Design method of subspace predictive controller with variable window length

Abstract

The invention relates to a design method of a subspace predictive controller with a variable window length, and relates to the technical field of industrial process control, and the method comprises the steps: constructing a control right matrix; acquiring operation data at the current moment; inputting a control power matrix according to the operation data type to obtain the operation data after power adjustment; constructing a Hankel matrix according to the operation data after weight adjustment and the length value of the current rolling window; the historical operation data and the input data at the current moment are substituted into the subspace prediction model to obtain an output optimal prediction value; solving a difference between the output optimal predicted value and the current output value through an error calculation algorithm to obtain a correction deviation; respectively judging the window correction deviation with an upper limit deviation reference and a lower limit deviation reference, and adjusting the length value of the corresponding rolling window; and waiting for new operation data and returning the new operation data to the operation data after weight adjustment. The method has the effect of enabling the obtained model to accurately reflect the current system characteristics.

Description

Technical Field

[0001] The present application relates to the technical field of industrial process control, and in particular to a design method for a subspace predictive controller with a variable window length. Background Art

[0002] Industrial process control refers to the automatic control of production processes that are continuous or carried out according to a certain periodic procedure in industrial production such as petroleum, chemical industry, shipbuilding, nuclear energy, electricity, water conservancy, metallurgy, light industry, textile, pharmaceutical, building materials, and environmental engineering. Industrial process control is an important branch of the field of automation.

[0003] In the related art, in the process of industrial process control, subspace predictive controllers are generally used to effectively control nonlinear, time-varying and multivariable industrial processes. Subspace predictive controllers process data through data preprocessing, feature extraction, dimensionality reduction and other steps, thereby retaining the main features of the data and simplifying the calculation and processing of the data. Then, the input and output data are processed and predictively controlled through a neural network to obtain the optimal control quantity.

[0004] With respect to the above-mentioned related technologies, the applicant has found the following defects: when using the subspace predictive controller for control, the input and output of the industrial process control are easily disturbed by variables such as hardware failure, valve position saturation or direct intervention of operators, so that the obtained model is not easy to accurately reflect the current system characteristics. Summary of the invention

[0005] In order to make the obtained model accurately reflect the current system characteristics, the present application provides a subspace prediction controller design method with variable window length.

[0006] In a first aspect, the present application provides a method for designing a subspace prediction controller with a variable window length, which adopts the following technical solution: A method for designing a subspace predictive controller with a variable window length, comprising: The weight matrix construction step is to construct the control weight matrix; Data acquisition step, obtaining the current operating data; Data weight adjustment step, inputting the control weight matrix according to the type of operating data to obtain the weighted operating data; Hankel matrix construction step, constructing the Hankel matrix according to the weighted operation data and the current rolling window length value; output prediction step, bringing the historical operation data and the input data at the current moment into the preset subspace prediction model to obtain the output optimal prediction value; Deviation calculation step, the output optimal prediction value and the output value at the current moment are calculated by a preset error calculation algorithm to obtain a corrected deviation; Window correction step: if the window correction deviation is greater than the preset upper limit deviation benchmark, the corresponding rolling window length value is reduced by one step; if the window correction deviation is less than the preset lower limit deviation benchmark, the corresponding rolling window length value is increased by one step; loop configuration step, wait for new operating data, and return to data weight adjustment step.

[0007] Optional, weight matrix construction step, control weights include primary variables and secondary variables; The data weighting step is to input the main variables in the operating data into the main variable control weight matrix, and input the secondary variables in the operating data into the secondary variable control weight matrix.

[0008] Optionally, the subspace prediction model is configured as Among them, L=[L1L2] is the corresponding subspace matrix, which is obtained based on the linear relationship model through a preset problem decomposition algorithm.

[0009] Optionally, the linear relationship model is configured as xk +1 =Axk+Buk+kf*ek,yk=Cxk+Duk+e k .

[0010] Optionally, the problem decomposition algorithm is: Constructing a least squares problem for solving the optimal prediction value; The linear algebra tools RQ decomposition and SVD decomposition are used to solve the above least squares problem.

[0011] Optionally, the error calculation algorithm is: Assume that the prediction error of the system at the moment is The correction deviation is calculated according to the preset error calculation formula:

[0012] The old data are replaced by the new data, and the optimal control law is obtained by minimizing the quadratic objective function.

[0013] In summary, the present application includes at least one of the following beneficial technical effects: 1. By constructing the control weight matrix and acquiring the current operating data, the operating data is weighted, and the Hankel matrix is ​​constructed through the weighted operating data and the current rolling window length value. The historical operating data and the current input data are then brought into the preset subspace prediction model to obtain the output optimal prediction value. The output optimal prediction value and the output value at the current moment are then subtracted by the preset error calculation algorithm to obtain the correction deviation, and the rolling window length value is adjusted according to the correction deviation, so that the obtained model accurately reflects the current system characteristics; BRIEF DESCRIPTION OF THE DRAWINGS

[0014] Figure 1 It is a flow chart of the method for designing a subspace prediction controller with a variable window length according to an embodiment of the present application.

[0015] Figure 2 This is a diagram of input-output simulation results obtained by the subspace prediction controller design method without using a variable window length in an embodiment of the present application.

[0016] Figure 3 This is a diagram of input-output simulation results obtained by the subspace prediction controller design method with variable window length in an embodiment of the present application. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of this application more clear, the following Figure 1-3 It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0018] An embodiment of the present application discloses a method for designing a subspace prediction controller with a variable window length.

[0019] Reference Figure 1 , a design method of a subspace prediction controller with a variable window length includes: Step S100: construct a control right matrix.

[0020] Among them, control rights include main variables and secondary variables. The main variable control rights matrix refers to the information used to indicate the matrix construction of the main variables at the current time, and the secondary variable control rights matrix refers to the information used to indicate the matrix construction of the secondary variables at the current time. By qualitatively analyzing the input and output variables of the process, the control rights matrix is ​​constructed to facilitate subsequent use.

[0021] Step S200: Acquire the current operation data.

[0022] The operating data at the current moment refers to the operating data information collected under normal operating conditions of the industrial process at the current moment, and the operating data at the current moment is obtained after collection.

[0023] Step S300: inputting a control weight matrix according to the type of operating data to obtain weighted operating data.

[0024] Among them, the weighted operating data refers to the data information used to indicate the weight adjustment of the operating data. By inputting the main variables in the operating data into the main variable control weight matrix, and the secondary variables in the operating data into the secondary variable control weight matrix, the main variable control weight matrix q and the secondary variable and interference control weight matrix p are constructed respectively. q and p as weight coefficients reflect the influence of the main and secondary variables and interference control effects on the controller at different times. In this way, the influence of unmeasurable interference noise and secondary variables on the performance of predictive control weight can be reduced by strengthening the main control weight q and weakening the secondary variables and error weight p. According to the mechanism analysis, multiple control weight matrices can be added to the input variables to reflect the influence of different variables on the controller.

[0025] It is also possible to add multiple control weight matrices to the input variables through mechanism analysis to reflect the impact of different variables on the controller. For example, the input at each time point is defined as: u = [α i u i Lβ j u j ], u is the input at the time of adoption, u i is the input of the i-th row, α i is the control weight matrix of the input in row i, u j is the input of the jth column, β j is the control weight matrix of the j-th column input, L is the input between the i-th row input and the j-th column input and the control weight matrix.

[0026] Step S400: constructing a Hankel matrix according to the weighted operating data and the current rolling window length value.

[0027] The current rolling window length value refers to the length value corresponding to the window when collecting and detecting real-time operation data information. The weighted operation data is detected and the Hankel matrix is ​​constructed through the current rolling window length value, so as to facilitate subsequent use.

[0028] Step S500: Bring the historical operation data and the input data at the current moment into the preset subspace prediction model to obtain the output optimal prediction value.

[0029] Among them, the subspace prediction model is configured as It is obtained based on the linear relationship model through the preset problem decomposition algorithm, where L = [L1 L2] is the corresponding subspace matrix, L1 is the subspace matrix of the past input and output, and W p is the past input and output, L2 is the subspace matrix of future input, U f For future input, For the optimal prediction value of future output, define the past input and output as

[0030] The linear relationship model is configured as xk +1 =Axk+Buk+kf*ek, yk=Cxk+Duk+ek, where, x k+1 is the state variable of the system at the next moment, A is the system matrix of the state variable of the system at the current moment, x k is the state variable of the system at the current moment, B is the system matrix of the system input at the current moment, u k is the input of the system at the current moment, k f is the Kalman filter gain for system stability, e k is the unknown new information sequence, y k is the output of the system at the current moment, C is the system matrix of the system state variables at the current moment, D is the system matrix of the system input at the current moment, and e k The covariance matrix of At this time, it is assumed that the input and output data sets with a measurement length of k∈{1,2,…2i+j-1} are given. At this time, u k ,y k It is measurable in k∈{1,2,…2i+j-1}.

[0031] For example, the Hankel matrix information of the data can be constructed, U p For past input, U f is the future input, and the past output is Y p With the future output Y f Also make similar definitions.

[0032] The problem decomposition algorithm is to construct a least squares problem based on the subspace prediction model, and use linear algebra tools RQ decomposition and SVD decomposition to solve the least squares problem to obtain the subspace prediction value.

[0033] The least squares problem is defined as The optimal prediction value of the output of future output Can be described as column space Y f To W p and U f The orthogonal mapping of Zhang Cheng's column space is Then, the orthogonal mapping is analyzed by RQ decomposition to obtain Then there is in Denotes pseudo-inverse, and then SVD decomposition is used to solve it.

[0034] Step S600: The output optimal prediction value and the output value at the current moment are subtracted by a preset error calculation algorithm to obtain a correction deviation; Among them, it is assumed that the prediction error of the system at the moment is ε(k) is the prediction error value at time k, y k is the process output at time k, is the predicted output at time k-1 for time k.

[0035] The correction deviation is calculated according to the preset error calculation formula. The error change formula is: η is the error change value at time k, and ε(k-1) is the prediction error value at time k-1.

[0036] Step S700: If the window correction deviation is greater than the preset upper limit deviation benchmark, the corresponding rolling window length value is reduced by one step; if the window correction deviation is less than the preset lower limit deviation benchmark, the corresponding rolling window length value is increased by one step.

[0037] The upper limit deviation benchmark refers to the threshold value that can be tolerated when the system changes too fast, and the upper limit deviation benchmark is obtained by querying from the database storing the upper limit deviation benchmark. The lower limit deviation benchmark refers to the threshold value that can be tolerated when the system changes too slow, and the lower limit deviation benchmark is obtained by querying from the database storing the lower limit deviation benchmark.

[0038] By comparing the window correction deviation with the upper limit deviation benchmark and the lower limit deviation benchmark respectively, when the window correction deviation is greater than the preset upper limit deviation benchmark, the corresponding rolling window length value is reduced by one step until the corresponding rolling window length value is consistent with the preset rolling window length minimum adjustment value. When the window correction deviation is less than the preset lower limit deviation benchmark, the corresponding rolling window length value is increased by one step until the corresponding rolling window length value is consistent with the preset rolling window length maximum adjustment value.

[0039] When the window correction deviation is less than the preset upper limit deviation benchmark and greater than the preset lower limit deviation benchmark, determine whether the prediction error is within the preset prediction error benchmark interval. When the prediction error is within the preset prediction error benchmark interval, the window length at this time is optimal. When the prediction error is not within the preset prediction error benchmark interval, the corresponding rolling window length value is increased by one step until the corresponding rolling window length value is consistent with the preset maximum adjustment value of the rolling window length.

[0040] At each sampling moment, the Hankel matrix is ​​updated in time according to the rolling window, the updated data replaces the old data, and the data is processed by minimizing the quadratic objective function to obtain the optimal control law, so that the obtained model accurately reflects the current system characteristics.

[0041] Step S800: Wait for new operation data and jump to execute step S300.

[0042] Among them, by waiting for new operating data and re-inputting the control weight matrix according to the type of operating data to obtain the weighted operating data, the model obtained at each moment can accurately reflect the current system characteristics.

[0043] For example, As an example, the subspace prediction controller design method with variable window length was not adopted, and the simulation results were as follows Figure 2 As shown in the figure, the subspace prediction controller design method with variable window length is adopted, and the simulation results are as follows Figure 3 shown.

[0044] It can be seen from the simulation results that, compared with the system input and output without variable window length, the tracking effect of this embodiment is better, the obtained model can better reflect the current system characteristics, and the system can also effectively suppress noise in the presence of noise, which proves the effectiveness of the algorithm.

[0045] The above are all preferred embodiments of the present application, and are not intended to limit the protection scope of the present application. Any feature disclosed in this specification (including the abstract and drawings), unless otherwise stated, can be replaced by other equivalent or alternative features with similar purposes. That is, unless otherwise stated, each feature is only an example of a series of equivalent or similar features.

Claims

1. A method for designing a subspace predictive controller with a variable window length, characterized in that: include: The weight matrix construction step is to construct the control weight matrix; Data acquisition step, obtaining the current operating data; Data weight adjustment step, inputting the control weight matrix according to the type of operating data to obtain the weighted operating data; Hankel matrix construction step: constructing the Hankel matrix according to the weighted running data and the current rolling window length value; Output prediction step, bringing historical operation data and current input data into the preset subspace prediction model to obtain the output optimal prediction value; Deviation calculation step, the output optimal prediction value and the output value at the current moment are calculated by a preset error calculation algorithm to obtain a corrected deviation; The window correction step includes: if the window correction deviation is greater than the preset upper limit deviation benchmark, the corresponding rolling window length value is reduced by one step; if the window correction deviation is less than the preset lower limit deviation benchmark, the corresponding rolling window length value is increased by one step; Loop through the configuration steps, wait for new running data, and return to the data adjustment step.

2. The method for designing a subspace prediction controller with a variable window length according to claim 1, characterized in that: The weight matrix construction step, the control weights include the main variables and the secondary variables; The data weighting step is to input the main variables in the operating data into the main variable control weight matrix, and input the secondary variables in the operating data into the secondary variable control weight matrix.

3. The method for designing a subspace prediction controller with a variable window length according to claim 1, characterized in that: The subspace prediction model is configured as Among them, L=[L1L2] is the corresponding subspace matrix, which is obtained based on the linear relationship model through a preset problem decomposition algorithm.

4. The method for designing a subspace prediction controller with a variable window length according to claim 3, characterized in that: The linear relationship model is configured as x k+1 =Ax k +Bu k +k f *e k ,y k =Cx k +Du k +e k .

5. The method for designing a subspace prediction controller with a variable window length according to claim 3, characterized in that: The problem decomposition algorithm is: Constructing a least squares problem for solving the optimal prediction value; The linear algebra tools RQ decomposition and SVD decomposition are used to solve the above least squares problem.

6. The method for designing a subspace prediction controller with a variable window length according to claim 1, characterized in that: The error calculation algorithm is: Assume that the prediction error of the system at the moment is The correction deviation is calculated according to the preset error calculation formula:

7. Replace the old data with the new data and obtain the optimal control law by minimizing the quadratic objective function.

Images