Permanent magnet synchronous motor neural network model prediction control system and control method

Abstract

The invention provides a permanent magnet synchronous motor neural network model predictive control system and a control method, and relates to the technical field of motor control and artificial intelligence crossing, and the method comprises the steps: firstly, collecting the stator current, DC bus voltage, inverter switching signals, motor operation temperature and other multi-dimensional physical quantities in a motor driving system; an advanced architecture of deep fusion neural network and finite control set model prediction control is designed. According to the architecture, an offline trained neural network is utilized, the state sensing ability of a system is dynamically adjusted on line, and full-speed-domain rotating speed and position identification without a physical speed sensor is achieved; and then through online real-time prediction, a weight factor and a voltage vector selection strategy of model prediction control are adaptively adjusted, parameter mismatch and external disturbance are effectively overcome, and the dynamic response speed and the control precision of the system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of motor control and artificial intelligence, specifically to a neural network model predictive control system and control method for permanent magnet synchronous motors. Background Technology

[0002] Permanent magnet synchronous motors (PMSMs) are widely used in demanding applications such as electric propulsion aircraft and new energy vehicles due to their high power density and efficiency. However, traditional control technologies face three major bottlenecks:

[0003] Perception bottleneck: Mechanical speed sensors have poor reliability in harsh environments such as vibration and extreme temperatures, increase system cost and weight, and their estimation accuracy drops sharply under changes in motor parameters (parameter mismatch) or strong interference, especially in the low-speed range.

[0004] Control bottleneck: Although Model Predictive Control (MPC) has excellent performance, its performance is highly dependent on the manual tuning of weighting factors, which is a cumbersome process and cannot adapt to dynamic operating conditions. At the same time, its predictive model is also affected by parameter mismatch and has poor robustness.

[0005] Architecture bottleneck: Different control architectures need to be designed for different motor and inverter topologies (such as open winding and multi-level), resulting in high system redundancy, poor versatility, and difficulty in engineering implementation.

[0006] Existing technologies lack an integrated solution that can fundamentally simulate the real dynamics of motors and is inherently robust to model mismatch and disturbances. Summary of the Invention

[0007] In view of this, embodiments of this application provide a neural network model predictive control system and control method for permanent magnet synchronous motors. By constructing a data-driven virtual motor model predictive control model, high-precision state perception and intelligent control decision-making are achieved, completely eliminating the dependence on physical sensors and fixed mathematical models, effectively overcoming parameter mismatch and external disturbances, and significantly improving the dynamic response speed and control accuracy of the system.

[0008] This application provides the following technical solution: a neural network model predictive control system for a permanent magnet synchronous motor, comprising: a signal collection module, a sensorless speed module, a neural network simulation predictive control module, and an inverter module.

[0009] Furthermore, the signal acquisition module is used to collect physical quantities of the motor drive system in real time, including stator phase current, DC bus voltage, and inverter switching status. This module provides an accurate real-time data foundation for all subsequent calculations.

[0010] Furthermore, the sensorless module has its input connected to the signal acquisition module. This module employs a scheme based on the traditional observer algorithm, the Extended Kalman Filter (EKF), to receive electrical signals from the signal acquisition module and output preliminary estimates of the rotor's speed and position. This module serves as a backup or basic sensing unit for the system.

[0011] Furthermore, the neural network simulation predictive control module is the core innovation of this invention. This module contains a neural network acting as a digital twin, designed as a high-precision virtual replica of the dynamic characteristics of a permanent magnet synchronous motor. Its inputs receive electrical quantities (such as dq-axis current and voltage) from the signal acquisition module, and its output directly provides a predicted value of the system's future state, primarily the predicted current. The module's workflow is as follows: First, it receives preliminary speed / position estimates from the sensorless module and real-time electrical quantities from the signal acquisition module; then, its internal neural network digital twin, based on these inputs, simulates the dynamic response of the motor under different voltage vectors in the future, providing a high-precision predictive model for the rolling optimization process of model predictive control (MPC); finally, based on the prediction results of this digital twin, MPC evaluates and selects the optimal voltage vector that minimizes the cost function.

[0012] Furthermore, the inverter module has its input terminal connected to the output terminal of the neural network analog predictive control module, which is used to receive the optimal switching signal generated by the module and drive the power switching transistors to control the operation of the permanent magnet synchronous motor.

[0013] Secondly, the present invention provides a control method for the above-mentioned system, comprising the following steps: First, a neural network digital twin for MPC is constructed and trained offline. Historical motor operating data covering all operating conditions is collected. Historical voltage vectors, historical currents, and speeds are used as input features, and future current responses are used as training labels to train the neural network model. This enables the model to accurately predict the motor's dynamic behavior under given inputs, forming a prediction model specifically for MPC. Specifically, stator phase currents are collected. , , , For phase a stator current, For phase b stator current, This refers to the c-phase stator current; the inverter switching state is... , , The current can be transformed using the Clark transformation to obtain the current in the rotating coordinate system, as shown in Equation 1 below. Then, it can be transformed using the Park transformation to obtain the current in the stationary coordinate system, as shown in Equation 2 below. Simultaneously, the voltage vector can be obtained from the inverter's switching state, as shown in Equation 3 below.

[0014] Clarke transform (ABC→αβ): Formula 1 in, and Representing the coordinates in the stationary coordinate system shaft current and shaft current; , , These are the three-phase stator current components of the motor.

[0015] Park transform (αβ→dq): Formula 2 in The rotor electrical angle is given.

[0016] Inverter output voltage vector (in the ABC coordinate system): Formula 3 in, , , The motor is in Three-phase stator voltage in coordinate system; This is the bus voltage on the DC side of the inverter; , , These are the three-phase bridge arm switch status signals of the inverter. This voltage vector It also needs to go through Clarke and Park transformations to obtain and .

[0017] Step 2, Online sensorless real-time control: Secondly, the online sensorless real-time control process acquires stator phase current, DC bus voltage, and inverter switching status in real time through the signal acquisition module. The sensorless speed module then processes and outputs preliminary rotor speed and position estimates. These estimates, along with real-time electrical quantities, are input to the neural network simulation predictive control module. The module uses its internal neural network digital twin to simulate predicted currents under different candidate voltage vectors. The model predictive controller selects the optimal voltage vector based on the error assessment between the predicted and reference currents. Finally, the corresponding switching signal is applied to the motor through the inverter module to complete closed-loop control. The above control calculation process utilizes the electromagnetic torque equation, specifically for the integrated permanent magnet synchronous motor (IPMSM): Formula 4 For surface-mounted permanent magnet synchronous motors (SPMSM), ): Formula 5 in, This refers to the electromagnetic torque of the motor. This represents the number of pole pairs of the motor. For permanent magnet flux linkage; It is a direct-axis inductor; It is a quadrature axis inductor.

[0018] The relationship between rotor position and speed can be calculated using the following formula: Formula Six Formula 7 Formula 8 in, For the mechanical angle of the motor; This refers to the mechanical angular velocity of the motor. Let be the electric angular velocity of the motor.

[0019] Combining the above formulas, we can finally obtain the mechanical motion equations, and then consider the relationships between the system's moment of inertia, load torque, electromagnetic torque, coefficient of friction, and mechanical angular velocity: Formula Nine in, Let the system's moment of inertia be denoted by . This refers to the external load torque. It is the coefficient of viscous friction; This is the derivative of the mechanical angular velocity with respect to time.

[0020] Compared with the prior art, the beneficial effects that at least one technical solution adopted in the embodiments of this specification can achieve include at least: (1) This invention proposes a fundamental change in the construction of predictive models by using a deep learning-based neural network digital twin to completely replace the linear predictive model used in traditional model predictive control. This innovative architecture establishes a direct mapping relationship from the multidimensional electrical signals of the motor to its dynamic response through an end-to-end learning approach, achieving high-precision modeling of the complex nonlinear characteristics, multivariable coupling relationships, and runtime-varying parameters of the permanent magnet synchronous motor. This digital twin can simulate the real dynamic behavior of the motor in real time, fundamentally overcoming the modeling error problem of traditional predictive models based on fixed parameter equations, resulting in an order-of-magnitude improvement in the prediction accuracy of model predictive control, and laying a solid foundation for high-performance control.

[0021] (2) This invention proposes an innovative solution for system robustness design. It utilizes the strong generalization capability of a data-driven neural network digital twin to construct a predictive control system with inherent adaptive characteristics. This digital twin is trained using massive amounts of multi-condition data covering the entire operating domain, and its network weight parameters naturally contain a compensation mechanism for changes in motor parameters. When the system encounters parameter mismatch conditions caused by drastic temperature fluctuations, magnetic saturation effects, etc., this architecture can automatically adjust the predictive characteristics to maintain the accuracy of the predictive output. This achieves a fundamental breakthrough in maintaining the stability and dynamic quality of the control system under complex operating environments, solving the technical problems of parameter identification lag and high model complexity in traditional adaptive control.

[0022] (3) This invention proposes an innovative modular collaborative scheme at the system-level architecture design level. By decoupling the sensorless module from the neural network predictive control module, a perception-decision separation intelligent control architecture is constructed. This design allows the two core components, state estimation and optimal control, to be independently and deeply optimized. It retains the computational efficiency and reliability advantages of traditional observers under normal operating conditions while fully leveraging the intelligent decision-making capabilities of neural network predictive control under complex operating conditions. The two modules interact through information to create a performance synergy effect, ultimately constructing a high-performance control platform that combines engineering practicality, advanced algorithms, and system scalability, providing a new architectural paradigm for the design of next-generation motor control systems. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a block diagram of the overall architecture of the system of the present invention; Figure 2 This is a flowchart illustrating the offline training and online deployment process of the neural network digital twin of the present invention. Figure 3 This is a schematic diagram of a sensorless Simulink simulation model incorporating sliding mode control in an embodiment of the present invention. Figure 4 This is a schematic diagram of the dynamic response curve of the EKF-based sensorless system in an embodiment of the present invention. Figure 5 This is a schematic diagram of the EKF speed estimation error curve in an embodiment of the present invention; Figure 6This is a Simulink simulation design block diagram for neural network model predictive control in an embodiment of the present invention; Figure 7 This is a schematic diagram of the dynamic performance test results of Model Predictive Control (MPC) in an embodiment of the present invention. Detailed Implementation

[0025] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0026] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0027] The present invention proposes a sensorless permanent magnet synchronous motor neural network model predictive control system architecture as follows: Figure 1 As shown, the architecture includes a permanent magnet synchronous motor, a three-phase inverter, and a neural network model predictive control system. The control system is completely isolated from external speed sensors, enabling full-state observation and sensorless control. The neural network model predictive control system is directly connected to the motor's electrical system, forming an intelligent system that integrates digital twins and predictive control. This includes, but is not limited to, a hybrid predictive architecture that combines adaptive observer state estimation with neural network forward prediction. The control system includes a signal acquisition module, a sensorless state observation module, a neural network digital twin prediction module, and a multi-objective optimization decision module. The signal acquisition module is responsible for acquiring stator phase current and DC bus voltage. The sensorless state observation module outputs preliminary estimates of rotor position and speed. The neural network digital twin prediction module generates a system state prediction sequence through forward inference. The multi-objective optimization decision module outputs the optimal voltage vector based on the prediction results. Finally, the inverter drive module achieves closed-loop control of the motor.

[0028] This embodiment presents a sensorless permanent magnet synchronous motor neural network model predictive control system architecture. It features a motor control platform centered on a digital signal processor (DSP), with a high-performance DSP main controller responsible for all computation and control tasks. A sensor system includes at least two Hall effect current sensors to measure stator phase current. and A voltage Hall sensor is used to measure the DC bus voltage. A standard three-phase voltage source inverter consists of six IGBT or SiC MOSFET power transistors and their driving circuits, along with a signal conditioning circuit that amplifies, biases, and filters the sensor output signal to adapt it to the input range of the DSP's internal ADC.

[0029] In this embodiment, the flowchart of the offline training and online deployment of the neural network digital twin is as follows: Figure 2 As shown. Offline training of the neural network digital twin is a key preparatory step for realizing intelligent predictive control, and it is completed on the host computer before the system is put into use. First, a high-precision experimental platform is built and a full-condition historical dataset is collected. In the experimental platform, a high-precision photoelectric encoder is temporarily installed on the motor to obtain the real rotor position. and rotational speed As training labels. Simultaneously, stator phase current, DC bus voltage, and inverter switching status are collected. As input features, further data preprocessing and construction are performed. The collected raw data is transformed using Clarke and Park transforms to calculate... , , , and electrical angle Electric angular velocity Build dataset , , that is, use State prediction at time The current at any given moment. Then, reconstruct and train the neural network model, building a fully connected deep neural network model on the host computer. ,in and Here are the network weights and biases. The mean squared error (MSE) is used as the loss function. The network is trained using the Adam optimizer. Finally, the model is deployed by exporting the trained neural network model parameters and converting them into C code that can run on a DSP.

[0030] The online deployment of the neural network digital twin is the main process of system operation. After the system is powered on, the DSP performs operations in each interrupt cycle. The internal signal acquisition module will first trigger the ADC to read the signal. , and The sampled value. The inverter switching state from the last cycle is directly obtained from the register. , , Then, the sensorless module operates, performing Clarke and Park transforms on the current signal to obtain... and Combined with the voltage vector of the previous cycle , The observer algorithm based on the extended Kalman filter (EKF) is run to output a preliminary estimate of the rotor position for the current cycle. and preliminary estimate of rotational speed .

[0031] To verify the feasibility of the sensorless control scheme described in this invention and to provide a benchmark for subsequent neural network enhancement, a sensorless control simulation model of a permanent magnet synchronous motor (PMSM) based on extended Kalman filter (EKF) was constructed, and its dynamic performance was tested. The architecture of the simulation model is as follows: Figure 3 As shown, a surface-mounted permanent magnet synchronous motor (SPMSM) is used as the controlled object. Its key parameters are strictly set according to the mathematical model of the motor, and the number of pole pairs is n. p =4, stator resistance is R s =1.2Ω, stator inductance is L s =8mH, permanent magnet magnetic 0.18Wb An EKF state estimation module that integrates the discretized state equations of the PMSM. Its input is obtained after Clark and Park transforms. The coordinate system current signal has been used to obtain the real-time estimated speed of the output rotor. With electrical angle It includes an Active Disturbance Rejection Control (ADRC) current loop to improve the system's disturbance rejection performance and provide a more stable current environment for EKF estimation. Furthermore, ADRC controllers are designed separately for the d-axis and q-axis currents. adrc and An ADRC module is used to replace a traditional PI controller. An SVPWM modulation module utilizes the electrical angle output from the EKF. Perform inverse Park transform to generate The voltage command in the coordinate system is used to generate the switching signal to drive the inverter through space vector pulse width modulation (SVPWM) technology, thus forming a complete control closed loop.

[0032] To address the common load surge conditions in low-altitude electric propulsion systems, dynamic performance tests were conducted on the simulation model to verify its disturbance rejection capability. The test system started from rest (0 rpm) with a target steady-state speed of 1000 rpm. At t=0.2s, a step load torque of 10 N·m was instantaneously applied. The system's dynamic speed response curve is shown below. Figure 4As shown in the figure. The results indicate that after starting from 0 rpm, the system can rapidly increase to 1200 rpm within approximately 0.02 s and smoothly converge to a steady-state value of 1000 rpm, exhibiting excellent start-up characteristics. After a sudden load change is introduced at 0.2 s, the speed drops momentarily to approximately 950 rpm. However, the system demonstrates excellent recovery capability, and after only a brief fluctuation of approximately 0.03 s, the speed quickly and accurately recovers to a steady-state of 1000 rpm. Figure 5 The error between the EKF-estimated rotational speed and the actual rotational speed was further quantitatively demonstrated.

[0033] Therefore, this simulation test effectively verifies that the sensorless speed control scheme based on EKF has good dynamic response and disturbance rejection capability under sudden load changes, and can provide basic usable state feedback for the control system. Meanwhile, Figure 5 The revealed estimation error provides a clear performance benchmark and basis for improvement for the subsequent introduction of neural network digital twins for high-precision correction and prediction optimization, highlighting the progressiveness of the present invention.

[0034] The neural network simulation predictive control module can... , , and all possible candidate voltage vectors , Combined into input vector .in, , express A set of candidate voltage vectors under the axis, which is mapped from all possible combinations of switching states of the inverter during the current sampling period. The digital twin invokes a deployed neural network digital twin. For each candidate voltage vector, forward inference is performed to predict the current value at the next moment after the vector is applied: Formula 10 in, For the forward inference function of a neural network digital twin; , for Candidate voltage vector components under the axis; subscript Indicates the current moment. This indicates the predicted next moment.

[0035] Optimizing Decisions: Constructing a Cost Function For each candidate vector, calculate: Formula Eleven in, This is the evaluation value of the cost function; , They are respectively shaft and Shaft current reference value; As a weighting factor; This is the switching frequency evaluation function. This is used to balance the constraint relationship between current tracking error and inverter switching frequency. It iterates through all candidate vectors and selects the one that makes the current tracking error equal to the inverter switching frequency. The smallest voltage vector is taken as the optimal vector. And convert it into a specific switching signal. , , To construct a performance benchmark for the neural network-enhanced model predictive control (MPC) scheme described in this invention and to verify the superiority of its core architecture, a current loop simulation model based on finite control set MPC (FCS-MPC) was developed and tested. The FCS-MPC simulation model is constructed based on the predictive model, objective function, and optimization process, and is used to control a permanent magnet synchronous motor driven by a two-level voltage source inverter (2L-VSI). Its core modules are as follows: Figure 6 As shown. The prediction model integrates the discretized state equations of the PMSM, discretized using the forward Euler method, with a sampling period (T). s =1s). Its input includes dq-axis current. , dq axis voltage , Rotor speed electrical angle and complete set of motor parameters , , , This is used to achieve accurate prediction of current at future moments. The motor parameters include: stator resistance. d-axis inductance q-axis inductance and permanent magnet flux . The input q-axis current is set, and the execution and feedback module receives the switching signal output from the MPC to drive the 2-Level H-bridge inverter; the motor model outputs the speed and current. , ),Voltage( , ), electrical angle ( Torque Real-time status variables, such as , form a closed-loop control system.

[0036] The core algorithm of the model embodies the complete workflow of FCS-MPC. The prediction model is constructed by discretizing the dq-axis current equation with a sampling period of T=1μs, achieving multi-step current prediction. The algorithm first calculates the current at time k+1. , Then, iterates through all effective voltage vectors to predict the current value at time k+2. Voltage vector traversal and transformation: First, seven effective basic voltage vectors of the two-level inverter are defined, and then mapped to the dq rotating coordinate system through Clark and Park transformations, realizing the transformation from a three-phase topology to a decoupled coordinate system. With "minimizing current tracking error" as the core objective, the predicted current and reference current under the action of each voltage vector are calculated. , The absolute error (output from the speed loop) is calculated. The optimal voltage vector that minimizes the objective function is selected by minimizing the function, and the corresponding three-phase switch states are directly output, achieving a simplified computational burden strategy combining "vector pre-selection and enumeration optimization". Dynamic operating condition tests were conducted based on this model to verify the performance of the MPC in low-altitude electric propulsion scenarios. The results are as follows: Figure 7 As shown in the figure. The results show that after the system starts from 0 rpm, it can quickly reach a steady state of 1000 rpm within about 0.01 s without overshoot, and the steady-state speed error is ≤0.5%, which fully demonstrates the high accuracy and speed of FCS-MPC in speed tracking. After a sudden load change of 10 N·m is introduced at t=0.2 s, the speed fluctuation amplitude is <1%, and it can instantly recover to the steady state with a recovery time of <0.01 s. The recovery time of traditional PI controller is >0.05 s, and the overshoot is 5%~8%. In comparison, its dynamic performance is significantly improved, verifying the strong anti-interference capability of MPC under multi-objective optimization (current tracking and switching frequency suppression).

[0037] Finally, the DSP writes the calculated optimal switching signal into the corresponding register of its ePWM module. Based on the register value, the ePWM module automatically generates PWM pulses to control the inverter to drive the permanent magnet synchronous motor.

[0038] Through the aforementioned hardware and software co-design, this invention successfully embeds a complex neural network digital twin into the real-time control loop of a DSP, achieving high-performance, highly intelligent sensorless predictive control. This solves a series of comprehensive bottleneck problems faced by traditional control systems of permanent magnet synchronous motors in terms of accuracy, robustness, real-time performance, and engineering complexity under the harsh operating conditions of high-end equipment such as low-altitude electric propulsion aircraft.

[0039] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A permanent magnet synchronous motor neural network model predictive control system, characterized in that, The system comprises: a signal collection module for collecting electrical quantities of the permanent magnet synchronous motor in real time, the electrical quantities including stator phase current, DC bus voltage and switch state of the inverter; a speed sensorless module, an input end of the speed sensorless module being connected to the signal collection module, for receiving the electrical quantities and outputting preliminary estimation values of rotor speed and position based on an observer algorithm; a neural network model predictive control module, input ends of the neural network model predictive control module being connected to output ends of the signal collection module and the speed sensorless module respectively; the neural network model predictive control module being internally provided with a neural network digital twin, for predicting the state of the motor at a future set time according to the preliminary estimation values of the speed and position and real-time electrical quantities; based on the predicted state, performing rolling optimization through a model predictive control algorithm to output optimal switch signals; an inverter module, an input end of the inverter module being connected to an output end of the neural network model predictive control module, for executing the optimal switch signals to drive the permanent magnet synchronous motor.

2. The system of claim 1, wherein, The neural network digital twin is obtained through offline training, wherein the training data includes historical voltage vector, historical current and historical speed of the permanent magnet synchronous motor, and the training target is the current response of the permanent magnet synchronous motor at a future set time, so that the neural network digital twin can serve as a high-precision prediction model inside the model predictive control algorithm.

3. The system of claim 1 or 2, wherein, The observer algorithm adopted by the speed sensorless module is an extended Kalman filter.

4. The system of claim 1 or 2, wherein, The working process of the neural network model predictive control module comprises: receiving the preliminary estimation values of the speed and position output by the speed sensorless module and the real-time current and voltage from the signal collection module; calling the neural network digital twin to simulate the dynamic response of the permanent magnet synchronous motor under the action of different candidate voltage vectors and output corresponding predicted current; constructing a cost function with the tracking error of the predicted current and the reference current as the core, and selecting the optimal voltage vector by evaluating the cost function; converting the optimal voltage vector into corresponding switch signals and outputting the switch signals.

5. The system of claim 1 or 2, wherein, The system further comprises a coordinate transformation module connected between the signal collection module and the speed sensorless module, for performing Clark transformation and Park transformation on the original data collected by the signal collection module to convert the original data from three-phase topology to decoupled coordinate system.

6. The system of claim 1 or 2, wherein, The stator phase current collected by the signal collection module is and and the third phase current is calculated by the formula wherein, is the a-phase stator current, is the b-phase stator current, is the c-phase stator current.​ 7. A control method for the system of any one of claims 1 to 6, characterized in that, The system comprises the following steps: a neural network model predictive control module offline training step: collecting historical operation data of the permanent magnet synchronous motor under multiple working conditions, taking historical voltage vector, historical current and historical speed as input, and taking current response at a future set time as output target, training a neural network to construct a neural network digital twin for model predictive control; an online control step: collecting real-time electrical quantities of the permanent magnet synchronous motor in real time through the signal collection module, the electrical quantities including stator phase current, DC bus voltage and inverter switch state; The data collected by the signal collection module is processed by the speed sensorless module to output a preliminary estimation of the rotor speed and position; The preliminary estimation of the rotor speed and position is input into the neural network model predictive control module along with real-time electrical quantities; In the neural network model predictive control module, the neural network digital twin predicts the state of the permanent magnet synchronous motor at a future set time and generates optimal switching signals based on a model predictive control algorithm; The optimal switching signals are applied to the permanent magnet synchronous motor through an inverter module.

8. The control method according to claim 7, characterized by, In the offline training step, the historical operation data covers the full speed range, full load range and dynamic process of the permanent magnet synchronous motor.

Images