Control Method of Power Converter Based on Subspace Predictive Control

Abstract

The present application discloses a control method for a power converter based on subspace predictive control, which is used in the technical field of power converter control. The method includes: constructing a subspace predictor model of the power converter based on the historical data matrix of the power converter; predicting the linear part of the model and predicting the nonlinear part through a neural network; obtaining a functional relationship of the future predicted output current based on the predicted value of the linear part and the predicted value of the nonlinear part in the current control period; determining the optimal input voltage in the current control period based on the objective function; constructing a cost function with the optimal input voltage in the current control period and the voltage vectors corresponding to the respective switching states of the power converter, and taking the switching state corresponding to the voltage vector when the value of the cost function is the smallest as the optimal switching state in the current control period. The present application can accurately predict the state of the power converter and reduce the negative impact of parameter mismatch on the power converter.

Description

Technical Field

[0001] The present application relates to the technical field of power converter control, and particularly to a control method for a power converter based on subspace predictive control. Background Art

[0002] With the progress of renewable energy technology, the electrification of modern society is gradually intensifying. As the interface for energy conversion and consumption in renewable energy systems, power converters play a key role in this process. Finite control set model predictive control (FCS-MPC) has received extensive attention in the academic community due to its advantages such as strong multitasking ability, simple implementation, and rapid dynamic response. In standard finite control set model predictive control, the finite switching states of the inverter and the discrete model of the controlled object are first used to predict the possible future trajectories of the system. These predicted trajectories are substituted into the cost function containing the reference trajectory, and all switching states are evaluated by an enumeration method. The optimal switching state that can minimize the cost function is selected and applied to the power converter. It can be observed that the accuracy of predicting future behavior depends on the accuracy of the information about the controlled object. However, in an industrial environment, with the continuous operation of production equipment, frequent switching of operating conditions, and continuous changes in external conditions (such as temperature, humidity, electromagnetic interference, etc.), the internal parameters of the system often drift, and at the same time, it is also affected by various unpredictable external disturbances. Therefore, modeling errors, unmodeled dynamics, and parameter changes are inevitable. In these cases, the load parameters will fluctuate around their nominal values. Due to the error between the prediction model in the controller and the actual system model, the prediction accuracy will be reduced, thus deteriorating the control performance. The high dependence of FCS-MPC on the prediction model severely limits its application in the industrial field.

[0003] There are four main schemes to enhance the parameter robustness of model predictive control, namely the current look-up table scheme, the super-local model scheme, the parameter estimation scheme, and the time series model scheme. In the current look-up table scheme, a current look-up table is established, and the voltage vector and its corresponding current increment are stored in the look-up table. During control, the voltage vector is screened according to the current increment in the look-up table, which can achieve control without relying on system parameters. However, when updating the look-up table, for some voltage vectors that have not been used for a long time, the corresponding current increments have not been updated, resulting in a stagnation effect. When selecting to update the stagnant voltage vector at this time, the control effect will decrease due to the incorrect corresponding current increment. In the super-local model scheme, the model of the control system is converted into a super-local model, where the future state of the system is represented as the sum of the lumped disturbance term and the control input term. An extended state observer is designed to estimate the lumped disturbance in real time to complete the observation of the system state to resist parameter disturbances. However, since it separates the system state into the lumped disturbance term and the control input term, although various observers can be designed to estimate the lumped disturbance term, the input gain of the control input term still depends on the prior knowledge of the system, which also makes this control scheme unable to complete control when the system parameters are unknown. In the parameter estimation scheme, the recursive least squares algorithm is used to estimate the parameters such as the inductance and resistance of the control load online to resist parameter disturbances. However, it is sensitive to the noise in the input data. If there is a lot of measurement noise in the input signal, it may cause fluctuations in the estimation results, thus affecting the control performance. In addition, the estimation algorithm cannot fully capture the nonlinear characteristics of the control system. In the time series model scheme, the output of the control system is represented as a subspace time series, and the recursive least squares is used to estimate the sequence coefficients. The control input can be obtained according to the future reference output. However, this scheme assumes that the system behavior is linear or can be linearly approximated. Therefore, it is difficult to capture the complex nonlinear dynamics in the actual system. The characterization ability of the time series subspace method for nonlinear dynamics is limited by the model structure, especially when the parameter estimation error is large or the external disturbance is significant, which may lead to fluctuations or degradation of the control performance. Summary of the Invention

[0004] To solve the deficiencies of the prior art, the purpose of this application is to provide a control method for a power converter based on subspace predictive control, so as to accurately predict the state of the power converter and reduce the negative impact of parameter mismatch on the power converter.

[0005] To achieve the above purpose, this application provides a control method for a power converter based on subspace predictive control, and the method includes:

[0006] Construct a subspace predictor model of the power converter based on the historical data matrix within the historical time window of the power converter, the output current within the future time window, and the input voltage within the future time window. The historical data matrix includes the historical input voltage and the corresponding historical output current, and the subspace predictor model includes a linear part and a non-linear part;

[0007] Construct an objective function for the linear part based on the historical data matrix, and determine the predicted value of the linear part for the current control period based on the objective function;

[0008] Construct a first non-linear function for the non-linear part, and a second non-linear function for the output current of the power converter based on the predicted value of the linear part for the current control period. Use a neural network to estimate the first non-linear function and the second non-linear function, and determine the predicted value of the non-linear part for the current control period;

[0009] Based on the predicted value of the linear part for the current control period, the predicted value of the non-linear part for the current control period, and the subspace predictor model, determine the functional relationship of the predicted output current within the future time window;

[0010] Construct an objective function with the functional relationship and the expected output current within the preset future time window, and determine the optimal input voltage for the current control period based on the objective function;

[0011] Construct a cost function with the optimal input voltage for the current control period and the voltage vectors corresponding to the respective switching states of the power converter. When the value of the cost function is minimized, determine the voltage vector corresponding to the optimal input voltage for the current control period, and use the switching state corresponding to this voltage vector as the optimal switching state of the power converter for the current control period.

[0012] Furthermore, constructing the subspace predictor model of the power converter includes:

[0013] Define the non-linear part of the subspace predictor model;

[0014] Construct a subspace predictor model of the power converter based on the historical data matrix within the historical time window of the power converter, the output current within the future time window, the input voltage within the future time window, and the non-linear part. The linear part of the subspace predictor model is determined based on the coefficient matrix of the subspace predictor model.

[0015] Furthermore, constructing the subspace predictor model of the power converter includes:

[0016] The subspace predictor model of the power converter is expressed as:

[0017] ;

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] where k represents the current control period, represents the length of the historical time window, represents the length of the future time window, represents the predicted output current matrix within the future time window, represents the output current increment matrix of adjacent control periods within the future time window, and represents the coefficient matrix, represents the input voltage matrix within the future time window, represents the input voltage increment matrix of adjacent control periods within the future time window, represents the historical data matrix, and represents the historical input voltage and the corresponding historical output current, represents the historical input voltage increment matrix and the corresponding historical output current increment matrix of two adjacent control periods within the historical time window, represents the nonlinear part of the subspace predictor model, represents a single nonlinear term, represents a matrix of all 1s with

[0023] rows and columns;

[0024] Furthermore, an objective function of the linear part is constructed based on the historical data matrix, and the predicted value of the linear part of the current control period is determined based on the objective function, including:

[0025] Construct an objective function of the coefficient matrix of the linear part with the increment of the historical data matrix and the predicted value of the nonlinear part of the previous control period;

[0026] Take the partial derivative of the objective function to obtain the coefficient matrix of the linear part of the current control period, and determine the predicted value of the linear part of the current control period based on the historical data matrix and the coefficient matrix of the linear part of the current control period.

[0027] Furthermore, an objective function of the linear part is constructed based on the historical data matrix, and the predicted value of the linear part of the current control period is determined based on the objective function, including:

[0028] The objective function of the coefficient matrix of the linear part is expressed as:

[0029] ;

[0030] ;

[0031] wherein, represents the weight factor, represents the increment of the historical data matrix at time represents the increment of the output current at time represents the increment of the input voltage at time represents the j-th row data of the coefficient matrix ; represents the predicted value of the coefficient matrix for the current control period, represents the predicted value of a single non-linear term for the previous control period;

[0032] Let , then the predicted value of the coefficient matrix of the linear part for the current control period is expressed as:

[0033] ;

[0034] ;

[0035] wherein, represents the step factor, represents the predicted value of the coefficient matrix for the previous control period;

[0036] Based on the predicted value of the coefficient matrix of the linear part for the current control period and the historical data matrix, determine the predicted value of the linear part for the current control period.

[0037] Furthermore, use a neural network to estimate the first non-linear function and the second non-linear function to determine the predicted value of the non-linear part for the current control period, including:

[0038] Define the weight matrix of the neural network and define the activation function of the neural network with current as a parameter;

[0039] Construct the first non-linear function of a single non-linear term according to the weight matrix and the activation function;

[0040] Construct the second non-linear function of the output current of the power converter according to the predicted value of the linear part for the current control period, the weight matrix, and the activation function;

[0041] Estimate the first non - linear function and the second non - linear function using a neural network, including the estimated value of the output current in the current control period, the estimated value of the weight matrix in the current control period, and the estimated values of each individual non - linear term in the current control period;

[0042] Based on the estimated values of each individual non - linear term in the current control period, determine the predicted value of the non - linear part in the current control period.

[0043] Furthermore, estimating the first non - linear function and the second non - linear function using a neural network to determine the predicted value of the non - linear part in the current control period includes:

[0044] The first non - linear function of the individual non - linear term d is expressed as:

[0045] ;

[0046] ;

[0047] The second non - linear function of the output current of the power converter is expressed as:

[0048] ;

[0049] ;

[0050] Wherein, represents the output current of the power converter, represents the predicted value of the linear part in the current control period, represents the weight matrix of the neural network, represents the activation function of the neural network, represents the minimum error of the non - linear term estimation, represents the minimum error of the output current estimation;

[0051] Estimate the first non - linear function and the second non - linear function using a neural network, and the estimated value of the non - linear term d and the estimated value of the output current of the power converter are respectively expressed as:

[0052] ;

[0053] ;

[0054] ;

[0055] Wherein, , , represent positive coefficients, Represents the estimated value of the linear part prediction value, Represents the estimated value of the weight matrix of the neural network, Represents the error between the estimated value of the output current of the power converter and the actual output current of the power converter, Represents the derivative of the estimated value of the weight matrix of the neural network;

[0056] The estimated value of the non-linear term d and the estimated value of the output current of the power converter as well as the derivative of the estimated value of the weight matrix are respectively discretized by first-order Euler to obtain the following expressions:

[0057] ;

[0058] wherein, represents the estimated value of the output current in the current control cycle, represents the estimated value of the weight matrix in the current control cycle, represents the estimated value of the linear part prediction value in the current control cycle, represents the length of the control cycle, represents the estimated value of a single non-linear term in the current control cycle;

[0059] The non-linear part prediction value of the subspace predictor model in the current control cycle is expressed as:

[0060] .

[0061] Furthermore, based on the linear part prediction value in the current control cycle, the non-linear part prediction value in the current control cycle, and the subspace predictor model, a functional relationship for predicting the output current within a future time window is determined, including:

[0062] Substitute the linear part prediction value in the current control cycle and the non-linear part prediction value in the current control cycle into the expression of the subspace predictor model to obtain the functional relationship of the predicted output current matrix in the future time window as:

[0063] ;

[0064] ;

[0065] ;

[0066] wherein, represents an m×m dimensional matrix, the positive diagonal elements of which are 1 and the remaining elements are 0, and m represents the data dimension of the system input and output;

[0067] Construct an objective function based on the functional relationship and the expected output current within a preset future time window, and determine the optimal input voltage for the current control period, including:

[0068] Using the predicted output current matrix Construct an objective function with the functional relationship and the expected output current matrix within a preset future time window It is expressed as:

[0069] ;

[0070] ;

[0071] Among them, Represents the expected output current matrix within a preset future time window, Represents the expected output current at time k + 1 obtained by extrapolating based on the vector angle, Represents the input voltage matrix within the future time window, And Represents the diagonal weight matrix;

[0072] When , according to the objective function Obtain the input voltage increment matrix within the future time window as:

[0073] ;

[0074] ;

[0075] ;

[0076] Using the sum of the first matrix element of and the optimal input voltage of the previous control period as the optimal input voltage for the current control period.

[0077] Furthermore, construct a cost function based on the optimal input voltage of the current control period and the voltage vectors corresponding to the respective switching states of the power converter, including:

[0078] Construct a cost function based on the optimal input voltage of the current control period and the voltage vectors corresponding to the respective switching states of the power converter The expression of which is:

[0079] ;

[0080] Among them, Denote the voltage vectors corresponding to the respective switching states of the power converter, and \(n\) represents \(n\) switching states;

[0081] Based on the cost function Calculate the square of the absolute value between each voltage vector and the optimal input voltage of the current control period Determine the voltage vector corresponding to the minimum value of the square of the absolute value, and use the switching state corresponding to this voltage vector as the optimal switching state of the power converter in the current control period.

[0082] This application also provides a computer device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the steps of the control method of the power converter based on subspace predictive control as described above.

[0083] This application constructs a subspace predictor model of the power converter with a linear part and a nonlinear part, predicts its linear part and nonlinear part, and solves for the predicted output current within a future time window based on the linear part prediction value, the nonlinear part prediction value, and the subspace predictor model. The nonlinear dynamic information of the device captured by the neural network estimator can be integrated into the subspace predictor updated in real time, so as to accurately predict the state of the power converter and reduce the negative impact of parameter mismatch on the control device. By determining the optimal switching state of the power converter in the current control period, the optimal input voltage of the future control input of the power converter can be obtained, improving the accuracy of device control, and enhancing the robustness of the device to parameter changes without relying on the physical model and prior knowledge of the controlled object. BRIEF DESCRIPTION OF THE DRAWINGS

[0084] Figure 1 It is the first flowchart of the control method of the power converter based on subspace predictive control provided by the embodiment of this application;

[0085] Figure 2 It is the second flowchart of the control method of the power converter based on subspace predictive control provided by the embodiment of this application;

[0086] Figure 3 It is the third flowchart of the control method of the power converter based on subspace predictive control provided by the embodiment of this application;

[0087] Figure 4 It is the fourth flowchart of the control method of the power converter based on subspace predictive control provided by the embodiment of this application;

[0088] Figure 5 It is the system block diagram of the power converter based on subspace predictive control provided by the embodiment of this application;

[0089] Figure 6 Schematic diagram of the steady-state performance of the output current and the DC-side capacitor voltage of the traditional finite set model predictive control method provided by the embodiment of the present application under parameter mismatch;

[0090] Figure 7 Schematic diagram of the transient performance of the output current and the DC-side capacitor voltage of the traditional finite set model predictive control method provided by the embodiment of the present application under parameter mismatch;

[0091] Figure 8 Schematic diagram of the steady-state performance of the output current and the DC-side capacitor voltage of the method of the present application provided by the embodiment of the present application under parameter mismatch;

[0092] Figure 9 Schematic diagram of the transient performance of the output current and the DC-side capacitor voltage of the method of the present application provided by the embodiment of the present application under parameter mismatch;

[0093] Figure 10 System block diagram of the computer device provided by the embodiment of the present application. Detailed implementation manners

[0094] In order to enable those skilled in the art to better understand the solution of the present application, the technical solutions in the specific implementation manners of the present application will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present application.

[0095] Please refer to Figure 1 , a control method for a power converter based on subspace predictive control provided by the present application, the method includes the steps of:

[0096] S101. Construct a subspace predictor model of the power converter based on the historical data matrix within the historical time window of the power converter, the output current within the future time window, and the input voltage within the future time window, wherein the historical data matrix includes the historical input voltage and the corresponding historical output current, and the subspace predictor model includes a linear part and a nonlinear part;

[0097] S102. Construct an objective function for the linear part based on the historical data matrix, and determine the predicted value of the linear part for the current control period based on the objective function;

[0098] S103. Construct a first nonlinear function for the nonlinear part, and a second nonlinear function for the output current of the power converter according to the predicted value of the linear part for the current control period, and use a neural network to estimate the first nonlinear function and the second nonlinear function to determine the predicted value of the nonlinear part for the current control period;

[0099] S104. Determine the functional relationship of the predicted output current within the future time window based on the predicted value of the linear part in the current control period, the predicted value of the non-linear part in the current control period, and the subspace predictor model;

[0100] S105. Construct an objective function with the functional relationship and the expected output current within the preset future time window, and determine the optimal input voltage for the current control period based on the objective function;

[0101] S106. Construct a cost function with the optimal input voltage for the current control period and the voltage vectors corresponding to the respective switching states of the power converter. When the value of the cost function is minimized, determine the voltage vector corresponding to the optimal input voltage for the current control period, and use the switching state corresponding to this voltage vector as the optimal switching state of the power converter in the current control period.

[0102] Exemplarily, the length of the historical time window can be the length of multiple control periods. By collecting the input voltage and output current of the power converter in each control period as data samples, that is, based on each data sample obtained within the historical time window as the historical data matrix, the historical data matrix includes the historical input voltage and the corresponding historical output current. The length of the future time window can be the length of multiple control periods. The historical time window and the future time window can be continuous in time. The power converter can be an inverter.

[0103] Exemplarily, the subspace predictor model can be a model that describes the future output using past input-output data and future control input data. Represent the controlled power converter as a subspace predictor model. Therefore, construct the subspace predictor model of the power converter based on the historical data matrix within the historical time window of the power converter, the output current within the future time window, and the input voltage within the future time window. Since in the industrial production environment, with the continuous operation of the equipment, the frequent switching of working conditions, and the continuous change of external conditions (such as temperature, humidity, electromagnetic interference, etc.), the internal parameters of the equipment often drift, and at the same time, the equipment is also affected by various unpredictable external disturbances, resulting in some errors in the non-linear dynamic information between the model and the actual equipment. Therefore, when constructing the subspace predictor model of the power converter in this application, the non-linear part of the model is considered to be able to better predict the state of the equipment through the model.

[0104] In this embodiment, by constructing a subspace predictor model of a power converter with a linear part and a non-linear part, predicting its linear part, and predicting its non-linear part through a neural network estimator, the predicted output current within a future time window is solved based on the predicted value of the linear part, the predicted value of the non-linear part, and the subspace predictor model. The non-linear dynamic information of the device captured by the neural network estimator can be integrated into the subspace predictor that is updated in real time, so as to accurately predict the state of the power converter and reduce the negative impact of parameter mismatch on the control device. By constructing an objective function for the input voltage within a future time window, determining the optimal input voltage for the current control period based on the objective function, and constructing a cost function with the optimal input voltage for the current control period and the voltage vectors corresponding to the respective switching states of the power converter to determine the optimal switching state of the power converter for the current control period, the optimal input voltage for the future control input of the power converter can be obtained, improving the accuracy of device control. Without relying on the physical model and prior knowledge of the controlled object, the robustness of the device to parameter changes is enhanced.

[0105] In one embodiment, as Figure 2 shown, constructing a subspace predictor model of a power converter includes the steps of:

[0106] S201. Define the non-linear part of the subspace predictor model;

[0107] S202. Based on the historical data matrix within the historical time window of the power converter, the output current within the future time window, the input voltage within the future time window, and the non-linear part, construct a subspace predictor model of the power converter, wherein the linear part of the subspace predictor model is determined based on the coefficient matrix of the subspace predictor model.

[0108] A typical subspace predictor model can be expressed as:

[0109] ;

[0110] In one embodiment, the above subspace predictor model is transformed by multiplying both sides of the formula by an increment, which represents the incremental form of the independent variable in the formula. The subspace predictor model of the power converter can be expressed as:

[0111] ;

[0112] ;

[0113] ;

[0114] ;

[0115] ;

[0116] Among them, \(k\) represents the current control period, represents the length of the historical time window, represents the length of the future time window, represents the predicted output current matrix within the future time window, represents the output current increment matrix of adjacent control periods within the future time window, and represents the coefficient matrix, represents the input voltage matrix within the future time window, represents the input voltage increment matrix of adjacent control periods within the future time window, represents the historical data matrix, and represents the historical input voltage and the corresponding historical output current, represents the historical input voltage increment matrix and the corresponding historical output current increment matrix of two adjacent control periods within the historical time window, represents the nonlinear part of the subspace predictor model, represents a single nonlinear term, represents a column vector of all 1s with \(m\) rows;

[0117] Taking represents the linear part of the subspace predictor model.

[0118] In this embodiment, the subspace predictor model is constructed through historical input-output data, taking into account both the linear part and the nonlinear part in the model, and introducing an unknown residual nonlinear part into the model, enhancing the model's resistance to unknown disturbances of the device and its compensation ability for device uncertainties.

[0119] In one embodiment, as Figure 3 shown, the objective function of the linear part is constructed based on the historical data matrix, and the predicted value of the linear part for the current control period is determined based on the objective function, including the steps of:

[0120] S301. Construct the objective function of the coefficient matrix of the linear part with the increment of the historical data matrix and the predicted value of the nonlinear part of the previous control period;

[0121] S302. Take the partial derivative of the objective function to obtain the coefficient matrix of the linear part for the current control period, and determine the predicted value of the linear part for the current control period based on the historical data matrix and the coefficient matrix of the linear part for the current control period.

[0122] It can be seen from the above expression of the linear part that represents the historical input voltage increment matrix and the corresponding historical output current increment matrix, represents the input voltage increment matrix within the future time window. Therefore, the estimation of the linear part can be understood as the estimation of the coefficient matrix.

[0123] In one embodiment, the objective function of the coefficient matrix of is expressed as:

[0124] ;

[0125] ;

[0126] where, represents the weight factor, represents the increment of the historical data matrix at time represents the output current increment at time represents the input voltage increment at time represents the coefficient matrix the j-th row data of represents the predicted value of the coefficient matrix for the current control period, represents the predicted value of a single non-linear term for the previous control period;

[0127] Take the partial derivative of the objective function, that is, let , then the predicted value of the coefficient matrix of the linear part for the current control period is expressed as:

[0128] ;

[0129] ;

[0130] where, represents the step factor, represents the predicted value of the coefficient matrix for the previous control period;

[0131] Based on the predicted value of the coefficient matrix of the linear part for the current control period and the historical data matrix, determine the predicted value of the linear part for the current control period.

[0132] In this embodiment, through the historical input voltage and output current data, the subspace predictor model captures the linear dynamic information of the device, enhancing the real-time prediction ability of the device behavior.

[0133] In one embodiment, as Figure 4As shown, the steps of using a neural network to estimate the first non - linear function and the second non - linear function and determining the predicted value of the non - linear part in the current control period include:

[0134] S401. Define the weight matrix of the neural network and define the activation function of the neural network with current as a parameter;

[0135] S402. Construct the first non - linear function of a single non - linear term according to the weight matrix and the activation function;

[0136] S403. Construct the second non - linear function of the output current of the power converter according to the predicted value of the linear part in the current control period, the weight matrix, and the activation function;

[0137] S404. Use the neural network to estimate the first non - linear function and the second non - linear function, the estimated value of the output current in the current control period, the estimated value of the weight matrix in the current control period, and the estimated values of each single non - linear term in the current control period;

[0138] S405. Based on the estimated values of each single non - linear term in the current control period, determine the predicted value of the non - linear part in the current control period.

[0139] In this embodiment, the non - linear part is estimated by using a neural network.

[0140] Specifically, the first non - linear function of a single non - linear term d is expressed as:

[0141] ;

[0142] ;

[0143] Wherein, represents the activation function of the neural network, satisfies , represents the optimal weight matrix, and is a positive constant, represents the minimum error of non - linear term estimation, and is a positive constant.

[0144] The second non - linear function of the output current of the power converter is expressed as:

[0145] ;

[0146] ;

[0147] Wherein, represents the output current of the power converter, represents the predicted value of the linear part in the current control period, represents the weight matrix of the neural network, represents the activation function of the neural network, represents the minimum error of the non - linear term estimation, represents the minimum error of the output current estimation;

[0148] The above non - linear function is estimated using a neural network predictor. The first non - linear function and the second non - linear function are estimated using a neural network. The estimated value of the non - linear term d and the estimated value of the output current of the power converter are respectively expressed as:

[0149] ;

[0150] ;

[0151] ;

[0152] where, , , represent positive coefficients, represents the estimated value of the linear part prediction, represents the estimated value of the weight matrix of the neural network, represents the error between the estimated value of the output current of the power converter and the actual output current of the power converter, represents the derivative of the estimated value of the weight matrix of the neural network;

[0153] For the estimated value of the non - linear term d , the estimated value of the output current of the power converter and the derivative of the estimated value of the weight matrix are respectively discretized by first - order Euler to obtain the following expressions:

[0154] ;

[0155] where, represents the estimated value of the output current in the current control period, represents the estimated value of the weight matrix in the current control period, represents the estimated value of the linear part prediction in the current control period, represents the length of the control period, represents the estimated value of a single non - linear term in the current control period;

[0156] The non - linear part prediction value of the subspace predictor model in the current control period is expressed as:

[0157] 。

[0158] In this embodiment, according to the predicted value of the linear part that has been obtained, the neural network is used to estimate the non-linear part. By combining the neural network predictor and the subspace predictor, based on the weight matrix update and error compensation mechanism of the neural network, the online approximation of the dynamic non-linear disturbance is realized, so that the non-linear dynamic information and linear dynamic information of the device can be captured, and the processing ability of the model for unknown disturbances and non-linear transformations is improved. Especially when there are parameter mismatches, the device still maintains high control performance.

[0159] In one embodiment, the steps of determining the functional relationship of the predicted output current within the future time window based on the predicted value of the linear part in the current control cycle, the predicted value of the non-linear part in the current control cycle, and the subspace predictor model include:

[0160] Substitute the predicted value of the linear part in the current control cycle and the predicted value of the non-linear part in the current control cycle into the expression of the subspace predictor model to obtain the predicted output current matrix The functional relationship is:

[0161] ;

[0162] ;

[0163] ;

[0164] where represents an m×m dimensional matrix, the positive diagonal elements of this matrix are 1, and the remaining elements are 0, and m represents the data dimension of the system input and output.

[0165] In this embodiment, based on the above-obtained predicted value of the linear part and the predicted value of the non-linear part, the expression of the complete subspace predictor model can be obtained.

[0166] In order to obtain the optimal input voltage of the device, an objective function of the input voltage is constructed, that is, an objective function is constructed with the functional relationship of the predicted output current and the expected output current within the preset future time window, and the optimal input voltage of the current control cycle is determined based on this objective function.

[0167] In one embodiment, an objective function is constructed with the functional relationship of the predicted output current matrix and the expected output current matrix within the preset future time window which is expressed as:

[0168] ;

[0169] ;

[0170] Among them, represents the expected output current matrix within a preset future time window, represents the expected output current at the (k + 1)-th moment obtained by extrapolation based on the vector angle, represents the input voltage matrix within the future time window, and represents the diagonal weight matrix.

[0171] It can be seen from the above objective function that if the input voltage is to be optimized, the optimization of the objective function needs to satisfy the following two points as much as possible: formula represents minimizing the deviation between the future output current and the expected output current as much as possible to ensure that the output current can be consistent with the expected output current. Formula represents minimizing the change rate of the input voltage as much as possible, making the adjustment range of the input voltage as small as possible, ensuring that the switching times of the switching state of the power converter are reduced, thereby reducing the switching frequency.

[0172] Predicted output current is substituted into the objective function , and it can be seen that the future behavior of the device is determined by the future input voltage increment matrix . Therefore, taking the partial derivative of the objective function with respect to the input voltage increment, the input voltage increment matrix can be obtained.

[0173] That is, when , according to the objective function , the input voltage increment matrix within the future time window is:

[0174] ;

[0175] ;

[0176] ;

[0177] Because only one input voltage can act in each control period of the power converter, the sum of the first matrix element of and the optimal input voltage of the previous control period is used as the optimal input voltage for the current control period.

[0178] In this embodiment, the minimum error between the predicted future output current of the subspace predictor model and the desired output current is used as the optimization objective to derive the optimal value of the future input voltage of the power converter. Using this optimal value as the input voltage for the current control cycle can improve the accuracy of device control and achieve good controllability of the device in the case of device parameter mismatch.

[0179] A cost function is constructed using the optimal input voltage for the current control cycle and the voltage vectors corresponding to the respective switching states of the power converter. When the value of the cost function is minimized, the voltage vector corresponding to the optimal input voltage for the current control cycle is determined, and the switching state corresponding to this voltage vector is used as the optimal switching state of the power converter for the current control cycle, thereby achieving control of the power converter.

[0180] In one embodiment, since different switching states of the power converter output different voltage vectors, a finite number of switching states correspond to a finite number of discrete voltage vectors. A cost function is constructed using the optimal input voltage for the current control cycle and the voltage vectors corresponding to the respective switching states of the power converter The expression is:

[0181] ;

[0182] where represents the voltage vectors corresponding to the respective switching states of the power converter, and n represents n switching states.

[0183] Based on the cost function calculate the square of the absolute value between each voltage vector and the optimal input voltage for the current control cycle. When the square of the absolute value is minimized, that is, when the value of the cost function is minimized, it indicates that the optimal input voltage for the current control cycle is closest to the corresponding voltage vector. Furthermore, determine the voltage vector corresponding to the optimal input voltage for the current control cycle, determine the corresponding switching state based on this voltage vector, and use this switching state as the optimal switching state of the power converter for the current control cycle, thereby achieving precise control of the power converter.

[0184] In this embodiment, the construction of the cost function is based on the voltage vectors corresponding to the respective switching states, and these voltage vectors are based on a finite discrete selection method. In another embodiment, the voltage vectors can use space vector modulation of the control variables in a continuous space range to obtain continuous voltage vectors, thereby further improving the smoothness and accuracy of the device output and reducing switching losses and harmonic distortion.

[0185] Such as Figure 5As shown in the figure, it is a system block diagram of a power converter based on subspace predictive control according to an embodiment of the present application. Taking an inverter as an example, the input voltage and output current data of the inverter are collected as a historical data matrix. Based on this historical data matrix, a subspace predictor model of the inverter is constructed. The linear part of the model is estimated to obtain a linear part prediction value, and a neural network is used to estimate the nonlinear part of the model to obtain a nonlinear part prediction value. Based on the linear part prediction value and the nonlinear part prediction value, a complete subspace predictor model can be obtained, and then the predicted output current within a future time window can be obtained. An objective function of the input voltage within the future time window is constructed according to the predicted output current within the future time window and the desired output current within the future time window. According to this objective function, the optimal input voltage of the current control period is determined. A cost function is constructed with the optimal input voltage of the current control period and the voltage vectors corresponding to the respective switching states of the inverter. When the value of the cost function is minimized, the voltage vector corresponding to the optimal input voltage of the current control period is determined, and the switching state corresponding to this voltage vector is used as the optimal switching state of the inverter in the current control period, thereby realizing the control of the inverter.

[0186] In a specific embodiment of the present application, a neutral-point clamped three-level inverter is used as the control device, and its variables are shown in the following table:

[0187]

[0188] Figure 6 It is a schematic diagram of the steady-state performance of the output current and the DC-side capacitor voltage of the traditional finite set model predictive control method under parameter mismatch. Figure 7 It is a schematic diagram of the transient performance of the output current and the DC-side capacitor voltage of the traditional finite set model predictive control method under parameter mismatch. Figure 8 It is a schematic diagram of the steady-state performance of the output current and the DC-side capacitor voltage of the method of the present application under parameter mismatch. Figure 9 It is a schematic diagram of the transient performance of the output current and the DC-side capacitor voltage of the method of the present application under parameter mismatch. The load inductance value under the parameter mismatch condition is 80% of the standard inductance. In the figure represents the amplitude of the reference current. The steady-state performance means that the amplitude of the reference current is maintained at 12 A all the time, and the transient performance means that the reference current changes suddenly from 8 A to 12 A. i a 、i b and i c represent three-phase currents, and v p and v l represent the upper-side capacitor voltage and the lower-side capacitor voltage of the DC side of the inverter. From Figure 6 and Figure 7It can be seen that under the condition of parameter mismatch, the total harmonic distortion (THD) of the output current in the traditional control scheme reaches 4.65%, showing large current fluctuations and poor steady-state performance. And from Figure 8 and Figure 9 it can be seen that under the same parameter mismatch condition, the control scheme of this application can not only effectively cope with the load inductance deviation, but also resist external disturbances, maintain the stability and high performance of the system, and the THD of the output current is significantly reduced, only 1.97%. This result shows that compared with the traditional control method, the output current quality of this application's control scheme has been improved by 57.63%. This further verifies that the scheme proposed in this application has extremely high robustness in dealing with load parameter mismatch, and at the same time can effectively improve the output current quality, enabling the device to still have excellent control performance and reliability under complex working conditions.

[0189] The embodiment of this application also provides a computer-readable storage medium, in which a computer program is stored, and when the computer program is executed by a processor, the steps of any one of the above control methods of the power converter based on subspace prediction control are implemented.

[0190] Figure 10 It is a schematic hardware structure diagram of a computer device provided by the embodiment of this application. Figure 10 The shown computer device includes: a processor 1001, a communication interface 1002, a memory 1003, and a communication bus 1004. The processor 1001, the communication interface 1002, and the memory 1003 complete mutual communication through the communication bus 1004. Among them, Figure 10 The connection manner between the shown processor 1001, communication interface 1002, and memory 1003 is only exemplary. During implementation, the processor 1001, communication interface 1002, and memory 1003 can also communicate and connect with each other using other connection manners besides the communication bus 1004.

[0191] The memory 1003 can be used to store computer programs, which may include instructions and data, and implement the steps of any of the above control methods for power converters based on subspace predictive control. In the embodiments of this application, the memory 1003 can be various types of storage media, such as random access memory (RAM), read only memory (ROM), non-volatile RAM (NVRAM), programmable ROM (PROM), erasable PROM (EPROM), electrically erasable PROM (EEPROM), flash memory, optical memory, and registers, etc. The memory 1003 can include a hard disk and / or memory.

[0192] The processor 1001 can be a general-purpose processor, which can be a processor that executes specific steps and / or operations by reading and executing a computer program (such as the computer program) stored in a memory (such as the memory 1003). The general-purpose processor may use the data stored in the memory (such as the data in the memory 1003) during the execution of the steps and / or operations. The general-purpose processor can be, for example but not limited to, a central processing unit (CPU). In addition, the processor 1001 can also be a dedicated processor, which can be a processor specifically designed to execute specific steps and / or operations. The dedicated processor can be, for example but not limited to, ASIC and FPGA, etc. In addition, the processor 1001 can also be a combination of multiple processors, such as a multi-core processor.

[0193] The communication interface 1002 can include interfaces such as input / output (I / O) interfaces, physical interfaces, and logical interfaces for implementing the interconnection of components inside the network device, as well as interfaces for implementing the interconnection between the network device and other devices (such as network devices). The communication network can be Ethernet, radio access network (RAN), wireless local area networks (WLAN), etc. The communication interface 1002 can be a module, circuit, transceiver, or any device capable of implementing communication.

[0194] In the implementation process, each step of the above method can be completed by the integrated logic circuit of the hardware in the processor 1001 or the instructions in the form of software. The method disclosed in combination with the embodiments of the present application can be directly embodied as being executed and completed by the hardware processor, or executed and completed by the combination of the hardware and software modules in the processor. The software module can be located in a mature storage medium in the art such as random access memory flash memory, read-only memory, programmable read-only memory, or electrically erasable programmable memory, registers, etc. This storage medium is located in the memory 1003, and the processor 1001 reads the information in the memory 1003 and combines its hardware to complete the steps of the above method. To avoid repetition, it will not be described in detail here.

[0195] The above description of the embodiments is to enable those of ordinary skill in the art to understand and apply the present application. Obviously, those skilled in the art can easily make various modifications to the above embodiments and apply the general principles described herein to other embodiments without creative efforts. Therefore, the present application is not limited to the above embodiments, and all improvements and modifications made by those skilled in the art based on the disclosure of the present application should be within the protection scope of the present application.

Claims

1. A control method for a power converter based on subspace predictive control, characterized in that: The method comprises: Constructing a subspace predictor model of the power converter based on a historical data matrix in a historical time window of the power converter, an output current in a future time window, and an input voltage in a future time window, wherein the historical data matrix includes a historical input voltage and a corresponding historical output current, and the subspace predictor model includes a linear part and a nonlinear part; Constructing the objective function of the linear part based on the historical data matrix, and determining the linear part prediction value of the current control cycle based on the objective function; Constructing a first nonlinear function of the nonlinear part, and constructing a second nonlinear function of the output current of the power converter according to the linear part prediction value of the current control cycle, estimating the first nonlinear function and the second nonlinear function by using a neural network, and determining the nonlinear part prediction value of the current control cycle; Determine a functional relationship for predicting output current in a future time window based on the linear part prediction value of the current control cycle, the nonlinear part prediction value of the current control cycle, and a subspace predictor model; Constructing an objective function with the functional relationship and the expected output current in a preset future time window, and determining an optimal input voltage of a current control cycle based on the objective function; A cost function is constructed using the optimal input voltage of the current control cycle and the voltage vectors corresponding to the various switching states of the power converter. When the value of the cost function is minimum, the voltage vector corresponding to the optimal input voltage of the current control cycle is determined, and the switching state corresponding to the voltage vector is used as the optimal switching state of the power converter in the current control cycle.

2. The control method of a power converter based on subspace predictive control according to claim 1, wherein the subspace predictor model of the power converter is constructed, comprising: defining a nonlinear portion of the subspace predictor model; A subspace predictor model of the power converter is constructed based on the historical data matrix within the historical time window of the power converter, the output current within the future time window, the input voltage within the future time window, and the nonlinear part, wherein the linear part of the subspace predictor model is determined based on the coefficient matrix of the subspace predictor model.

3. The control method of a power converter based on subspace predictive control according to claim 2, characterized in that: The constructing of the subspace predictor model of the power converter comprises: The subspace predictor model of the power converter is expressed as: ； ； ； ； ； Among them, k represents the current control cycle, represents the length of the historical time window, represents the length of the future time window, represents the predicted output current matrix in the future time window, represents the output current increment matrix of adjacent control cycles in the future time window, and represents the coefficient matrix, represents the input voltage matrix in the future time window, represents the input voltage increment matrix of adjacent control cycles in the future time window, represents the historical data matrix, and Represents the historical input voltage and the corresponding historical output current, Represents the historical input voltage increment matrix and the corresponding historical output current increment matrix of two adjacent control cycles in the historical time window, represents the nonlinear part of the subspace predictor model, represents a single nonlinear term, express A matrix with all 1s in one row and one column; by represents the linear part of the subspace predictor model.

4. The control method of a power converter based on subspace predictive control according to claim 2, characterized in that: The step of constructing the objective function of the linear part based on the historical data matrix and determining the linear part prediction value of the current control cycle based on the objective function includes: The objective function of the coefficient matrix of the linear part is constructed by using the increment of the historical data matrix and the predicted value of the nonlinear part of the previous control cycle; The objective function is partially derived to obtain a coefficient matrix of the linear part of the current control cycle, and a predicted value of the linear part of the current control cycle is determined based on the historical data matrix and the coefficient matrix of the linear part of the current control cycle.

5. The control method of a power converter based on subspace predictive control according to claim 3, characterized in that: The step of constructing the objective function of the linear part based on the historical data matrix and determining the linear part prediction value of the current control cycle based on the objective function includes: The objective function of the coefficient matrix of the linear part It is expressed as: ； ； in, represents the weight factor, express The increment of the historical data matrix at time , express The output current increment at the moment, express The input voltage increment at time Representation coefficient matrix The j-th row of data, represents the coefficient matrix prediction value of the current control cycle, Represents the predicted value of a single nonlinear term in the previous control period; make , then the coefficient matrix prediction value of the linear part of the current control cycle is It is expressed as: ； ； in, represents the step size factor, Represents the coefficient matrix prediction value of the previous control cycle; The coefficient matrix predictions based on the linear part of the current control cycle The linear part prediction value of the current control period is determined by using the historical data matrix.

6. The control method of a power converter based on subspace predictive control according to claim 4, characterized in that: The first nonlinear function and the second nonlinear function are estimated by using a neural network to determine a nonlinear part prediction value of a current control cycle, including: Define the weight matrix of the neural network and define the activation function of the neural network with current as parameter; Constructing a first nonlinear function of a single nonlinear term according to the weight matrix and the activation function; Constructing a second nonlinear function of the output current of the power converter according to the linear part prediction value of the current control cycle, the weight matrix and the activation function; Using a neural network to estimate the first nonlinear function and the second nonlinear function, an estimated value of the output current of the current control cycle, an estimated value of the weight matrix of the current control cycle, and estimated values ​​of each single nonlinear term of the current control cycle; Based on the estimated values ​​of each single nonlinear term of the current control cycle, the predicted value of the nonlinear part of the current control cycle is determined.

7. The control method of a power converter based on subspace predictive control according to claim 5, characterized in that: The first nonlinear function and the second nonlinear function are estimated by using a neural network to determine a nonlinear part prediction value of a current control cycle, including: The first nonlinear function of a single nonlinear term d is expressed as: ； ； The second nonlinear function of the output current of the power converter is expressed as: ； in, represents the output current of the power converter, represents the linear part prediction value of the current control cycle, represents the weight matrix of the neural network, represents the activation function of the neural network, represents the minimum error of the nonlinear term estimate, represents the minimum error of output current estimation; The first nonlinear function and the second nonlinear function are estimated using a neural network, and the estimated value of the nonlinear term d is and the estimated value of the output current of the power converter Respectively expressed as: ； ； ； in, , , represents a positive coefficient, represents the estimated value of the linear part prediction value, represents the estimated value of the weight matrix of the neural network, represents the error between the estimated value of the output current of the power converter and the actual output current of the power converter, represents the derivative of the estimated value of the weight matrix of the neural network; The estimated value of the nonlinear term d , estimated value of output current of power converter and the derivative of the estimate of the weight matrix Perform first-order Euler discretization respectively and obtain the following expressions: ； in, represents the estimated value of the output current in the current control cycle, represents the estimated value of the weight matrix of the current control cycle, represents the estimated value of the linear part prediction value of the current control period, represents the length of the control cycle, represents the estimated value of a single nonlinear term for the current control period; The nonlinear part of the subspace predictor model predicted value for the current control period It is expressed as: 。 8. The control method of a power converter based on subspace predictive control according to claim 7, characterized in that: The step of determining a functional relationship for predicting the output current in a future time window based on the linear part prediction value of the current control cycle, the nonlinear part prediction value of the current control cycle, and the subspace predictor model includes: Substitute the linear part prediction value of the current control cycle and the nonlinear part prediction value of the current control cycle into the expression of the subspace predictor model to obtain the predicted output current matrix in the future time window: The functional relationship is: ； ； ； in, represents an m×m dimensional matrix, the diagonal elements of the matrix are 1, and the remaining elements are 0, and m represents the data dimension of the system input and output; The step of constructing an objective function based on the functional relationship and the expected output current in a preset future time window, and determining the optimal input voltage of the current control cycle based on the objective function includes: The predicted output current matrix The objective function is constructed by using the functional relationship of the expected output current matrix in the preset future time window. It is expressed as: ； ； in, represents the expected output current matrix within the preset future time window, represents the expected output current at time k+1 obtained based on vector angle extrapolation, represents the input voltage matrix in the future time window, and represents the diagonal weight matrix; when According to the objective function Get the input voltage increment matrix in the future time window for: ； ； ； by The sum of the first matrix element of and the optimal input voltage of the previous control cycle is used as the optimal input voltage to the current control cycle. .

9. The control method of a power converter based on subspace predictive control according to claim 8, characterized in that: The cost function is constructed based on the optimal input voltage of the current control cycle and the voltage vector corresponding to each switch state of the power converter, including: The cost function is constructed by using the optimal input voltage of the current control cycle and the voltage vector corresponding to each switching state of the power converter. The expression is: ； in, represents the voltage vector corresponding to each switching state of the power converter, and n represents n switching states; Based on the cost function Calculate the optimal input voltage for each voltage vector and the current control cycle The square of the absolute value between and is used to determine the voltage vector corresponding to the minimum value, and the switching state corresponding to the voltage vector is used as the optimal switching state of the power converter in the current control cycle.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, the steps of the method for controlling a power converter based on subspace predictive control as described in any one of claims 1 to 9 are implemented.

Images