Online data-driven predictive control methods, devices and electronic equipment for converters

By constructing a data-driven prediction model, utilizing historical current tracking errors and voltage increments, defining the future control increment trajectory, calculating free errors and forced response coefficients, and optimizing control quantities, the problems of decreased prediction accuracy and high computational complexity caused by converter parameter dependence are solved, and real-time control at high switching frequencies is realized.

CN123074249APending Publication Date: 2026-11-10ZHEJIANG UNIV
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611590950.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-14
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing predictive control methods for converters rely on physical parameters, which leads to decreased prediction accuracy when parameters are mismatched. They also have high computational complexity and are difficult to meet the real-time control requirements at high switching frequencies.

Method used

A data-driven prediction model is constructed, using historical current tracking error increments and voltage increments as inputs to define the future control increment trajectory. By recursively calculating the free error trajectory and forced response coefficients, a dynamic matrix is ​​constructed, a prediction equation for the future total error trajectory is established, the control quantity is optimized, and the optimal switching state is selected.

Benefits of technology

Without relying on physical parameters, the online computation load is reduced, real-time control of the converter at high switching frequencies is achieved, and prediction accuracy and control efficiency are improved.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123074249A_ABST
    Figure CN123074249A_ABST
Patent Text Reader

Abstract

This application discloses an online data-driven predictive control method, device, and electronic equipment for converters. The core of this invention is parameter robustness, noise suppression, and kernel function nonlinear capture: A data-driven predictive model is constructed using historical current tracking error increments and voltage increments as inputs, independent of converter physical parameters and avoiding the effects of parameter mismatch; an observer is introduced before modeling to smoothly reconstruct noisy currents to suppress measurement noise; the kernel function concept is used to explicitly upgrade the linear input to a nonlinear high-dimensional feature space, combined with online learning to capture system nonlinear dynamics such as dead-zone effects and tube voltage drops; and high-precision current tracking is achieved while maintaining low computational complexity during online operation by replacing the Diophantine equation with an online recursive dynamic matrix.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of converter control, and in particular to an online data-driven predictive control method, apparatus, and electronic device for converters. Background Technology

[0002] Finite set model predictive control (FCS-MPC) has gained widespread attention and application in the field of converter control due to its advantages such as fast dynamic response, strong multi-objective optimization capability, and ease of handling nonlinear constraints.

[0003] However, existing technologies still face challenges such as parameter dependence and high online computation costs: While prediction methods based on hyperlocal models simplify the model structure to some extent, they still rely on key electrical parameters such as inductance and resistance to establish the prediction model. In actual operation, parameters such as inductance will deviate significantly due to rising temperature, device aging, magnetic circuit saturation, and frequent switching of operating conditions. When parameters are mismatched or completely unknown, the prediction accuracy will drop sharply, leading to increased current harmonics, deterioration of steady-state error, and even system instability. To overcome the impact of parameter mismatch, data-driven methods attempt to approximate system behavior through online identification or recursive computation. However, their online operation requires explicit solution of Diophantine equations, which involves recursive computation and polynomial operations. The computational complexity increases significantly with the increase of the prediction step size, making it difficult to meet the real-time control requirements of power electronic converters at high switching frequencies.

[0004] Therefore, there is an urgent need for a predictive control method that does not rely on physical parameters and has low online computational costs. Summary of the Invention

[0005] This application provides an online data-driven predictive control method, apparatus, and electronic device for converters, which reduces the amount of online computation without relying on the physical parameters of the converter.

[0006] In a first aspect, this application provides an online data-driven predictive control method for converters, comprising: A data-driven prediction model is constructed, which uses the historical current tracking error increment vector and the historical voltage increment vector as input variables, and the current tracking error increment at the next moment as the prediction object. Define future control incremental trajectory Future control of incremental trajectory The voltage increment input variables in the data-driven prediction model consist of a sequence in the future prediction time domain; Using a data-driven predictive model, calculate the future control increment trajectory The free error increment trajectory under the condition that all voltage increments are zero and the corresponding free error trajectory Free error trajectory Configured as the current current tracking error trajectory of future free error increment The accumulation; Apply a unit perturbation to the voltage increment at the first moment in the future prediction time domain, and keep the voltage increment at subsequent moments zero. Use the data-driven prediction model to recursively calculate the predicted value of the current tracking error at each step in the future prediction time domain after the perturbation is injected. The forced response coefficient for each step is calculated by comparing the current tracking error at each step after the perturbation with the free response without perturbation. The free response is traversed via the free error trajectory. Obtain the forced response coefficient Characterize the effect of applying a perturbation at the current moment on the future... The impact of current tracking error in the step; Based on forced response coefficient Construct a dynamic matrix describing the impact of voltage increments on future multi-step current tracking errors. and the free error trajectory and dynamic matrix Combine to establish the future total error trajectory The prediction equation; Construct a trajectory containing the total future error With future control incremental trajectory The objective function is obtained by minimizing the objective function to obtain the optimal future control increment sequence. In order to obtain the optimal control quantity at the current moment. ; From the finite set of switching states of the converter, the switching state that is closest to the optimal control quantity is selected based on a preset value function and applied to the converter.

[0007] In one embodiment, a dynamic matrix is ​​constructed to describe the impact of voltage increments on future multi-step current tracking errors. ,include: Set the length of the future prediction time domain to be In the future, control the incremental trajectory Under the condition that all voltage increments are zero, the free error increments at each time point in the future prediction time domain are recursively calculated using a data-driven prediction model, and the free error increment trajectory is obtained by arranging them in chronological order. ; Based on free error increment trajectory Compared with the current current tracking error The free error trajectory was calculated. Free error trajectory This includes the free error prediction values ​​for each step within the future prediction time domain; Controlling Incremental Trajectory in the Future Apply a unit perturbation to the voltage increment at the first moment. And keep the voltage increment at the remaining time points in the future prediction time domain zero; The injected unit perturbation is calculated recursively using a data-driven prediction model. The predicted value of the current tracking error at each step in the post-future prediction time domain; The difference between the predicted current tracking error after injecting a unit perturbation and the corresponding predicted free error is calculated and then divided by the unit perturbation. The forced response coefficients at each step in the future prediction time domain are calculated. ; Arrange the forced response coefficients into a lower triangular matrix to obtain the dynamic matrix. Dynamic matrix The elements on and below the diagonal are the forced response coefficients. Elements above the diagonal are zero.

[0008] In one embodiment, based on the free error increment trajectory Compared with the current current tracking error The free error trajectory was calculated. ,include: Construct a lower triangular accumulation matrix Among them, the lower triangular accumulation matrix for In a 3D matrix, the elements at and below the diagonal of the lower triangular accumulator matrix are all 1, and the elements above the diagonal are all 0. Based on the current current tracking error Free error increment trajectory and lower triangular accumulation matrix The free error trajectory was calculated. : ; in, for A column vector of all 1s.

[0009] In one embodiment, constructing the objective function and solving for the optimal control quantity at the current time includes: Construct the objective function objective function For the future total error trajectory The inner product and the future control increment trajectory after regularization The sum of the inner products; Let the objective function Controlling the incremental trajectory of the future The partial derivatives are zero, and the optimal future control increment sequence, including the regularization term, is obtained by solving. ; The optimal future control increment sequence is obtained by using a preset initial selection vector. Perform weighted extraction to obtain the weighted control increment at the current time. Then, compare the weighted control increment at the current time with the voltage at the previous time. Add them together to obtain the optimal control quantity at the current moment. The coefficient of the first dimension of the initial selection vector is 1, and the coefficients of the remaining dimensions are all 0.

[0010] In one embodiment, selecting the switching state closest to the optimal control quantity from a finite set of switching states of the converter based on a preset value function includes: Based on the optimal future control increment sequence Determine the optimal control quantity for the next time step. ; Candidate switching states are selected sequentially from the finite set of switching states of the converter, and the candidate voltage vector corresponding to each candidate switching state is determined. and the midpoint voltage at the next moment ; Based on the optimal control quantity at the next time step With candidate voltage vector The tracking error term is calculated. ; Based on the midpoint voltage at the next time step The midpoint potential deviation term was calculated. ; Based on candidate switch states With the current switch state The switching frequency term was calculated. ; Tracking error term Midpoint potential deviation term and switching frequency term Multiply each value by its corresponding weighting coefficient and then sum them to obtain the value function; Select the switching state corresponding to the candidate voltage vector that minimizes the value function and apply it to the converter.

[0011] In one embodiment, noise suppression processing is also included on the historical current tracking error increment vector before constructing the data-driven prediction model: Obtain the noisy current value of the converter at the current moment. and the reference current at the current moment Calculate the original noisy current error ; Based on the historical current tracking error increment vector and the historical voltage increment vector, calculate the prior prediction increment of the current tracking error at the current moment. ; Calculate the estimated current tracking error after noise suppression at the previous time step. ( k- 1) Incremental with prior prediction The sum of these values ​​yields the prior prediction value of the current tracking error at the current moment. ; Introducing observer correction gain Utilizing the original noisy current error Prior prediction of the current tracking error at the current moment After correction, the current error estimate after noise suppression is obtained at the current time. ( k ); Based on the current error estimate after noise suppression at the current moment ( k ) Calculate the effective increment of error The historical current tracking error increment vector after noise suppression is obtained by utilizing the effective error increment. .

[0012] In one embodiment, the parameter vector of the data-driven prediction model is updated using an online learning method based on kernel functions, including: The historical current tracking error increment vector and the historical voltage increment vector after noise suppression are merged into a historical input-output data vector. Using the kernel function concept, the historical input and output data vectors are subjected to a three-dimensional up-mapping mapping to obtain a high-dimensional nonlinear mapping vector. The obtained high-dimensional nonlinear mapping vector is concatenated with the historical input and output data vectors to obtain an extended input vector. The extended input vector maps the original linear input space to a high-dimensional feature space containing cubic nonlinear terms. Based on the extended input vector and the parameter vector of the data-driven prediction model, a prediction model for the current tracking error increment at the current moment is constructed, thereby improving the data-driven prediction model from a linear model to a nonlinear model. The prediction residual at the current moment is calculated based on the effective error increment and the current tracking error increment at the current moment. The normalized least mean square algorithm is used to update the parameter vector of the data-driven prediction model online based on the prediction residual and the extended input vector, so as to track the system operating point and nonlinear dynamic changes.

[0013] In one embodiment, the converter is any one of a three-level midpoint clamping converter, a two-level converter, a multi-level converter, or a motor driver.

[0014] Secondly, this application also provides an online data-driven converter predictive control device, comprising: The model building module is used to build a data-driven prediction model. The data-driven prediction model uses the historical current tracking error increment vector and the historical voltage increment vector as input variables, and the current tracking error increment at the next moment as the prediction object. The trajectory definition module is used to define future control incremental trajectories. Future control of incremental trajectory The voltage increment input variables in the data-driven prediction model consist of a sequence in the future prediction time domain; The free response calculation module is used to calculate the future control increment trajectory using a data-driven predictive model. The free error increment trajectory under the condition that all voltage increments are zero and the corresponding free error trajectory Free error trajectory Configured as the current current tracking error trajectory of future free error increment The accumulation; The disturbance response module is used to apply a unit disturbance to the voltage increment at the first moment in the future prediction time domain and keep the voltage increment at subsequent moments zero. It uses a data-driven prediction model to recursively calculate the current tracking error at each step in the future prediction time domain after the disturbance is injected. The matrix construction module is used to calculate the forced response coefficients at each step by comparing the current tracking error at each step after the injected perturbation with the free response under no perturbation. The free response is traversed via the free error trajectory. Obtain the forced response coefficient Characterize the effect of applying a perturbation at the current moment on the future... The impact of current tracking error in the step; based on the forced response coefficient Construct a dynamic matrix describing the impact of voltage increments on future multi-step current tracking errors. and the free error trajectory and dynamic matrix Combine to establish the future total error trajectory The prediction equation; The optimization solution module is used to construct a trajectory containing the total future error. With future control incremental trajectory The objective function is obtained by minimizing the objective function to obtain the optimal future control increment sequence. In order to obtain the optimal control quantity at the current moment. ; The switch selection module is used to select the switch state that is closest to the optimal control quantity from the limited set of switch states of the converter based on a preset value function, and then apply it to the converter.

[0015] Thirdly, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the online data-driven converter predictive control method of the first aspect.

[0016] The aforementioned online data-driven converter predictive control method constructs a data-driven predictive model with historical current tracking error increment vectors and historical voltage increment vectors as inputs and the next-time current tracking error increment as the prediction object. It defines the future control increment trajectory and recursively calculates the free error increment trajectory and its corresponding free error trajectory under the condition that all voltage increments are zero. A unit disturbance is applied to the voltage increment at the first time step, and the current tracking error after the disturbance is recursively calculated. The forced response coefficient is obtained by comparing it with the free response, and then a dynamic matrix is ​​constructed to establish the predictive equation for the future total error trajectory. An objective function containing the future total error trajectory and the future control increment trajectory is constructed, and the optimal control quantity is obtained by minimizing the objective function. The closest switching state is selected from a finite set of switching states based on the value function and applied to the converter. This method, without relying on the converter's physical parameters, replaces the complex process of explicitly solving the Diophantine equations in traditional generalized predictive control by recursively constructing a dynamic matrix online, reducing the online computational load and achieving real-time control requirements for converters at high switching frequencies. Attached Figure Description

[0017] Figure 1 This is a flowchart of an online data-driven converter predictive control method in one embodiment; Figure 2 A flowchart for constructing a dynamic matrix describing the impact of voltage increments on future multi-step current tracking errors in one embodiment; Figure 3 This is a flowchart illustrating the selection of the switching state closest to the optimal control quantity based on a preset value function in one embodiment. Figure 4 This is a flowchart of the historical current tracking error increment vector noise suppression process in one embodiment; Figure 5 This is a control effect diagram of the traditional FCS-MPC method in one embodiment; Figure 6 This is a control effect diagram of an online data-driven converter predictive control method in one embodiment; Figure 7 This is a result diagram of an online data-driven converter predictive control device in one embodiment. Detailed Implementation

[0018] The present application will be described in detail below with reference to the specific embodiments shown in the accompanying drawings. However, these embodiments do not limit the present application. Any structural, methodological, or functional modifications made by those skilled in the art based on these embodiments are included within the protection scope of the present application.

[0019] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0020] In one embodiment, such as Figure 1 As shown, an online data-driven predictive control method for converters is provided, comprising the following steps: Step 101: Construct a data-driven prediction model. The data-driven prediction model uses the historical current tracking error increment vector and the historical voltage increment vector as input variables, and the current tracking error increment at the next moment as the prediction object. Specifically, in the embodiments of this application, the current tracking error is defined as the reference current. With actual current The difference is calculated as follows: ; in, The reference current at time k for The actual current at any given moment.

[0021] Define the current tracking error increment for Current tracking error at time and The difference in current tracking error at time: .

[0022] Based on the current tracking error increment at each time step in the historical error increment vector dimension, a historical current tracking error increment vector is constructed: ; in, is the dimension of the historical error increment vector.

[0023] Based on the voltage increments at each time point in the historical voltage increment vector dimension, construct the historical voltage increment vector: in, , is the dimension of the historical voltage increment vector.

[0024] Furthermore, a data-driven prediction model can be constructed to output the current tracking error increment at the next time step: ; in, The coefficient vector corresponding to the historical error increment vector. The coefficient vector corresponding to the historical voltage increment vector, and It can be updated in real time using the online recursive least squares method; This is the noise term.

[0025] The data-driven prediction model uses the historical current tracking error increment vector and historical voltage increment vector As an input variable, the current tracking error increment at the next moment. As the object of prediction, the prediction process of the data-driven prediction model does not depend on any converter physical parameters throughout the entire prediction process.

[0026] Step 102: Define the future control incremental trajectory Future control of incremental trajectory The voltage increment input variables in the data-driven prediction model consist of a sequence in the future prediction time domain; Specifically, in the embodiments of this application, the future control incremental trajectory It can be defined as starting from the current moment. Initially, the sequence vector composed of voltage increments at each time point in the future prediction time domain: .

[0027] in, Indicates in Time prediction Voltage increment at time t, Each element in this sequence vector corresponds to the value of the voltage increment input variable in the data-driven prediction model at future time points.

[0028] Step 103: Utilize a data-driven predictive model to calculate the future control increment trajectory. The free error increment trajectory under the condition that all voltage increments are zero and the corresponding free error trajectory Free error trajectory Configured as the current current tracking error trajectory of future free error increment The accumulation; Specifically, to control the incremental trajectory in the future. All elements in the matrix are zero. The current tracking error increment for the first future step is calculated using a data-driven prediction model. Further, the current tracking error increment for the first future step is used as historical data to update the historical error increment vector. Based on the updated historical error increment vector, the current tracking error increment for the second future step is recursively obtained; this process continues until the current tracking error increments for all steps within the future prediction time domain are obtained. The current tracking error increments for all steps are then arranged in chronological order to form the free error increment trajectory. .

[0029] Furthermore, based on the current current tracking error with free error increment trajectory The free error trajectory is obtained through cumulative calculation. .

[0030] Step 104: Apply a unit perturbation to the voltage increment at the first moment in the future prediction time domain, and keep the voltage increment at subsequent moments zero. Use the data-driven prediction model to recursively calculate the predicted value of the current tracking error at each step in the future prediction time domain after the perturbation. Specifically, controlling the incremental trajectory towards the future. Apply a unit perturbation to the voltage increment at the first moment in the future prediction time domain. And keep the voltage increment at the remaining time points in the future prediction time domain zero, that is, let .

[0031] Calculate the injected unit perturbation using a data-driven prediction model. The predicted current tracking error for the first future step is then calculated. This first-step current tracking error is used as historical data to update the historical error increment vector. Based on this updated historical error increment vector, the predicted current tracking error for the second future step is recursively calculated. This process continues until the injected unit perturbation is obtained. The predicted current tracking error for all steps in the future prediction time domain.

[0032] Step 105: Calculate the forced response coefficient for each step by comparing the current tracking error after the injected perturbation with the free response under no perturbation. The free response is traversed via the free error trajectory. Obtain the forced response coefficient Characterize the effect of applying a perturbation at the current moment on the future... The impact of current tracking error in the step; Specifically, for the first term in the future prediction time domain... Step, inject unit perturbation The predicted current tracking error and the free error trajectory in this step The corresponding number in The difference between the free error predictions of each step is calculated, and this difference is divided by the unit perturbation. The forced response coefficients for this step are obtained. ,in, =1, 2, ..., representing the step number in the future prediction time domain.

[0033] The forced response coefficient This indicates that after applying a unit perturbation to the voltage increment of the first time step at the current time step, the voltage increment of the next time step will be affected by the current perturbation. The extent of the impact caused by the current tracking error of the step. For example, This indicates the impact of the current disturbance on the current tracking error at the next moment. This indicates the effect on the next time step, and so on.

[0034] The forced response coefficients for each step in the future prediction time domain are obtained through the above calculations. , ...

[0035] Step 106: Based on the forced response coefficient Construct a dynamic matrix describing the impact of voltage increments on future multi-step current tracking errors. and the free error trajectory and dynamic matrix Combine to establish the future total error trajectory The prediction equation; Specifically, the forced response coefficients at each step Arranged in order of the number of steps in the future prediction time domain, they form a lower triangular matrix, resulting in the dynamic matrix. The dynamic matrix The elements on and below the middle diagonal are the forced response coefficients for each step. The elements above the diagonal are zero, which are used to describe the superposition effect of different voltage increments applied at the current moment on the current tracking error in future steps.

[0036] Furthermore, based on the free error trajectory With dynamic matrix Multiply by the future control increment trajectory Add the results together to establish the future total error trajectory. The prediction equation.

[0037] Step 107: Construct a trajectory containing the total future error With future control incremental trajectory The objective function is obtained by minimizing the objective function to obtain the optimal future control increment sequence. In order to obtain the optimal control quantity at the current moment. ; Specifically, based on the established future total error trajectory The prediction equation is used to construct the objective function. The objective function It includes two items: the first is the future total error trajectory. The inner product is used to minimize the current tracking error; the second term is the future control increment trajectory multiplied by the regularization coefficient λ. The inner product is used to limit the magnitude of changes in the control increment, thus preventing drastic fluctuations in the control quantity.

[0038] Let the objective function Controlling the incremental trajectory of the future The partial derivatives are zero, and the optimal future control increment sequence including the regularization term is obtained by solving. .

[0039] The optimal future control increment sequence is obtained by using a preset initial selection vector. Weighted extraction is performed to obtain the weighted control increment at the current time step. The first dimension coefficient of the initial selection vector is 1, and the coefficients of the remaining dimensions are all 0. The weighted control increment at the current time step is then compared with the voltage at the previous time step. Add them together to obtain the optimal control quantity at the current moment. .

[0040] Step 108: From the finite set of switching states of the converter, select the switching state that is closest to the optimal control quantity based on the preset value function, and apply it to the converter.

[0041] Specifically, the converter has a finite number of switching states, each corresponding to a specific voltage vector. The approximation between the voltage vector corresponding to each candidate switching state and the optimal control variable at the current moment is evaluated one by one.

[0042] The value function is used to comprehensively measure the degree of matching between the candidate switch state and the optimal control quantity. The smaller the value function value, the closer the candidate switch state is to the optimal control quantity.

[0043] Select the switching state that minimizes the value function from all candidate switching states, and use it as the optimal switching state at the current moment. Apply this switching state to the converter to achieve real-time control of the converter.

[0044] In this embodiment, the method constructs a data-driven prediction model with historical current tracking error increment vectors and historical voltage increment vectors as inputs and the next time-instance current tracking error increment as the prediction object; defines the future control increment trajectory, and recursively calculates the free error increment trajectory and the corresponding free error trajectory under the condition that the voltage increments are all zero; applies a unit disturbance to the voltage increment at the first time-instance, recursively calculates the current tracking error after the disturbance, obtains the forced response coefficient by comparing it with the free response, and then constructs a dynamic matrix and establishes the prediction equation for the future total error trajectory; constructs an objective function containing the future total error trajectory and the future control increment trajectory, and obtains the optimal control quantity by minimizing the solution; and selects the closest switching state from the finite set of switching states based on the value function to act on the converter. This method, without relying on the converter's physical parameters, replaces the complex process of explicitly solving the Diophantine equations in traditional generalized predictive control by online recursive construction of a dynamic matrix, reducing the online computation load and realizing the real-time control requirements of the converter at high switching frequencies.

[0045] In one embodiment, such as Figure 2 As shown, a dynamic matrix describing the impact of voltage increments on future multi-step current tracking errors is constructed. This includes the following steps: Step 201: Set the length of the future prediction time domain to be... In the future, control the incremental trajectory Under the condition that all voltage increments are zero, the free error increments at each time point in the future prediction time domain are recursively calculated using a data-driven prediction model, and the free error increment trajectory is obtained by arranging them in chronological order. ; Step 202: Based on the free error increment trajectory Compared with the current current tracking error The free error trajectory was calculated. Free error trajectory This includes the free error prediction values ​​for each step within the future prediction time domain; Step 203: Controlling the incremental trajectory towards the future Apply a unit perturbation to the voltage increment at the first moment. And keep the voltage increment at the remaining time points in the future prediction time domain zero; Step 204: Recursively calculate the injected unit perturbation using a data-driven prediction model. The predicted value of the current tracking error at each step in the post-future prediction time domain; Step 205: Subtract the predicted current tracking error value after injecting a unit perturbation from the corresponding predicted free error value, and divide by the unit perturbation. The forced response coefficients at each step in the future prediction time domain are calculated. ; Step 206: Arrange the forced response coefficients into a lower triangular matrix to obtain the dynamic matrix. Dynamic matrix The elements on and below the diagonal are the forced response coefficients. Elements above the diagonal are zero.

[0046] Specifically, the length of the future prediction time domain is set as... ,make Using a data-driven prediction model, the free error increment for the first step is obtained as follows: .

[0047] Free error increment in step 1 The updated historical error increment vector is then incorporated into the data, resulting in the updated historical error increment vector. Furthermore, based on the updated historical error increment vector, a data-driven prediction model is used to obtain the predicted free error increment value for the second step. Similarly, the free error increment at each time point in the future prediction time domain is calculated recursively. , , The free error increment trajectories are obtained by arranging them in chronological order. : .

[0048] Define the lower triangular accumulation matrix as The lower triangular accumulation matrix is 3D matrix: .

[0049] The formula shows that the lower triangular accumulation matrix... All elements at or below the diagonal are 1, and all elements above the diagonal are 0.

[0050] Furthermore, based on the free error increment trajectory Compared with the current current tracking error The free error trajectory is obtained through lower triangular accumulation matrix operations. Free error trajectory The calculation formula is: .

[0051] in, for A column vector of all 1s, i.e. .

[0052] It should be noted that the free error trajectory This includes the predicted free error value at each step within the future prediction time domain. In other words, the predicted free error value at each step within the future prediction time domain is... , ,

[0053] Going further, controlling the incremental trajectory in the future. Apply a unit perturbation to the voltage increment at the first moment. And keep the voltage increment at all other times in the future prediction time domain zero. The injected unit perturbation is recursively calculated using a data-driven prediction model. Predicted current tracking error at each step in the post-future prediction time domain , , .

[0054] Furthermore, based on the current tracking error after injecting a unit perturbation With the corresponding free error prediction value Divide by the unit disturbance. The forced response coefficients for each step in the future prediction time domain are calculated: .

[0055] Ultimately, the forced response coefficient will be... Arranged into a lower triangular matrix, the dynamic matrix is ​​obtained. : .

[0056] The dynamic matrix The elements on and below the diagonal are the forced response coefficients. The elements above the diagonal are zero, which describes the superposition effect of different voltage increments applied at the current moment on the current tracking error in future steps.

[0057] In this embodiment, by setting the future prediction time domain, the free error increment trajectory is recursively obtained under zero control conditions. Then, a unit perturbation is applied to the voltage increment at the first time step, and the predicted current tracking error after the perturbation is recursively obtained. The forced response coefficient is obtained by subtracting the perturbation and arranging the coefficients to obtain the lower triangular dynamic matrix. This eliminates the need to explicitly solve the Diophantine equations in traditional generalized predictive control; the dynamic matrix can be constructed online with only two model recursions and one arithmetic operation, reducing the amount of online computation.

[0058] In one embodiment, constructing the objective function and solving for the optimal control quantity at the current time includes: Construct the objective function objective function For the future total error trajectory The inner product and the future control increment trajectory after regularization The sum of the inner products; Let the objective function Controlling the incremental trajectory of the future The partial derivatives are zero, and the optimal future control increment sequence, including the regularization term, is obtained by solving. ; The optimal future control increment sequence is obtained by using a preset initial selection vector. Perform weighted extraction to obtain the weighted control increment at the current time. Then, compare the weighted control increment at the current time with the voltage at the previous time. Add them together to obtain the optimal control quantity at the current moment. The coefficient of the first dimension of the initial selection vector is 1, and the coefficients of the remaining dimensions are all 0.

[0059] Specifically, based on free error trajectories With dynamic matrix and future control incremental trajectory The product of these factors is used to establish the future total error trajectory. The prediction equation: .

[0060] objective function It consists of the sum of the inner product of the future total error trajectory and the inner product of the regularized future control increment trajectory: ; in, This is the regularization coefficient, used to balance the weight between minimizing error and controlling the change in increment.

[0061] Let the objective function Controlling the incremental trajectory of the future The partial derivatives are zero, even if we get The optimal future control increment sequence, including the regularization term, is obtained by solving the problem. : ; in, for 3D identity matrix As a regularization term, it ensures that the matrix The reversibility of this avoids It is either strange or pathological, while limiting and controlling incremental changes.

[0062] By using a preset initial selection vector For the optimal future control increment sequence Perform weighted extraction to obtain the weighted control increment at the current time step. Initial selection vector. for A column vector with the first dimension having a coefficient of 1 and the coefficients of the remaining dimensions being 0. .

[0063] The weighted control increment at the current time is: The weighted control increment at the current time will be... Voltage compared to the previous moment Add them together to obtain the optimal control value at the current moment: .

[0064] The optimal control quantity This is the optimal voltage value that is expected to be applied to the converter at the current moment, which is used for subsequent selection of finite switching states.

[0065] In one embodiment, such as Figure 3 As shown, from the finite set of switching states of the converter, the switching state closest to the optimal control quantity is selected based on a preset value function, including: Step 301: Based on the optimal future control increment sequence Determine the optimal control quantity for the next time step. ; Step 302: Select candidate switching states sequentially from the finite set of switching states of the converter, and determine the candidate voltage vector corresponding to each candidate switching state. and the midpoint voltage at the next moment ; Step 303: Based on the optimal control quantity at the next time step With candidate voltage vector The tracking error term is calculated. ; Step 304: Based on the midpoint voltage at the next time step The midpoint potential deviation term was calculated. ; Step 305: Based on candidate switch states With the current switch state The switching frequency term was calculated. ; Step 306: Include the tracking error term Midpoint potential deviation term and switching frequency term Multiply each value by its corresponding weighting coefficient and then sum them to obtain the value function; Step 307: Select the switching state corresponding to the candidate voltage vector that minimizes the value function and apply it to the converter.

[0066] Specifically, based on the optimal future control increment sequence From the optimal future control increment sequence Extract the optimal control corresponding to the next time step. .

[0067] For a three-level NPC converter, the finite set of switching states contains 27 switching states, and each switching state corresponds to a set of three-phase switching signals. Candidate switching states are selected sequentially from the finite set of switching states of the converter. And based on the DC bus voltage Determine the candidate voltage vector corresponding to each candidate switch state. .

[0068] The midpoint voltage is determined by the voltage difference between the upper and lower capacitors on the DC side. Let the voltage of the upper capacitor be... The lower capacitor voltage is Current switch state The current midpoint voltage is: Based on the candidate switch state Predict the midpoint voltage at the next moment. .

[0069] Furthermore, based on the optimal control quantity at the next time step With candidate voltage vector Calculate the tracking error term : .

[0070] Based on the midpoint voltage at the next time step Calculate the midpoint potential deviation term : .

[0071] Based on candidate switch states With the current switch state Calculate the switching frequency term : .

[0072] Tracking error term Midpoint potential deviation term and switching frequency term The value function is obtained by multiplying each value by its corresponding weight coefficient and then summing the results. value function The calculation formula is as follows: .

[0073] in, , , These are the weighting coefficients for the tracking error term, the midpoint potential deviation term, and the switching frequency term, respectively, used to balance the priorities of the three control objectives.

[0074] Iterate through all candidate switch states, calculate the value function value corresponding to each candidate switch state, select the switch state corresponding to the candidate voltage vector that minimizes the value function, and apply it to the converter to achieve real-time control of the converter.

[0075] In one embodiment, such as Figure 4 As shown, before constructing the data-driven prediction model, noise suppression processing is also performed on the historical current tracking error increment vector. The noise suppression processing includes the following steps: Step 401: Obtain the current value of the converter with noise at the current moment. and the reference current at the current moment Calculate the original noisy current error ; Step 402: Based on the historical current tracking error increment vector and the historical voltage increment vector, calculate the prior prediction increment of the current tracking error at the current moment. ; Step 403: Calculate the estimated current tracking error after noise suppression processing at the previous time step. ( k- 1) Incremental with prior prediction The sum of these values ​​yields the prior prediction value of the current tracking error at the current moment. ; Step 404: Introduce an observer to correct the gain Utilizing the original noisy current error Prior prediction of the current tracking error at the current moment After correction, the current error estimate after noise suppression is obtained at the current time. ( k ); Step 405: Estimate the current error based on the noise suppression processing value at the current moment. ( k ) Calculate the effective increment of error The historical current tracking error increment vector after noise suppression is obtained by utilizing the effective error increment. .

[0076] Specifically, obtain the noisy current value of the converter at the current moment. and the reference current at the current moment Calculate the original noisy current error : ; in, The current value acquired by the sensor, which contains measurement noise. , The given reference current value.

[0077] Furthermore, based on the historical current tracking error increment vector and the historical voltage increment vector, the prior prediction increment of the current tracking error at the current moment is calculated. : .

[0078] Furthermore, the estimated current tracking error after noise suppression processing at the previous time step is calculated. ( k- 1) Incremental with prior prediction The sum of these values ​​yields the prior prediction value of the current tracking error at the current moment. : ( k- 1) .

[0079] Furthermore, an observer-corrected gain is introduced. Utilizing the original noisy current error Prior prediction of the current tracking error at the current moment After correction, the current error estimate after noise suppression is obtained at the current time. : .

[0080] in, Correct the gain for the observer, with a value ranging from 0 to 1. <1, used to balance the confidence level between prior predictions and measured values. The larger the value, the higher the confidence level in the measured value, and the weaker the filtering effect. The smaller the value, the higher the confidence in the prior prediction, and the stronger the filtering effect.

[0081] Based on the current error estimate after noise suppression at the current moment Calculate the effective increment of the error: .

[0082] Utilizing this effective error increment Update the historical current tracking error increment vector to obtain the historical current tracking error increment vector after noise suppression: .

[0083] in, This represents the dimension of the historical error increment vector. The historical current tracking error increment vector after noise suppression processing replaces the original historical current tracking error increment vector, improving the controller's ability to resist measurement noise.

[0084] Furthermore, based on the historical current tracking error increment vector after noise suppression processing This allows us to obtain a data-driven prediction model after noise suppression: .

[0085] In this embodiment, noise components in the sensor-acquired current can be effectively filtered out, providing a high-quality clean data source for the controller's prediction and parameter identification, and improving the data-driven controller's ability to resist measurement noise.

[0086] In one embodiment, the parameter vector of the data-driven prediction model is updated using an online learning method based on kernel functions, including: The historical current tracking error increment vector and the historical voltage increment vector after noise suppression are merged into a historical input-output data vector. Using the kernel function concept, the historical input and output data vectors are subjected to a three-dimensional up-mapping mapping to obtain a high-dimensional nonlinear mapping vector. The obtained high-dimensional nonlinear mapping vector is concatenated with the historical input and output data vectors to obtain an extended input vector. The extended input vector maps the original linear input space to a high-dimensional feature space containing cubic nonlinear terms. Based on the extended input vector and the parameter vector of the data-driven prediction model, a prediction model for the current tracking error increment at the current moment is constructed, thereby improving the data-driven prediction model from a linear model to a nonlinear model. The prediction residual at the current moment is calculated based on the effective error increment and the current tracking error increment at the current moment. The normalized least mean square algorithm is used to update the parameter vector of the data-driven prediction model online based on the prediction residual and the extended input vector, so as to track the system operating point and nonlinear dynamic changes.

[0087] Specifically, the historical current tracking error increment vector and the historical voltage increment vector, after noise suppression processing, are merged into a historical input-output data vector. Historical input / output data vector The calculation formula is as follows: .

[0088] To enhance the ability to capture the nonlinear dynamics of the system, this application employs the kernel function concept, introducing a three-dimensional upscaling mapping. The historical input and output data vectors are then subjected to a three-dimensional upscaling mapping to obtain... The result obtained after three dimensionality-upgrading mappings With historical input and output data vectors By concatenating the vectors, we obtain the expanded input vector. , The calculation formula is as follows: .

[0089] in, This represents the element-wise cube of a vector. .

[0090] Based on extended input vector With the parameter vector of the data-driven prediction model Calculate the current tracking error increment at the current moment. Prediction model: .

[0091] in, This is the parameter vector for the data-driven prediction model.

[0092] Define the prediction residual as The formula for calculating the prediction residual is the effective increment of the error plus the current tracking error increment at the current moment. The difference is calculated as follows: .

[0093] The parameter vector is processed using NLMS (Normalized Least Mean Square algorithm). Perform online updates to track changes in the system's operating point: .

[0094] in, For learning rate, To prevent positive numbers with excessively small denominators.

[0095] In this embodiment, the method performs a three-dimensional upscaling mapping using a kernel function, extending the linear data-driven model to a nonlinear high-dimensional space. Combined with the normalized least mean square algorithm, the method updates parameters online, thereby improving the controller's ability to capture and track nonlinear dynamics such as dead zone effects and tube voltage drops in the converter system.

[0096] Based on the online data-driven converter predictive control method provided in this application, in order to verify the performance of the control method provided in this application, the control method provided in this application was compared with the traditional finite set model predictive control method (FCS-MPC).

[0097] The system parameters used in the comparative experiment are as follows: AC output frequency of 50Hz, sampling time of 100μs, filter inductance of 10mH, reference current amplitude of 12A, DC side capacitor of 2700μF, load resistance of 1Ω, and DC bus voltage of 110V.

[0098] Under harsh operating conditions with parameter mismatch (inductance parameter mismatch from 10mH to 8mH) and injected random noise of amplitude 0.05, the performance of the two control methods is as follows: Figure 5 and Figure 6 As shown.

[0099] Figure 5 The figure shows the control effect of the traditional FCS-MPC method. As can be seen from the figure, under the combined effect of parameter mismatch and noise interference, there are obvious distortion and harmonic components in the three-phase current waveform of the traditional FCS-MPC method, with a total harmonic distortion (THD) of 3.83%.

[0100] Figure 6 The figure shows the control effect of the control method provided in this application. As can be seen from the figure, under the same harsh operating conditions, the three-phase current waveform of the control method provided in this application is smoother and has a higher sinusoidal degree, with a total harmonic distortion (THD) of only 1.38%.

[0101] The comparison results show that, compared with the traditional FCS-MPC method, the online data-driven converter predictive control method provided in this application reduces THD from 3.83% to 1.38% under harsh conditions of parameter mismatch and noise interference, and reduces harmonic content. This indicates that the control method provided in this application can effectively mitigate the impact of adverse factors such as parameter mismatch and noise interference on control performance, and has stronger robustness and higher control accuracy.

[0102] Furthermore, in some embodiments, the online data-driven converter predictive control method provided in this application can be extended in various ways. For example, the online data-driven converter predictive control method in this application can be combined with multi-vector synthesis technology. Instead of outputting only a single voltage vector in each control cycle, two or three voltage vectors are simultaneously selected for vector synthesis. The switching state corresponding to the synthesized virtual voltage vector is then applied to the converter, thereby effectively improving the quality of the output current.

[0103] Furthermore, the data-driven model prediction architecture in this application is applicable to various power electronic topologies, including two-level converters, modular multilevel converters (MMC), and permanent magnet synchronous motors (PMSM). By simply rewriting the model-based prediction equations into prediction equations based on the generalized data-driven model prediction architecture, the controlled system's anti-interference and anti-parameter variation capabilities can be effectively improved. This application is applicable not only to off-grid operation but also to grid-connected operation. Since the output indicators in grid-connected operation are still physical quantities such as current and voltage, which are the same as the control objectives of this application, this application can still achieve parameter and disturbance estimation, enabling effective operation of the controlled system without requiring system parameters.

[0104] Based on the same concept, such as Figure 7 As shown, this application also provides an online data-driven converter predictive control device, comprising: The model building module 701 is used to build a data-driven prediction model. The data-driven prediction model uses the historical current tracking error increment vector and the historical voltage increment vector as input variables, and the current tracking error increment at the next moment as the prediction object. Trajectory definition module 702 is used to define future control incremental trajectories. Future control of incremental trajectory The voltage increment input variables in the data-driven prediction model consist of a sequence in the future prediction time domain; The free response calculation module 703 is used to calculate the future control increment trajectory using a data-driven predictive model. The free error increment trajectory under the condition that all voltage increments are zero and the corresponding free error trajectory Free error trajectory Configured as the current current tracking error trajectory of future free error increment The accumulation; The disturbance response module 704 is used to apply a unit disturbance to the voltage increment at the first moment in the future prediction time domain and keep the voltage increment at subsequent moments zero. It uses a data-driven prediction model to recursively calculate the current tracking error at each step in the future prediction time domain after the disturbance is injected. Matrix construction module 705 is used to calculate the forced response coefficients at each step by comparing the current tracking error at each step after the injected perturbation with the free response under no perturbation. The free response is traversed via the free error trajectory. Obtain the forced response coefficient Characterize the effect of applying a perturbation at the current moment on the future... The impact of current tracking error in the step; based on the forced response coefficient Construct a dynamic matrix describing the impact of voltage increments on future multi-step current tracking errors. and the free error trajectory and dynamic matrix Combine to establish the future total error trajectory The prediction equation; Optimization solver module 706 is used to construct a trajectory containing the total future error. With future control incremental trajectory The objective function is obtained by minimizing the objective function to obtain the optimal future control increment sequence. In order to obtain the optimal control quantity at the current moment. ; The switch selection module 707 is used to select the switch state that is closest to the optimal control quantity from the limited set of switch states of the converter based on a preset value function, and apply it to the converter.

[0105] Based on the same concept, this application also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement an online data-driven converter predictive control method.

[0106] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of this application is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means.

[0107] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The above are merely preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. An online data-driven predictive control method for converters, characterized in that, include: A data-driven prediction model is constructed, which uses the historical current tracking error increment vector and the historical voltage increment vector as input variables and the current tracking error increment at the next moment as the prediction object. Define future control incremental trajectory The future control incremental trajectory The data-driven prediction model consists of a sequence of voltage increment input variables in the future prediction time domain. Using the data-driven prediction model, the future control increment trajectory is calculated. The free error increment trajectory under the condition that all voltage increments are zero and the corresponding free error trajectory The free error trajectory Configured as the current current tracking error trajectory of future free error increment The accumulation; A unit perturbation is applied to the voltage increment at the first moment in the future prediction time domain, and the voltage increment at subsequent moments is kept to zero. The data-driven prediction model is used to recursively calculate the predicted value of the current tracking error at each step in the future prediction time domain after the perturbation is injected. The forced response coefficient for each step is calculated by comparing the current tracking error at each step after the perturbation with the free response without perturbation. The free response passes through the free error trajectory. The forced response coefficient is obtained. Characterize the effect of applying a perturbation at the current moment on the future... The impact of current tracking error in the step; Based on the forced response coefficient Construct a dynamic matrix describing the impact of voltage increments on future multi-step current tracking errors. and the free error trajectory and dynamic matrix Combine to establish the future total error trajectory The prediction equation; Construct a trajectory containing the total future error With future control incremental trajectory The objective function is defined, and the optimal future control increment sequence is obtained by minimizing the objective function. In order to obtain the optimal control quantity at the current moment. ; From the limited set of switching states of the converter, the switching state that is closest to the optimal control quantity is selected based on a preset value function and applied to the converter.

2. The online data-driven converter predictive control method according to claim 1, characterized in that, Construct a dynamic matrix describing the impact of voltage increments on future multi-step current tracking errors. ,include: The length of the future prediction time domain is set to In the future control increment trajectory Under the condition that all voltage increments are zero, the free error increments at each time point in the future prediction time domain are recursively calculated using the data-driven prediction model, and the free error increment trajectories are obtained by arranging them in chronological order. ; Based on the free error increment trajectory Compared with the current current tracking error The free error trajectory was calculated. The free error trajectory This includes the free error prediction values ​​for each step within the future prediction time domain; To the future control incremental trajectory Apply a unit perturbation to the voltage increment at the first moment. And keep the voltage increment at the remaining time points in the future prediction time domain zero; The injected unit perturbation is recursively calculated using the data-driven prediction model. The predicted value of the current tracking error at each step in the post-future prediction time domain; The difference between the predicted current tracking error after injecting a unit perturbation and the corresponding predicted free error is calculated and then divided by the unit perturbation. The forced response coefficients at each step in the future prediction time domain are calculated. ; Arrange the forced response coefficients into a lower triangular matrix to obtain the dynamic matrix. The dynamic matrix The elements on and below the diagonal are the forced response coefficients. Elements above the diagonal are zero.

3. The online data-driven converter predictive control method according to claim 2, characterized in that, Based on the free error increment trajectory Compared with the current current tracking error The free error trajectory was calculated. ,include: Construct a lower triangular accumulation matrix The lower triangular accumulation matrix for The lower triangular accumulation matrix is ​​a 3D matrix in which all elements at and below the diagonal are 1, and all elements above the diagonal are 0. Based on the current current tracking error The free error increment trajectory and the lower triangular accumulation matrix The free error trajectory is calculated. : ; in, for A column vector of all 1s.

4. The online data-driven converter predictive control method according to claim 1, characterized in that, Constructing the objective function and solving for the optimal control input at the current time includes: Construct the objective function The objective function The future total error trajectory The inner product of the regularized future control increment trajectory The sum of the inner products; Let the objective function Controlling the incremental trajectory of the future The partial derivatives are zero, and the optimal future control increment sequence, including the regularization term, is obtained by solving. ; The optimal future control increment sequence is obtained by using a preset initial selection vector. Perform weighted extraction to obtain the weighted control increment at the current time, and then compare the weighted control increment at the current time with the voltage at the previous time. Add them together to obtain the optimal control quantity at the current moment. The initial selection vector has a first dimension coefficient of 1 and all other dimension coefficients of 0.

5. The online data-driven converter predictive control method according to claim 1, characterized in that, From a finite set of switching states of the converter, the switching state closest to the optimal control quantity is selected based on a preset value function, including: Based on the optimal future control increment sequence Determine the optimal control quantity for the next time step. ; Candidate switching states are selected sequentially from the finite set of switching states of the converter, and the candidate voltage vector corresponding to each candidate switching state is determined. and the midpoint voltage at the next moment ; Based on the optimal control quantity at the next time step With the candidate voltage vector The tracking error term is calculated. ; Based on the midpoint voltage at the next time step The midpoint potential deviation term was calculated. ; Based on the candidate switch states With the current switch state The switching frequency term was calculated. ; The tracking error term The midpoint potential deviation term and the switching frequency term Multiply each value by its corresponding weighting coefficient and then sum them to obtain the value function; The switching state corresponding to the candidate voltage vector that minimizes the value function is selected and applied to the converter.

6. The online data-driven converter predictive control method according to claim 1, characterized in that, Before constructing the data-driven prediction model, noise suppression processing is also performed on the historical current tracking error increment vector: Obtain the noisy current value of the converter at the current moment. and the reference current at the current moment Calculate the original noisy current error ; Based on the historical current tracking error increment vector and the historical voltage increment vector, calculate the prior prediction increment of the current tracking error at the current moment. ; Calculate the estimated current tracking error after noise suppression at the previous time step. ( k - 1) with the prior prediction increment The sum of these values ​​yields the prior prediction value of the current tracking error at the current moment. ; Introducing observer correction gain Using the original noisy current error The prior prediction value of the current tracking error at the current moment After correction, the current error estimate after noise suppression is obtained at the current time. ( k ); Based on the current error estimate after noise suppression processing at the current moment ( k ) Calculate the effective increment of error The historical current tracking error increment vector after noise suppression is obtained using the effective error increment. .

7. The online data-driven converter predictive control method according to claim 6, characterized in that, The parameter vector of the data-driven prediction model is updated using an online learning method based on kernel functions, including: The historical current tracking error increment vector after noise suppression processing and the historical voltage increment vector are merged into a historical input-output data vector. Using the kernel function concept, the historical input and output data vector is subjected to a three-dimensional up-mapping mapping to obtain a high-dimensional nonlinear mapping vector. The obtained high-dimensional nonlinear mapping vector is concatenated with the historical input and output data vector to obtain an extended input vector. The extended input vector maps the original linear input space to a high-dimensional feature space containing cubic nonlinear terms. Based on the extended input vector and the parameter vector of the data-driven prediction model, a prediction model for the current tracking error increment at the current moment is constructed, thereby improving the data-driven prediction model from a linear model to a nonlinear model. The prediction residual at the current moment is calculated based on the effective error increment and the current tracking error increment at the current moment. The normalized least mean square algorithm is used to update the parameter vector of the data-driven prediction model online based on the prediction residual and the extended input vector, so as to track the system operating point and nonlinear dynamic changes.

8. The online data-driven predictive control method for converters according to claim 1, characterized in that, The converter is any one of a three-level midpoint clamping converter, a two-level converter, a multi-level converter, or a motor driver.

9. An online data-driven predictive control device for a converter, characterized in that, include: The model building module is used to build a data-driven prediction model. The data-driven prediction model uses the historical current tracking error increment vector and the historical voltage increment vector as input variables, and the current tracking error increment at the next moment as the prediction object. The trajectory definition module is used to define future control incremental trajectories. The future control incremental trajectory The data-driven prediction model consists of a sequence of voltage increment input variables in the future prediction time domain. The free response calculation module is used to calculate the future control increment trajectory using the data-driven prediction model. The free error increment trajectory under the condition that all voltage increments are zero and the corresponding free error trajectory The free error trajectory Configured as the current current tracking error trajectory of future free error increment The accumulation; The disturbance response module is used to apply a unit disturbance to the voltage increment at the first moment in the future prediction time domain and keep the voltage increment at subsequent moments zero. The data is used to drive the prediction model to recursively calculate the current tracking error at each step in the future prediction time domain after the disturbance is injected. The matrix construction module is used to calculate the forced response coefficients at each step by comparing the current tracking error at each step after the injected perturbation with the free response under no perturbation. The free response passes through the free error trajectory. The forced response coefficient is obtained. Characterize the effect of applying a perturbation at the current moment on the future... The impact of current tracking error in the step; based on the forced response coefficient Construct a dynamic matrix describing the impact of voltage increments on future multi-step current tracking errors. and the free error trajectory and dynamic matrix Combine to establish the future total error trajectory The prediction equation; The optimization solution module is used to construct a trajectory containing the total future error. With future control incremental trajectory The objective function is defined, and the optimal future control increment sequence is obtained by minimizing the objective function. In order to obtain the optimal control quantity at the current moment. ; The switch selection module is used to select the switch state that is closest to the optimal control quantity from the limited set of switch states of the converter based on a preset value function, and apply it to the converter.

10. An electronic device, characterized in that, include: Memory, used to store program instructions; A processor, coupled to the memory, is configured to execute the program instructions to implement the online data-driven converter predictive control method as described in any one of claims 1 to 8.