Continuous set model-free predictive control method for permanent magnet motors based on Bayesian optimization

Abstract

The present application provides a continuous set model-free predictive control method for permanent magnet motors based on Bayesian optimization, which relates to the field of permanent magnet synchronous motor control. The method obtains a hyperlocal model and an extended state observer; determines bandwidth and input gain as optimization variables, obtains a training set and a test set for Bayesian optimization; calculates the joint probability distribution of the training set and the test set using a kernel function, and determines the predicted mean and predicted variance of the optimization variables; screens the optimization variable combination based on an expected value acquisition function, and updates the bandwidth and input gain based on the optimization variable combination; calculates the reference voltage for the next control cycle based on the updated extended state observer and hyperlocal model and the deadbeat control principle; decomposes the reference voltage into basic voltage vectors through space vector modulation, and executes the switching sequence of the basic voltage vectors according to the minimum number of switching times to drive the permanent magnet motor. This method is not affected by changes in motor parameters and can improve the steady-state performance of current predictive control.

Description

Technical Field

[0001] The present invention relates to the field of motor control, and in particular to a continuous set model-free predictive control method for a permanent magnet motor based on Bayesian optimization. Background Art

[0002] In wind power generation systems, permanent magnet synchronous motors (PMSMs) are the preferred energy conversion device due to their high power density and high efficiency, making them particularly suitable for offshore wind power generation systems operating in harsh environments. The proportional-integral control method used in related technologies suffers from slow dynamic response. To overcome this, finite set model predictive control (FPMC) has been proposed. However, FPMC suffers from large steady-state errors, strong parameter dependence, and weak robustness. Summary of the Invention

[0003] In order to address the deficiencies of the prior art, the purpose of this application is to provide a continuous set model-free predictive control method for permanent magnet motors based on Bayesian optimization, which is not affected by changes in motor parameters and can improve the steady-state performance and dynamic characteristics of current predictive control.

[0004] In a first aspect, the present application provides a continuous set model-free predictive control method for a permanent magnet motor based on Bayesian optimization, the method comprising:

[0005] Obtaining a hyperlocal model and a pre-established extended state observer; wherein the hyperlocal model is determined based on a mathematical model of a permanent magnet synchronous motor in a two-phase rotating coordinate system;

[0006] The bandwidth of the extended state observer and the input gain of the hyperlocal model are determined as optimization variables, and a training set and a test set are obtained for Bayesian optimization. The training set is obtained by sampling the historical bandwidth and historical input gain through a sliding window.

[0007] Calculate the joint probability distribution of the training set and the test set through the kernel function, and determine the predicted mean and predicted variance of the optimized variable;

[0008] Screening the optimization variable combination based on the expected value acquisition function, and updating the bandwidth of the extended state observer and the input gain of the hyperlocal model based on the screened optimization variable combination;

[0009] Based on the updated extended state observer and hyperlocal model, and using the deadbeat control principle, the reference voltage applied to the next control cycle is calculated;

[0010] The reference voltage is decomposed into basic voltage vectors through space vector modulation, and the inverter is controlled to execute the switching sequence of the basic voltage vectors according to the minimum switching number to drive the permanent magnet motor.

[0011] In one embodiment, determining a hyperlocal model based on a mathematical model of a permanent magnet synchronous motor in a two-phase rotating coordinate system includes:

[0012] A mathematical model of the permanent magnet synchronous motor in a two-phase rotating coordinate system is established; the mathematical model describes the dynamic characteristics of the current through voltage equations on the d-axis and q-axis. The mathematical model includes the following parameters: the d-axis and q-axis components of the permanent magnet synchronous motor stator resistance, the d-axis and q-axis components of the permanent magnet synchronous motor stator inductance, the mechanical speed of the permanent magnet synchronous motor rotor, and the permanent magnet flux linkage of the permanent magnet synchronous motor.

[0013] When there is an error between the stator resistance in the mathematical model and the actual resistance value and / or there is an error between the stator inductance and the actual inductance, the mathematical model is converted into a hyperlocal model through a first-order linear transformation.

[0014] In one embodiment, the hyperlocal model is the product of an input gain and an input voltage of the permanent magnet synchronous motor plus a lumped disturbance; wherein the input gain is affected by a stator inductance error, and the lumped disturbance includes at least one of the following: a stator resistance error, an external load variation, and unmodeled dynamics.

[0015] In one embodiment, establishing the extended state observer includes:

[0016] The current prediction error is defined as the difference between the actual current of the permanent magnet synchronous motor and the current estimated value of the permanent magnet synchronous motor;

[0017] An observer equation is designed. The observer equation includes the product of the input gain and the input voltage of the permanent magnet synchronous motor, the lumped disturbance estimation, and the feedback compensation based on the current prediction error. The response speed of the lumped disturbance estimation can be controlled by adjusting the size of the observer bandwidth.

[0018] In one embodiment, the method further comprises: discretizing the hyperlocal model and the extended state observer;

[0019] The observer equation after discretization is: the current estimate at the current moment is associated with the current estimate at the previous moment, the product of the permanent magnet synchronous motor input voltage and input gain at the previous moment, the lumped disturbance estimate at the previous moment, and feedback compensation based on the current prediction error at the previous moment;

[0020] The hyperlocal model after discretization is as follows: the current estimate at the current moment is associated with the current value at the previous moment, the lumped disturbance at the previous moment, and the product of the permanent magnet synchronous motor input voltage and input gain at the previous moment.

[0021] In one embodiment, screening the optimization variable combination based on the expected value acquisition function includes:

[0022] Calculate the expected performance improvement of all candidate optimization variable combinations, where the performance improvement is the difference between the predicted value and the current optimal value;

[0023] Select the candidate optimization variable combination with the largest expected improvement as the optimal solution for the next control cycle;

[0024] If the expected improvement of all candidate optimization variable combinations is negative, the current optimization variable combination remains unchanged.

[0025] In one embodiment, the reference voltage is calculated as follows:

[0026] ;

[0027] in, represents a constant parameter, is the lumped disturbance estimate of the extended state observer, T is the time constant, Indicates the control cycle The reference voltage value when Indicates the value of input gain at control period k+P, Indicates that during the control cycle The current reference value when Indicates that during the control cycle The current reference value when Indicates that during the control cycle The estimated value of the lumped disturbance at Indicates that during the control cycle Estimated current value at .

[0028] In one embodiment, a reference voltage is decomposed into basic voltage vectors by space vector modulation, and an inverter is controlled to execute a switching sequence of the basic voltage vectors according to a minimum number of switching times to drive a permanent magnet motor, including:

[0029] Converting the reference voltage from the rotating coordinate system to the stationary coordinate system;

[0030] determining a basic voltage vector based on a location of a target sector;

[0031] Calculate the action time of the basic voltage vector;

[0032] Based on the action time, the inverter is controlled to execute the switching sequence of the basic voltage vector in the order of the minimum switching times to drive the permanent magnet motor.

[0033] In a second aspect, the present application further provides a continuous set model-free predictive control device for a permanent magnet motor based on Bayesian optimization, the control device comprising:

[0034] An acquisition module is used to acquire a hyperlocal model and a pre-established extended state observer; wherein the hyperlocal model is determined based on a mathematical model of a permanent magnet synchronous motor in a two-phase rotating coordinate system;

[0035] An update module is used to determine the bandwidth of the extended state observer and the input gain of the hyperlocal model as optimization variables, and obtain a training set and a test set for Bayesian optimization. The training set is obtained by sampling the historical bandwidth and historical input gain through a sliding window;

[0036] The prediction module is used to calculate the joint probability distribution of the training set and the test set through the kernel function and determine the predicted mean and predicted variance of the optimized variable;

[0037] A screening module selects optimization variable combinations based on the expected value acquisition function and updates the bandwidth of the extended state observer and the input gain of the hyperlocal model based on the screened optimization variable combinations. The screening module also calculates the reference voltage to be applied to the next control cycle based on the updated extended state observer and hyperlocal model and the deadbeat control principle.

[0038] The control module is used to decompose the reference voltage into basic voltage vectors through space vector modulation, and control the inverter to execute the switching sequence of the basic voltage vectors according to the minimum switching number to drive the permanent magnet motor.

[0039] In a third aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the continuous set model-free predictive control method for a permanent magnet motor based on Bayesian optimization according to the first aspect.

[0040] The above-mentioned continuous set model-free current predictive control method for a permanent magnet synchronous motor (PMSM) determines the observer bandwidth and model input gain as optimization variables by acquiring a hyperlocal model and an extended state observer based on the PMSM mathematical model. The method then uses a sliding window to sample historical data to construct the training set required for Bayesian optimization. A kernel function is used to calculate the joint probability distribution of the data and predict the mean and variance of the optimization variables. An expected value acquisition function is used to select the optimal variable combination and update the observer and model parameters. Based on the updated model and observer, the reference voltage for the next cycle is calculated in conjunction with the deadbeat control principle. The reference voltage is then decomposed into basic voltage vectors using space vector modulation. The inverter is then controlled to execute the switching sequence of these basic voltage vectors based on the minimum number of switching cycles to drive the PMSM. This method replaces the traditional PMSM model with an adaptive hyperlocal model and extended state observer for current predictive control. This method is unaffected by variations in the motor's own parameters and improves the steady-state performance and dynamic characteristics of the current predictive control. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1is a flow chart of a continuous set model-free predictive control method for a permanent magnet motor based on Bayesian optimization in one embodiment;

[0042] Figure 2 A flowchart for determining a hyperlocal model in one embodiment;

[0043] Figure 3 Flowchart for establishing an extended state observer in one embodiment

[0044] Figure 4 A flowchart of screening optimization variable combinations based on an expected value acquisition function in one embodiment;

[0045] Figure 5 A flowchart of calculating a voltage vector in one embodiment;

[0046] Figure 6 Schematic diagram of online updating of a training set in one embodiment;

[0047] Figure 7 A framework diagram of a continuous set model-free current predictive control of a magnetic synchronous motor based on Bayesian optimization in one embodiment;

[0048] Figure 8 FIG. 1 is a diagram of an apparatus for continuous set model-free predictive control of a permanent magnet motor based on Bayesian optimization in one embodiment. DETAILED DESCRIPTION

[0049] In order to make the purpose, technical solutions and advantages of this application more clear, the present application is further described in detail below with reference to the accompanying drawings and examples. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application. Unless otherwise defined, the technical terms or scientific terms involved in this application should have the general meaning understood by people with ordinary skills in the technical field to which this application belongs.

[0050] In one embodiment, Figure 1 As shown, a continuous set model-free predictive control method for a permanent magnet motor based on Bayesian optimization is provided, the method comprising the following steps:

[0051] Step 101: Obtain a hyperlocal model and a pre-established extended state observer; wherein the hyperlocal model is determined based on a mathematical model of a permanent magnet synchronous motor in a two-phase rotating coordinate system;

[0052] A mathematical model of the permanent magnet synchronous motor in a two-phase rotating coordinate system, or dq coordinate system, is established. This model describes the dynamic behavior of the permanent magnet synchronous motor, including the relationships between parameters such as stator voltage, current, inductance, resistance, and permanent magnet flux linkage. This mathematical model of the permanent magnet synchronous motor in a two-phase rotating coordinate system helps us understand its characteristics under different operating conditions.

[0053] Furthermore, the hyperlocal model is a simplified model that extracts key dynamic characteristics from the mathematical model of the permanent magnet synchronous motor through mathematical transformations, such as first-order affine transformations, while ignoring complex nonlinear factors or higher-order dynamics. The hyperlocal model retains only the input-output relationship directly related to current control and introduces parameters such as input gain to represent the dynamic characteristics of the permanent magnet synchronous motor.

[0054] A pre-established extended state observer (ESO) can estimate the current state and external disturbances of a permanent magnet synchronous motor in real time. This ESO estimates the motor's internal state and external disturbances independently of the motor model parameters. By feeding the ESO's estimation results back into the hyperlocal model, real-time monitoring of the permanent magnet synchronous motor's state and dynamic compensation for disturbances are possible.

[0055] Step 102: Determine the bandwidth of the extended state observer and the input gain of the hyperlocal model as optimization variables, and obtain a training set and a test set for Bayesian optimization. The training set is obtained by sampling the historical bandwidth and historical input gain through a sliding window.

[0056] The bandwidth of the extended state observer and the input gain of the hyperlocal model are optimization variables. To optimize the bandwidth and input gain, a dynamically updated training set is required. This training set is dynamically updated using a sliding window technique. The sliding window length represents the control step size, ensuring that the latest state information of the permanent magnet synchronous motor is always included.

[0057] For example, assuming the sliding window length is 5 and the current control cycle is the 10th cycle, the training set contains the historical bandwidth and input gain from the 6th to the 10th control cycle. When the control cycle advances to the 11th cycle, the sliding window moves forward one position, and the training set is updated to the bandwidth and input gain from the 7th to the 11th control cycle.

[0058] It's important to define the upper and lower bounds of the test data set, as well as its step size, and to maintain a constant test set. Before Bayesian optimization, clearly define the test set's data range and sampling interval. This ensures that the test set remains constant throughout the optimization process, providing a stable benchmark for performance evaluation. The upper and lower bounds of the test set define the data range, while the step size determines the interval between data points. Maintaining a constant test set prevents bias in the evaluation due to changes in the test data during the optimization process.

[0059] Step 103: Calculate the joint probability distribution of the training set and the test set using the kernel function, and determine the predicted mean and predicted variance of the optimized variable;

[0060] Select an appropriate kernel function, such as the radial basis kernel function (RBF), which measures the similarity between data points in the training and test sets, forming a kernel matrix. Using the kernel function, the data points in the training set can be mapped to a high-dimensional space. In this high-dimensional space, the distance between the data points in the training set can reflect the degree of similarity between the data points.

[0061] In high-dimensional space, the joint probability distribution of the training and test sets can be calculated. This joint probability distribution describes the range of possible values for the test set data given the training set data. From this joint probability distribution, the predicted mean and predicted variance of the optimization variable can be further derived. The predicted mean represents the expected value of the test set data given the training set data, while the predicted variance represents the uncertainty or variability of the test set data.

[0062] Step 104: Screening an optimization variable combination based on the expected value acquisition function, and updating the bandwidth of the extended state observer and the input gain of the hyperlocal model based on the screened optimization variable combination;

[0063] Expected value acquisition function ( EI ) is an indicator that measures the degree of improvement that a candidate parameter combination may bring relative to the current optimal solution. The expected value acquisition function can comprehensively consider the prediction mean and prediction variance, focusing on both potential performance improvements and uncertainty.

[0064] During the optimization process, the kernel function is used to calculate the joint probability distribution of the training and test sets to derive the predicted mean and predicted variance of the optimization variables. Based on these predicted mean and predicted variance, an expected value acquisition function is constructed to evaluate the potential improvement of each candidate parameter combination. Furthermore, the expected value acquisition function calculates an expected value for each candidate point, which reflects the expected improvement in the optimization objective. By comparing the expected values, the optimization algorithm selects the parameter combination with the highest expected improvement.

[0065] The selected optimization variable combination is the optimal value for the bandwidth of the extended state observer and the input gain of the hyperlocal model. These factors can influence the dynamic response and robustness of the permanent magnet synchronous motor. Furthermore, the selected optimal parameter combination is applied to the extended state observer and the hyperlocal model to update the bandwidth of the extended state observer and the input gain of the hyperlocal model. This update process enables the permanent magnet synchronous motor to adaptively adjust its parameters to better adapt to the current operating state and external disturbances.

[0066] Step 105: Based on the updated extended state observer and the hyperlocal model, and using the deadbeat control principle, calculate the reference voltage applied to the next control cycle;

[0067] The updated extended state observer can estimate the permanent magnet synchronous motor's current state and external disturbances in real time. These estimates reflect the motor's actual operating conditions and external disturbances at the current moment. Based on the estimates provided by the extended state observer and the motor's dynamic characteristics, the hyperlocal model can predict the motor's reference voltage for the next control cycle.

[0068] It should be noted that the deadbeat control principle is a prediction-based control method. Its core is to use the optimized variable combination of the current cycle to accurately predict the permanent magnet synchronous motor state in the next control cycle, allowing the control command to quickly and accurately track the desired value, thereby achieving control without steady-state error. For example, if the current control cycle is the 10th control cycle, the optimized variable combination of the 10th control cycle can be used to predict the reference voltage for the 11th control cycle.

[0069] During this process, the hyperlocal model uses the current estimates and lumped disturbance values provided by the extended state observer, combined with the hyperlocal model, to calculate the reference voltage required for the next control cycle. This reference voltage calculation takes into account the motor's dynamic characteristics, current state, and desired current trajectory, ensuring that the permanent magnet synchronous motor current quickly and accurately tracks the desired current value during the next control cycle.

[0070] Step 106: Decompose the reference voltage into basic voltage vectors through space vector modulation, and control the inverter to execute the switching sequence of the basic voltage vectors according to the minimum switching number to drive the permanent magnet motor.

[0071] In a three-phase inverter, there are eight basic voltage vectors, which are formed by the different on and off states of the inverter's switches. Six of these basic voltage vectors are non-zero, and two are zero. The reference voltage is converted from a rotating coordinate system to a stationary coordinate system, and its components in the stationary coordinate system are calculated. The sector in which the reference voltage is located is determined based on the positions of these components in the stationary coordinate system. Each sector corresponds to two adjacent non-zero and zero basic voltage vectors. Based on the volt-second balance principle, the interaction time of these two adjacent non-zero and zero basic voltage vectors within a sampling period is calculated to ensure that the combined voltage within that period is equivalent to the reference voltage.

[0072] After determining two adjacent non-zero basic voltage vectors and a zero basic voltage vector, along with their action times, the switching sequence of the basic voltage vectors is arranged according to the principle of minimum switching times to reduce switching losses and electromagnetic interference in the inverter's switching devices. Specifically, in two adjacent switching cycles, the state of only one switch is changed. For example, when switching from one basic voltage vector to another, only one switch is changed from on to off, or from off to on. Furthermore, according to the arranged switching sequence and corresponding action times, the inverter is driven to sequentially output the basic voltage vectors. The voltage waveform output by the inverter will approximate that of the reference voltage, thereby generating a suitable three-phase AC voltage to drive the permanent magnet synchronous motor.

[0073] In this embodiment, the method obtains a hyperlocal model and an extended state observer based on the mathematical model of a permanent magnet synchronous motor, determines the observer bandwidth and model input gain as optimization variables, and uses a sliding window to sample historical data to construct the training set required for Bayesian optimization. A kernel function is used to calculate the joint probability distribution of the data and predict the mean and variance of the optimization variables. An expected value acquisition function is used to select the optimal variable combination and update the observer and model parameters. Based on the updated model and observer, the reference voltage for the next cycle is calculated in conjunction with the deadbeat control principle. The reference voltage is then decomposed into basic voltage vectors through space vector modulation. The inverter is then controlled to execute the switching sequence of these basic voltage vectors based on the minimum number of switching cycles to drive the permanent magnet motor. This method uses an adaptively changing hyperlocal model and extended state observer to replace the traditional permanent magnet motor model for current predictive control. This method is unaffected by changes in the motor's own parameters and improves the steady-state performance and dynamic characteristics of current predictive control.

[0074] In one embodiment, Figure 2 As shown, determining the super-local model based on the mathematical model of the permanent magnet synchronous motor in the two-phase rotating coordinate system includes the following steps:

[0075] Step 201: Establishing a mathematical model of a permanent magnet synchronous motor in a two-phase rotating coordinate system; wherein the mathematical model describes the dynamic characteristics of the current through voltage equations on the d-axis and q-axis, and the mathematical model includes the following parameters: the d-axis and q-axis components of the permanent magnet synchronous motor stator resistance, the d-axis and q-axis components of the permanent magnet synchronous motor stator inductance, the mechanical speed of the permanent magnet synchronous motor rotor, and the permanent magnet flux linkage of the permanent magnet synchronous motor;

[0076] The operating principle of a permanent magnet synchronous motor (PMSM) is based on the law of electromagnetic induction and the fundamental laws of electrical circuits. These laws describe the electromagnetic phenomena and circuit behavior within a PMSM, enabling mathematical equations to describe its operating characteristics. In a two-phase rotating coordinate system (dq coordinate system), the d-axis, also known as the direct axis, aligns with the permanent magnet flux linkage of the PMSM rotor and represents the direction of the PMSM's magnetic flux. The q-axis, also known as the quadrature axis, is perpendicular to the d-axis and represents the direction of torque generation in the PMSM. In the dq coordinate system, physical quantities such as voltage, current, and flux linkage of the PMSM can be decomposed into components along the d- and q-axes. By controlling the currents along the d- and q-axes separately, the PMSM's magnetic flux and torque can be independently controlled, achieving precise control of the PMSM.

[0077] By applying the law of electromagnetic induction and the basic laws of circuits, the voltage equations on the d-axis and q-axis of the permanent magnet synchronous motor can be derived: .

[0078] The d-axis voltage equation describes the rate of change of the d-axis current and is primarily affected by the d-axis resistance, d-axis inductance, and q-axis current. The q-axis voltage equation describes the rate of change of the q-axis current and is primarily affected by the q-axis resistance, q-axis inductance, d-axis current, and permanent magnet flux linkage.

[0079] in, and They represent the d-axis and q-axis components of the stator output voltage of the permanent magnet synchronous motor respectively; Indicates the stator resistance of the permanent magnet synchronous motor; and They represent the d-axis and q-axis components of the permanent magnet synchronous motor stator current respectively; and They represent the d-axis and q-axis components of the permanent magnet synchronous motor stator inductance respectively; Indicates the mechanical speed of the permanent magnet synchronous motor rotor; Represents the permanent magnet flux of permanent magnet synchronous motor.

[0080] Step 202: When there is an error between the stator resistance in the mathematical model and the actual resistance value and / or the stator inductance and the actual inductance, the mathematical model is converted into a hyperlocal model through a first-order linear transformation. The hyperlocal model is the product of the input gain and the input voltage of the permanent magnet synchronous motor plus a lumped disturbance. The input gain is affected by the stator inductance error, and the lumped disturbance includes at least one of the following: stator resistance error, external load variation, and unmodeled dynamics.

[0081] When the stator inductance and resistance parameters of the permanent magnet motor do not match the actual values, the mathematical model of the voltage on the d-axis and q-axis of the permanent magnet synchronous motor is:

[0082] .

[0083] in, and They represent the parameter errors of stator resistance and stator inductance respectively.

[0084] Furthermore, by using the first-order affine transformation, that is, rearranging the mathematical model, the time derivative of the current is obtained as:

[0085] .

[0086] Simplify the coefficients in the above equation and introduce a new input gain α d ,α q , and the lumped disturbance F d , F q .in, , , = , = Furthermore, by using the first-order affine transformation and simplifying the coefficients, the obtained hyperlocal model is:

[0087] ;

[0088] in, and represents the model input gain that changes due to the inductance parameter disturbance; and represents the lumped disturbance.

[0089] In one embodiment, Figure 3 As shown in Figure 2, the establishment of the extended state observer includes the following steps:

[0090] Step 301: define the current prediction error as the difference between the actual current of the permanent magnet synchronous motor and the current estimated value of the permanent magnet synchronous motor;

[0091] ;

[0092] in, is the estimated current value of the permanent magnet synchronous motor, is the actual current of the permanent magnet synchronous motor, is the current prediction error of the extended state observer. Further, =[ , ] T , is the d-axis current error. is the q-axis current error. T is the transpose operator, used to transpose the matrix. =[ , ] T express is and A column vector consisting of .

[0093] Step 302: Design an observer equation; wherein the observer equation includes the product of the input gain and the permanent magnet synchronous motor input voltage, the lumped disturbance estimation, and the feedback compensation based on the current prediction error, and the response speed of the lumped disturbance estimation can be controlled by adjusting the size of the observer bandwidth.

[0094] The design of the extended state observer equation based on lumped disturbance is as follows:

[0095] .

[0096] in, is the voltage reference value, is the current prediction error of the extended state observer; is the estimated value of the current; is the estimated value of the lumped disturbance; and are the feedback gains of the observer respectively; represents the bandwidth of the extended state observer, is the input gain.

[0097] In this embodiment, through this design, the observer can estimate the external disturbance in real time and perform feedback compensation, thereby improving the dynamic performance of current control.

[0098] In one embodiment, the method further includes: discretizing the hyperlocal model and the extended state observer; wherein the observer equation after the discretization processing is: the current estimate value at the current moment is associated with the current estimate value at the previous moment, the product of the permanent magnet synchronous motor input voltage and input gain at the previous moment, the lumped disturbance estimate at the previous moment, and feedback compensation based on the current prediction error at the previous moment; and the hyperlocal model after the discretization processing is: the current estimate value at the current moment is associated with the current value at the previous moment, the lumped disturbance at the previous moment, and the product of the permanent magnet synchronous motor input voltage and input gain at the previous moment.

[0099] After discretizing the extended observer equation, we can obtain:

[0100] .

[0101] in, is the current prediction error at the current moment k is the estimated current value at the current moment k, is the estimated current value at time k+1. is the actual current at the current moment k. is the lumped disturbance estimate at the current time k, is the lumped disturbance estimate at time k+1. is the input gain at the current moment k. is the permanent magnet synchronous motor voltage reference value at the current time k. is the bandwidth of the extended state observer at the current time k, is a single control cycle.

[0102] After discretizing the first-order hyperlocal model, we get:

[0103] .

[0104] in, is the actual current at the current moment k, is the actual current at time k+1. is the input gain at the current moment k, is the permanent magnet synchronous motor voltage reference value at the current time k. is the aggregate disturbance at the current time k, is a single control cycle.

[0105] It should be noted that the discretized observer equation and hyperlocal model both consider the current estimate at the previous moment, maintaining the continuity and stability of the current estimate. The product of the input gain and the permanent magnet synchronous motor input voltage at the previous moment reflects the impact of the motor input voltage on the current and is used to predict the current change at the current moment. The lumped disturbance estimate can reflect various disturbance factors, such as stator resistance error, external load changes, and unmodeled dynamics. By considering the lumped disturbance estimate at the previous moment, the accuracy of the current estimate can be improved. The current prediction error is used for feedback compensation to adjust the observer output and improve the estimation accuracy. Feedback compensation can reduce the current estimation error.

[0106] In one embodiment, the method for online updating the test set and training set for Bayesian optimization is to determine the upper and lower limits of the test set and their step size, while keeping the test set constant. Each time the control sequence is shifted backward, the data in the training set is shifted forward one position. The data initially indexed as 1 is discarded and replaced by the data indexed as 2. This process continues until the number of data shifts equals the control step size. The optimization parameters collected during the shift are then stored, equal to the number of control step sizes, in the corresponding data positions in the training set. The expression is as follows:

[0107] .

[0108] in, =[ , ] represents the training set, whose size is t ×3; Represents the acquisition data storage matrix, the size is t ×3; P To control the step size.

[0109] Furthermore, the radial basis kernel function (RBF) is used to calculate the similarity between each data point in the updated training set and the test set. The kernel function formula is:

[0110] .

[0111] in and Represent two correlation vectors, is a constant parameter. The higher the similarity, the closer the kernel function value is to 1; the lower the similarity, the closer the kernel function value is to 0. It should be noted that and It can be represented as a training set and a data set. Using the kernel function, the similarity between the training set and the data set can be calculated, providing a basis for the subsequent construction of the joint probability.

[0112] The control objective of the hyperlocal model is to minimize the current tracking error. The randomly selected training set data can be considered to form a high-dimensional normal distribution with the tracking error:

[0113] .

[0114] in, Represents the control target corresponding to the training set data, the inverse of the steady-state current tracking cumulative error; and Respectively represent the predicted mean and predicted variance corresponding to the training set data; The number of cumulative error integration cycles set; is the reference current value; is the actual current value. Indicates input gain and observer bandwidth The function follows the mean , the covariance matrix is Gaussian distribution, reflected in different and The predicted distribution of performance under the combination. is the input gain , which represents the predicted value of performance under different input gains. is the observer bandwidth ω 0, which represents the predicted value of performance under different observer bandwidths. Indicates the current control cycle. Indicates the step size. Indicates the offset of the time window. Indicates the control cycle arrive Input gain The value of . Indicates the control cycle arrive The observer bandwidth The value of .

[0115] It should be noted that it is assumed that there is a high-dimensional normal distribution relationship between the training set data and the control target, which means that under different control cycles, the input gain and observer bandwidth Combination and corresponding control objectives f There are certain statistical laws between them. and the prediction variance It can describe the central tendency and discrete degree of the training set data, providing a basis for subsequent Bayesian optimization. r The integration period of the accumulated error can be controlled, which can affect the calculation of the control target.

[0116] Furthermore, as the control time increases, the steady-state error under fixed control instructions will gradually decrease and approach 0, so the high-dimensional normal distribution of the training set can be approximated as a zero-mean normal distribution. Similarly, the test set data also satisfies the zero-mean normal distribution.

[0117] Furthermore, the kernel function can be used to construct the joint probability distribution of the training set and the test set. The joint probability distribution can describe the possible value range of the test set data given the training set data. Among them, the training set and the test set are constructed according to the operating rules of the normal distribution, and their joint probability distribution also conforms to the rules of the normal distribution. According to the definition of the normal joint distribution, the joint probability distribution of the training set and the test set can be obtained. The mean of this joint probability distribution is 0 (the target expectation of optimization); the variance is a matrix composed of the kernel function K. The mathematical expression of the joint probability distribution is:

[0118] .

[0119] in, Represents the control target corresponding to the test set data. Represents the training dataset X The kernel function matrix between itself is used to measure the similarity between training set data points. Represents the training dataset With the test dataset The kernel function matrix between is used to measure the similarity between the training set data points and the test set data points. Represents the test dataset The kernel matrix between itself is used to measure the similarity between the test set data points.

[0120] Furthermore, the joint probability distribution of the training set and the test set is calculated using the standard zero-mean normal distribution formula to obtain the target optimization parameters, namely the prior Gaussian distribution of the input gain and bandwidth:

[0121] .

[0122] in represents the predicted mean of the prior Gaussian distribution, is the standard deviation of the prior Gaussian distribution.

[0123] Furthermore, in order to speed up the iteration of the above-mentioned prior Gaussian distribution, the expected acquisition function (EI) can be used as the optimal function for judging the next control cycle. and Criteria for determining the value:

[0124] .

[0125] in, Indicates that in all possible X In the example, find the objective function (such as the expected value acquisition function or performance indicators Reaching the maximum value X . , Represents the training set data set; Represents the current moment training set objective function vector The maximum value in ; Represents the data added to the training dataset after the next step starts when the optimization iteration loop is not terminated; It represents the optimal parameter combination obtained by Bayesian optimization under the current number of iterations. Calculate the optimal parameter combination of all candidate optimization variables. value, and select The combination with the largest value is taken as the optimal solution for the next control cycle. Furthermore, the bandwidth of the extended state observer and the input gain of the hyperlocal model are updated based on the selected optimization variable combinations.

[0126] It should be noted that the expected value acquisition function screens the optimization variable combination, and the process based on the screened optimization variable combination is:

[0127] Filter out relevant objective functions within the training set The maximum value in and its corresponding optimal parameter set ;Will and Bring in The calculation is carried out in the calculation formula of the function. Further, we can introduce Transform the training set and simplify the formula to The calculation formula of the normal distribution. The calculated result is the test set The optimization goal for each data point The expected value of , this process requires the participation of the results of the previously calculated joint Gaussian distribution. The calculated data vector takes the maximum value, and the test set corresponding to the maximum value The data results in the result are the filtered optimization results. The specific calculation formula of the expected acquisition function EI function is as follows:

[0128] .

[0129] in, is the standard normal probability density function; The distribution function of the standard normal distribution is represented. The optimal combination of optimization variables is applied to the extended state observer and the hyperlocal model to update their bandwidth and input gain.

[0130] In one embodiment, Figure 4 As shown in FIG, screening the optimization variable combination based on the expected value acquisition function includes the following steps:

[0131] Step 401: Calculate the expected performance improvement of all candidate optimization variable combinations, where the performance improvement is the difference between the predicted value and the current optimal value;

[0132] For each candidate combination, the expected value acquisition function is used to predict its performance value; the predicted value is compared with the currently known optimal performance value, and the difference between each candidate group is calculated. The difference is the expected performance improvement of the candidate combination.

[0133] Step 402: Select the candidate optimization variable combination with the largest expected improvement as the optimal solution for the next control cycle;

[0134] The candidate optimization variable combination with the greatest expected improvement is selected as the optimal solution for the next control cycle. This means that at the end of each control cycle, from all possible optimization variable combinations, the combination most likely to achieve the greatest performance improvement under the predicted conditions is selected. This combination can be applied to the next control cycle, hoping to achieve a significant improvement in control performance.

[0135] Step 403: If the expected improvement of all candidate optimization variable combinations is negative, the current optimization variable combination is maintained unchanged.

[0136] If the calculated expected performance improvement for all candidate combinations is negative, it means that none of the candidate combinations will improve performance, and may even result in performance degradation. To ensure system stability and control performance, you can choose to maintain the current optimization variable combination and continue to use the current parameters for control until a new candidate combination that can improve performance is found during subsequent optimization.

[0137] In this embodiment, the method can effectively improve optimization efficiency and effect, ensure the stability and optimality of performance, and avoid performance degradation caused by blind attempts.

[0138] In one embodiment, the reference voltage is calculated as follows:

[0139] ;

[0140] in, is the voltage reference value, is the lumped disturbance estimate of the extended state observer, represents the control period, represents the step length, is the time constant used to adjust the integral part of the control input.

[0141] Indicates that in the control cycle The reference voltage value when Indicates the input gain during the control cycle The value of Indicates that during the control cycle The current reference value when Indicates that during the control cycle The current reference value when Indicates that during the control cycle The estimated value of the lumped disturbance at Indicates that during the control cycle Estimated current value at .

[0142] Calculate the next control cycle Current reference value With the current control cycle Estimated current Using proportional control gain The current error is adjusted to quickly respond to current changes. Subtract the current cycle The perturbation estimate of , to compensate for the disturbance, by dividing the control period , calculate the rate of change of the current error. Divide the above terms by the current cycle Input gain , get the final reference voltage value .

[0143] In this embodiment, the formula can accurately calculate the reference voltage value required by the permanent magnet synchronous motor by combining proportional control, disturbance compensation and current error change rate, so as to achieve precise control of the current.

[0144] In one embodiment, Figure 5 As shown, the reference voltage is decomposed into the action time of the basic voltage vector through space vector modulation, and the inverter output is controlled according to the principle of minimum switching times, including the following steps:

[0145] Step 501: Convert the reference voltage from the rotating coordinate system to the stationary coordinate system;

[0146] Reference voltage In the rotating coordinate system (dq coordinate system), it is expressed as and , can be converted into the stationary coordinate system (α-β coordinate system) by inverse Clark transformation and The formula for the inverse Clark transform is as follows:

[0147] .

[0148] in, is the rotor position angle, which indicates the current position of the permanent magnet synchronous motor rotor.

[0149] Step 502: Determine a basic voltage vector based on the location of the target sector;

[0150] Use the arctan function to calculate the angle of the reference voltage in the stationary coordinate system (α-β coordinate system) φ , the formula is:

[0151] ;

[0152] because The range of the function is (−2 π , 2 π ), in order to obtain the complete [0,2 π ) range, you need to and The specific rules are as follows:

[0153] when >0 and ≥0, ;

[0154] when <0, ;

[0155] when >0 and <0, .

[0156] According to the angle of the reference voltage in the stationary coordinate system The range determines the sector where the reference voltage is located. The coordinate system is divided into six sectors, each sector corresponds to a certain angle range, and each sector corresponds to two specific adjacent non-zero basic voltage vectors and a zero basic voltage vector. The basic voltage vectors include V 0(000), V 1(100), V 2(110), V 3(010), V 4(011), V 5(001), V 6(101) and V 7(111), V 0 andV 7 is a zero basic voltage vector, and the remaining six basic voltage vectors are non-zero basic voltage vectors. The entire voltage space is divided into six sectors by the six non-zero basic voltage vectors. Based on the sector where the reference voltage is located, two adjacent non-zero basic voltage vectors and the zero basic voltage vector in the sector are determined as the required basic voltage vectors.

[0157] Step 503: Calculate the action time of the basic voltage vector;

[0158] For example, u s Take the first sector as an example:

[0159] .

[0160] in, is the amplitude of the reference voltage vector in the two-phase stationary coordinate system; is the DC side voltage of the inverter. and are two adjacent non-zero basic voltage vectors.

[0161] Furthermore, the action time of each basic voltage vector in the next control cycle is:

[0162] .

[0163] Among them, the adjustment system Indicates the reference voltage amplitude and DC side voltage The ratio is used to control the output voltage. The value range is usually between 0 and 1, indicating the ratio of the output voltage to the DC voltage. represents the action time of the non-zero basic voltage vector V1, Indicates the action time of the non-zero basic voltage vector V2. Indicates the action time of the zero basic voltage vector V0 or V7.

[0164] It should be noted that according to the volt-second balance principle, in one sampling period T s The volt-second product of the reference voltage vector is equal to the sum of the volt-second products of each basic voltage vector, that is:

[0165] VT s = V x T x + V y T y +V 0 T 0.

[0166] in, and are two adjacent non-zero basic voltage vectors, and They are the action times corresponding to two adjacent non-zero basic voltage vectors, where the two adjacent non-zero basic voltage vectors can be V 1 and V 2; V 0 is the zero vector ( V 0 or V 7), is the action time of the zero basic voltage vector and satisfies + + = .

[0167] Step 504: Based on the action time, the inverter is controlled to execute the switching sequence of the basic voltage vector in the arrangement sequence of the minimum switching times to drive the permanent magnet motor.

[0168] To reduce losses in the inverter's switching devices, the switching order of the basic voltage vectors is arranged based on the principle of minimizing the number of switching cycles. When switching between two adjacent basic voltage vectors, only one switch state is changed. For example, when switching from one basic voltage vector to the next, only one switch state is changed from on to off, or vice versa.

[0169] The inverter outputs the basic voltage vectors in sequence according to the programmed minimum switching times and corresponding action times. The voltage waveform output by the inverter will approach the waveform of the reference voltage vector, thereby generating a suitable three-phase AC voltage.

[0170] For example, if the reference voltage vector is located in sector I, the corresponding non-zero basic voltage vector is V 1 and V 2. The zero basic voltage vector is V 0 and V 7. Calculated V 1. V 2. V 0 and V After the action time of 7, the switching sequence is determined according to the principle of minimum switching times: V 0- V 1- V 2- V 7- V 7- V 1- V 0. The inverter outputs firstV 0 corresponding action time, then output V 1. V 2. V 7 and other corresponding action times, and this process is continuously repeated to drive the permanent magnet synchronous motor to operate.

[0171] In one embodiment, the test set and training set structure diagram of online Bayesian optimization are as follows: Figure 6 As shown in the figure, the training set (storing historical input gain and bandwidth), the test set (fixed value range for verification) and the data acquisition and storage module (real-time recording of operating data such as current and speed) realize data management.

[0172] The sliding window mechanism is used to update the training set data, with the current control moment on the time axis As a benchmark, set the control step size To optimize the interval, the training set window size is (Storage history data length), the prediction period is ,in The data at that moment will be discarded. arrive The data at the moment is used as training data, and the latest k-time data is added to the storage, and arrive Status prediction for the time period.

[0173] The data flow follows the order of discarded data (old data outside the window) → training data (valid historical data) → measured data (latest real-time data) → predicted data (future state prediction). Through the computational iteration loop, the closed-loop process of "model update → test verification → state prediction → parameter optimization → control implementation → new data acquisition" is continuously executed. This incremental update mechanism ensures data timeliness while improving computational efficiency, and can dynamically track parameter changes. The hyperlocal model provides a dynamic, simplified representation, while the extended state observer estimates unmeasurable states and disturbances in real time. The two work together to ensure excellent control performance and robustness under various operating conditions.

[0174] In one embodiment, the framework diagram of the continuous set model-free current predictive control of the magnetic synchronous motor based on Bayesian optimization is as follows: Figure 7 As shown in the figure, it is used for current control of permanent magnet synchronous motor (PMSM). It includes modules such as hyperlocal model, EI acquisition function, Gaussian process model, SVPWM inverter, state observation and PI controller.

[0175] The current measurement module collects the phase current of the permanent magnet synchronous motor 、 、 , the encoder measures the speed of the permanent magnet synchronous motor and rotor position θ . Convert the three-phase current 、 、 for Current in the coordinate system , from αβ to Coordinate transformation requires the voltage in the αβ coordinate system to be Convert to Voltage in the coordinate system . Extended state observer lumped disturbance estimate and current estimates The hyperlocal model receives the optimization variables and , and output Shaft voltage component The Gaussian process model makes predictions based on historical data and current data, and outputs the predicted mean. and the predicted standard deviation The EI acquisition function receives the predicted mean and the predicted standard deviation , calculate the expected performance improvement of each candidate optimization variable combination and select the best combination. The inverter converts the reference voltage into a precise PWM signal through the SVPWM control module, controls the switching state of the inverter, and thus adjusts the power supply voltage and frequency of the motor. Control the inverter output. The PI controller controls the speed of the permanent magnet synchronous motor. and speed reference Adjusting the current reference value , ensuring the accuracy of current tracking.

[0176] Based on the same inventive concept, embodiments of the present application also provide a device for model-free predictive control of a continuous set of permanent magnet motors based on Bayesian optimization. The implementation solution provided by this device is similar to the implementation solution described in the aforementioned method. Therefore, the specific limitations of one or more embodiments of model-free predictive control of a continuous set of permanent magnet motors based on Bayesian optimization provided below can be found in the limitations of the model-free predictive control method of a continuous set of permanent magnet motors based on Bayesian optimization described above, and will not be repeated here.

[0177] In one embodiment, Figure 8 As shown, the embodiment of the present application also provides a continuous set model-free predictive control device for a permanent magnet motor based on Bayesian optimization, the control device comprising:

[0178] An acquisition module 801 is configured to acquire a hyperlocal model and a pre-established extended state observer; wherein the hyperlocal model is determined based on a mathematical model of a permanent magnet synchronous motor in a two-phase rotating coordinate system;

[0179] Update module 802 uses the bandwidth of the extended state observer and the input gain of the hyperlocal model as optimization variables to obtain a training set and a test set for Bayesian optimization. The training set is obtained by sampling the historical bandwidth and historical input gain through a sliding window.

[0180] Prediction module 803, used to calculate the joint probability distribution of the training set and the test set through the kernel function, and determine the predicted mean and predicted variance of the optimized variable;

[0181] A screening module 804 screens optimization variable combinations based on the expected value acquisition function and updates the bandwidth of the extended state observer and the input gain of the hyperlocal model based on the screened optimization variable combinations. The screening module also calculates a reference voltage for the next control cycle based on the updated extended state observer and hyperlocal model and using the deadbeat control principle.

[0182] The control module 805 is configured to decompose the reference voltage into basic voltage vectors through space vector modulation, and control the inverter to execute a switching sequence of the basic voltage vectors according to a minimum switching number to drive the permanent magnet motor.

[0183] In one embodiment, the acquisition module 801 determines a hyperlocal model based on a mathematical model of the permanent magnet synchronous motor in a two-phase rotating coordinate system, and is specifically used to: establish a mathematical model of the permanent magnet synchronous motor in a two-phase rotating coordinate system; wherein the mathematical model describes the dynamic characteristics of the current through voltage equations of the d-axis and q-axis, and the mathematical model includes the following parameters: the d-axis and q-axis components of the stator resistance of the permanent magnet synchronous motor, the d-axis and q-axis components of the stator inductance of the permanent magnet synchronous motor, the mechanical speed of the rotor of the permanent magnet synchronous motor, and the permanent magnet flux linkage of the permanent magnet synchronous motor; when there is an error between the stator resistance in the mathematical model and the actual resistance value and / or there is an error between the stator inductance and the actual inductance, the mathematical model is converted into a hyperlocal model through a first-order linear transformation.

[0184] In one embodiment, the acquisition module 801 models the hyperlocal model as the product of an input gain and an input voltage of the permanent magnet synchronous motor plus a lumped disturbance; wherein the input gain is affected by a stator inductance error, and the lumped disturbance includes at least one of the following: a stator resistance error, an external load change, and unmodeled dynamics.

[0185] In one embodiment, the acquisition module 801 defines the current prediction error as the difference between the actual current of the permanent magnet synchronous motor and the current estimate of the permanent magnet synchronous motor; and designs an observer equation; wherein the observer equation includes the product of the input gain and the input voltage of the permanent magnet synchronous motor, the lumped disturbance estimation, and the feedback compensation based on the current prediction error, and the response speed of the lumped disturbance estimation can be controlled by adjusting the size of the observer bandwidth.

[0186] In one embodiment, the acquisition module 801 discretizes the hyperlocal model and the extended state observer; wherein the observer equation after discretization is: the current estimate at the current moment is associated with the current estimate at the previous moment, the product of the permanent magnet synchronous motor input voltage and input gain at the previous moment, the lumped disturbance estimate at the previous moment, and feedback compensation based on the current prediction error at the previous moment; the hyperlocal model after discretization is: the current estimate at the current moment is associated with the current value at the previous moment, the lumped disturbance at the previous moment, and the product of the permanent magnet synchronous motor input voltage and input gain at the previous moment.

[0187] In one embodiment, the screening module 804 screens the optimization variable combinations based on the expected value acquisition function, specifically for: calculating the expected performance improvement of all candidate optimization variable combinations, where the performance improvement is the difference between the predicted value and the current optimal value; selecting the candidate optimization variable combination with the largest expected improvement as the optimal solution for the next control cycle; if the expected improvement of all candidate optimization variable combinations is negative, maintaining the current optimization variable combination unchanged.

[0188] In one embodiment, the calculation formula for the reference voltage calculated by the screening module 804 is:

[0189] ;

[0190] in, represents a constant parameter, is the lumped disturbance estimate of the extended state observer, T is the time constant, Indicates that during the control cycle The reference voltage value when Indicates the input gain during the control cycle The value of Indicates that during the control cycle The current reference value when Indicates that during the control cycle The current reference value when Indicates that during the control cycle The estimated value of the lumped disturbance at Indicates that during the control cycle Estimated current value at .

[0191] In one embodiment, the control module 805 decomposes the reference voltage into basic voltage vectors through space vector modulation, and controls the inverter to execute the switching sequence of the basic voltage vectors according to the minimum number of switching times to drive the permanent magnet motor. Specifically, it is used to: convert the reference voltage from the rotating coordinate system to the stationary coordinate system; determine the basic voltage vector based on the position of the target sector; calculate the action time of the basic voltage vector; and control the inverter to execute the switching sequence of the basic voltage vectors according to the arrangement order of the minimum number of switching times based on the action time to drive the permanent magnet motor.

[0192] Based on the same concept, the present application also provides a computer-readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to provide a continuous set model-free predictive control method for a permanent magnet motor based on Bayesian optimization.

[0193] The technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0194] The above embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present application. It should be noted that a person of ordinary skill in the art may make various modifications and improvements without departing from the spirit of the present application, and these modifications and improvements fall within the scope of protection of the present application. Therefore, the scope of protection of the present application shall be determined by the appended claims.

Claims

1. A continuous set model-free predictive control method for permanent magnet motors based on Bayesian optimization, characterized in that: The method comprises: Obtaining a hyperlocal model and a pre-established extended state observer; wherein the hyperlocal model is determined based on a mathematical model of a permanent magnet synchronous motor in a two-phase rotating coordinate system; Determining the bandwidth of the extended state observer and the input gain of the hyperlocal model as optimization variables, and obtaining a training set and a test set for Bayesian optimization, wherein the training set is obtained by sampling historical bandwidth and historical input gain through a sliding window; Calculating the joint probability distribution of the training set and the test set by a kernel function, and determining the predicted mean and predicted variance of the optimized variable; Screening an optimization variable combination based on an expected value acquisition function, and updating a bandwidth of the extended state observer and an input gain of the hyperlocal model based on the screened optimization variable combination; Based on the updated extended state observer and the hyperlocal model, and using the deadbeat control principle, calculating a reference voltage to be applied to the next control cycle; The reference voltage is decomposed into basic voltage vectors through space vector modulation, and the inverter is controlled to execute the switching sequence of the basic voltage vectors according to the minimum switching number to drive the permanent magnet motor.

2. The method for continuous set model-free predictive control of a permanent magnet motor based on Bayesian optimization according to claim 1, characterized in that: Determining the super-local model based on a mathematical model of the permanent magnet synchronous motor in a two-phase rotating coordinate system includes: Establishing a mathematical model of the permanent magnet synchronous motor in a two-phase rotating coordinate system; wherein the mathematical model describes the dynamic characteristics of the current through the voltage equations of the d-axis and q-axis, and the mathematical model includes the following parameters: the d-axis and q-axis components of the stator resistance of the permanent magnet synchronous motor, the d-axis and q-axis components of the stator inductance of the permanent magnet synchronous motor, the mechanical speed of the rotor of the permanent magnet synchronous motor, and the permanent magnet flux linkage of the permanent magnet synchronous motor; When there is an error between the stator resistance in the mathematical model and the actual resistance value and / or there is an error between the stator inductance and the actual inductance, the mathematical model is converted into the hyperlocal model through a first-order linear transformation.

3. The method for continuous set model-free predictive control of a permanent magnet motor based on Bayesian optimization according to claim 2, characterized in that: The hyperlocal model is the product of the input gain and the permanent magnet synchronous motor input voltage plus a lumped disturbance; wherein the input gain is affected by the stator inductance error, and the lumped disturbance includes at least one of the following: stator resistance error, external load change, and unmodeled dynamics.

4. The method for continuous set model-free predictive control of a permanent magnet motor based on Bayesian optimization according to claim 1, characterized in that: The establishment of the extended state observer includes: The current prediction error is defined as the difference between the actual current of the permanent magnet synchronous motor and the current estimation value of the permanent magnet synchronous motor; An observer equation is designed; wherein the observer equation includes the product of the input gain and the input voltage of the permanent magnet synchronous motor, a lumped disturbance estimate, and feedback compensation based on the current prediction error, and the response speed of the lumped disturbance estimate can be controlled by adjusting the size of the observer bandwidth.

5. The method for continuous set model-free predictive control of a permanent magnet motor based on Bayesian optimization according to any one of claims 1 to 4, characterized in that: The method further comprises: discretizing the hyperlocal model and the extended state observer; The observer equation after discretization is: the current estimate at the current moment is associated with the current estimate at the previous moment, the product of the permanent magnet synchronous motor input voltage and input gain at the previous moment, the lumped disturbance estimate at the previous moment, and feedback compensation based on the current prediction error at the previous moment; The discretized hyperlocal model is as follows: the current estimation value at the current moment is associated with the current value at the previous moment, the lumped disturbance at the previous moment, and the product of the permanent magnet synchronous motor input voltage and input gain at the previous moment.

6. The method for continuous set model-free predictive control of a permanent magnet motor based on Bayesian optimization according to claim 1, characterized in that: Screening optimization variable combinations based on expected value acquisition functions, including: Calculate the expected performance improvement of all candidate optimization variable combinations, where the performance improvement is the difference between the predicted value and the current optimal value; Select the candidate optimization variable combination with the largest expected improvement as the optimal solution for the next control cycle; If the expected improvement of all candidate optimization variable combinations is negative, the current optimization variable combination remains unchanged.

7. The method for continuous set model-free predictive control of a permanent magnet motor based on Bayesian optimization according to claim 1, characterized in that: The reference voltage is calculated as follows: ； in, represents a constant parameter, is the lumped disturbance estimate of the extended state observer, T is the time constant, Indicates that during the control cycle The reference voltage value when Indicates the input gain during the control cycle The value of Indicates that during the control cycle The current reference value when Indicates that during the control cycle The current reference value when Indicates that during the control cycle The estimated value of the lumped disturbance at Indicates that during the control cycle Estimated current value at .

8. The method for continuous set model-free predictive control of a permanent magnet motor based on Bayesian optimization according to claim 1, characterized in that: Decomposing the reference voltage into basic voltage vectors through space vector modulation, and controlling the inverter to execute the switching sequence of the basic voltage vectors according to the minimum number of switching times to drive the permanent magnet motor, comprising: converting the reference voltage from a rotating coordinate system to a stationary coordinate system; determining a target sector based on the voltage in the stationary coordinate system; determining a basic voltage vector based on a location of the target sector; Calculating the action time of the basic voltage vector; Based on the action time, the inverter is controlled to execute the switching sequence of the basic voltage vector in an arrangement sequence with a minimum number of switching times to drive the permanent magnet motor.

9. A continuous set model-free predictive control device for a permanent magnet motor based on Bayesian optimization, characterized in that: The control device comprises: An acquisition module, configured to acquire a hyperlocal model and a pre-established extended state observer; wherein the hyperlocal model is determined based on a mathematical model of a permanent magnet synchronous motor in a two-phase rotating coordinate system; an updating module, configured to determine the bandwidth of the extended state observer and the input gain of the hyperlocal model as optimization variables, and obtain a training set and a test set for Bayesian optimization, wherein the training set and the test set are both obtained by sampling historical bandwidth and historical input gain through a sliding window; A prediction module, configured to calculate the joint probability distribution of the training set and the test set by using a kernel function, and determine the predicted mean and predicted variance of the optimized variable; a screening module that screens optimization variable combinations based on an expected value acquisition function and updates the bandwidth of the extended state observer and the input gain of the hyperlocal model based on the screened optimization variable combinations; the screening module also calculates a reference voltage for a next control cycle based on the updated extended state observer and the hyperlocal model and using a deadbeat control principle; A control module is used to decompose the reference voltage into basic voltage vectors through space vector modulation, and control the inverter to execute the switching sequence of the basic voltage vectors according to the minimum switching number to drive the permanent magnet motor.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the continuous set model-free predictive control method for a permanent magnet motor based on Bayesian optimization according to any one of claims 1 to 8 are implemented.

Images