method is used to solve the system equations. The specific steps are as follows:
[0081] Step 2.1: First, a Hankel matrix is established based on the collected input and output data of the controlled system. The established input matrix is as follows:
[0082]
[0083] Step 2.2: Similarly, the output Hankel matrix and the perturbation Hankel matrix can be defined as follows:
[0084]
[0085] Step 2.3: Next, iterate the state variables in the equation starting from time t=k:
[0086] x(k+1)=Ax(k)+Bu(k)+Ke(k)
[0087]
[0088] Step 2.4: Let k = k+1, k+2, ... k+N-1, and combine the equations in columns, that is:
[0089]
[0090] Step 2.5: Finally, the system matrix is solved using regression method and least square method to obtain the state space model of the system.
[0091] The objective function of the established model predictive controller is:
[0092]
[0093] in
[0094]
[0095] Q and R are positive definite matrices, and the weights given to the system state and control input can be determined by adjusting Q and R; τ is a positive constant, the smaller its value, the better the closed-loop system control performance, and the appropriate τ value is calculated when the state terminal set is selected.
[0096] Step 3: Use H ∞ The cost function is a type of cost function, which determines the weight of the system state or control input by adjusting the parameter matrix of the cost function.
[0097] First, define the state, input and interference sequence of the system at time k as follows:
[0098] x(k)=[x(k),x(k+1∣k),…,x(k+N k ∣k)] T
[0099] u(k)=[u(k),u(k+1|k),…,u(k+N k -1|k)] T
[0100] e(k)=[e(k),e(k+1∣k),…,e(k+N k -1|k)] T
[0101] Then the system state can be described by the matrix
[0102] x(k+1)=Ax(k)+Bu(k)+Ce(k).
[0103] Step 4: Use the model predictive control algorithm with decreasing prediction step size to achieve the minimization optimization of the cost function and reduce the control time of the overall system.
[0104] The specific steps of the process of the decreasing step size model predictive control algorithm are as follows:
[0105] S4.1: Select the system state input and output and the system parameter values;
[0106] S4.2: At time k, detect the system state x(k). If the system state has not ended, stop the calculation, otherwise set the prediction step size N k is k-1, and proceed to step 4.3;
[0107] S4.3: Solve the optimization problem, then apply the first element of the optimization control input sequence to the system and proceed to step 4.4;
[0108] S4.4: Wait for sampling time k+1 to arrive and repeat S4.2.
[0109] In the following examples, a discrete-time linear system model is considered in the specific implementation scheme:
[0110] x(k+1)=Ax(k)+Bu(k)+Ce(k)
[0111] in, C=|1|
[0112] Consider system input constraints and disturbances:
[0113] The H∞ type cost function is used to optimize the system performance and determine the cost function parameters. The values of Q, R, and τ are respectively 0.1 and 2.
[0114] It is worth mentioning that at each moment the computer collects the system status and first determines whether the system status enters the terminal set. If so, the optimization is stopped; if not, the prediction step size is reduced by 1 step, and the first element of the control input sequence is applied to the system for the next sampling.
[0115] For different initial states, the patented method can guide the system state to enter the terminal set within a limited step size, while the fixed step size method fails. Due to the decreasing prediction step size, the execution time of the patented method is also significantly reduced.
[0116] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art should understand that the above embodiments do not limit the protection scope of the present invention in any form, and all technical solutions obtained by equivalent replacement and the like fall within the protection scope of the present invention. The parts not involved in the present invention are the same as the prior art or can be implemented by the prior art.
Claims
1. A DMO distillation temperature model predictive control method based on LSTM feature extraction and decreasing prediction step size, Features The steps include: S1: Collect chemical industry data and use LSTM time series model to extract and preprocess data features; S2: Perform mechanism modeling on the data obtained in step S1 to determine the system model of predictive control; S3: Using H ∞ Type cost function, which determines the weight of system state or control input by adjusting the parameter matrix of the cost function; S4: Use the model predictive control algorithm with decreasing step size to achieve the minimization optimization of the cost function and reduce the control time of the overall system.
2. A DMO distillation temperature model predictive control method based on LSTM feature extraction and decreasing prediction step according to claim 1, It is characterized in that In step S1, LSTM time series model prediction is introduced, and the specific steps are as follows: S1.1: Measure and compile historical data of the input indicators of the model built on site to form input variables, take temperature data at different times and in different environments as data sets and perform data preprocessing; S1.2: The data is normalized using the maximum and minimum values and mapped to the range [0,1]. The normalization formula is: Where x * is the normalized data; x min and x max are the minimum and maximum values of the sample data set respectively; x is the original sample data; S1.3: Divide the dataset into training set and test set; S1.4: Determine the number of LSTM input nodes and output nodes according to the number of input variables and output variables, set the number of hidden layer nodes and learning rate, and establish an LSTM prediction model; S1.5: Train the established LSTM temperature prediction model through the training set samples to obtain the optimal LSTM prediction model parameters that minimize the least squares error E between the predicted temperature and the actual temperature; where: Where n represents the sample size of the training set; X act (i) and X pred (i) is the actual value and predicted value of the temperature at the i-th moment; S1.6: After obtaining the LSTM temperature prediction model with the optimal parameters, predict the test set samples to obtain the predicted temperature value to be predicted, and evaluate the accuracy of the prediction results according to the evaluation indicators.
3. A DMO distillation temperature model predictive control method based on LSTM feature extraction and decreasing prediction step according to claim 2, It is characterized in that After obtaining the prediction results, the prediction results are converted back to the scale of the original data through the denormalization operation to facilitate interpretation and use in the subsequent steps.
4. A DMO distillation temperature model predictive control method based on LSTM feature extraction and decreasing prediction step according to claim 1, It is characterized in that In step S2, the Hankel matrix of the model is constructed, and the system equation is solved using the least squares method. The specific steps are as follows: S2.1: First, the Hankel matrix is established based on the collected input and output data. is the state vector of the system, is the input vector of the system, is the output vector of the system, is the process noise vector, k=1,2,…,M,…,N; the input matrix established is as follows: Where p represents the past and f represents the future, that is, U p represents the past input matrix, U f represents the future input matrix; S2.2: Similarly, the output Hankel matrix and the perturbation Hankel matrix are defined as follows: Where p represents the past and f represents the future, that is, Y p represents the past output matrix, Y f represents the future output matrix; and E p represents the past noise matrix, E f represents the future noise matrix; S2.3: Then, the state variables in the equation are iterated starting from time t=k: x(k+1)=Ax(k)+Bu(k)+Ke(k) S2.4: Let k = k+1, k+2, ... k+N-1, and combine the equations in columns, that is: S2.5: Finally, the regression method and least squares method are used to solve the system matrix and obtain the state space model of the system.
5. A DMO distillation temperature model predictive control method based on LSTM feature extraction and decreasing prediction step according to claim 1, It is characterized in that In step 3, the objective function of the established model predictive controller is: in k is the sampling time, i is 0-N k The middle natural number, N k is the prediction step length; Q and R are positive definite matrices, and the weights given to the system state and control input are determined by adjusting Q and R; τ is a positive constant, the smaller its value is, the better the control performance of the closed-loop system is. The appropriate τ value is calculated when the state terminal set is selected.
6. A DMO distillation temperature model predictive control method based on LSTM feature extraction and decreasing prediction step according to claim 1, It is characterized in that In step S3, the state of the system at time k is defined, x(k) is the state vector of the system, u(k) is the input vector of the system, e(k) is the process noise vector, and the input and interference sequences are as follows: x(k)=[x(k),x(k+1∣k),…,x(k+N k ∣k)] T u(k)=[u(k),u(k+1∣k),…,u(k+N k -1∣k)] T e(k)=[e(k),e(k+1∣k),…,e(k+N k -1∣k)] T Then the system state can be described by the matrix: x(k+1)=Ax(k)+Bu(k)+Ce(k).
7. A DMO distillation temperature model predictive control method based on LSTM feature extraction and decreasing prediction step according to claim 1, It is characterized in that In step S4, the specific steps of the process of the decreasing step size model predictive control algorithm are as follows: S4.1: Select the system state input and output and the system parameter values; S4.2: At time k, detect the system state x(k). If the system state has not ended, stop the calculation, otherwise set the prediction step size N k is k-1, and proceed to step 4.3; S4.3: Solve the optimization problem, then apply the first element of the optimization control input sequence to the system and proceed to step 4.4; S4.4: Wait for sampling time k+1 to arrive and repeat step S4.2.
Citation Information
Cited By
LSTM-based rectification tower sensitive plate temperature model prediction control method and device, processor and storage medium thereof
CN119758716A
