Permanent Magnet Motor Predictive Control Method Based on Self-Learning Ultra-Local Model

Abstract

The present invention provides a predictive control method for a permanent magnet synchronous motor based on a self-learning hyper-local model, which relates to the field of permanent magnet motor control. The method includes: constructing a historical data queue of the input-output states of the controlled motor, the queue satisfying the first-in first-out rule, and its outputs being respectively used for parameter self-learning and model prediction; according to the self-learning algorithm, online updating the parameters of the lumped disturbance term and the control input gain in the speed hyper-local model and the current hyper-local model; using the updated hyper-local model and the current state to implement the robust predictive control of the controlled motor; wherein, the lumped disturbance term and the control input gain are approximated by a neural network, and at the same time, relevant stability conditions and analyses are provided. The method only uses historical data to approximate the actual dynamics of the system, alleviating the negative impact on the control performance caused by parameter changes and unmodeled disturbances; there is no need to design the control input gain and the observer bandwidth according to prior physical parameters, further broadening the applicable range of the robust algorithm.

Description

Technical Field

[0001] The present invention relates to the field of predictive control of permanent magnet motors, and specifically provides a predictive control method and device for permanent magnet motors based on a self-learning hyperlocal model. Background Art

[0002] The finite set model predictive control (FCS-MPC) of permanent magnet motors has been widely applied because of its simple logic, convenience for implementation in digital systems, fast dynamic response, and ability to handle multi-variable and multi-constraint systems. However, its performance depends on an accurate system model, and parameter variations and unmodeled dynamics will greatly affect the control performance.

[0003] In related technologies, the hyperlocal model combined with an observer is the main approach to improving the robust performance of traditional predictive control of permanent magnet motors. However, the design value of the control input gain in the hyperlocal model is a fixed value and depends on prior physical parameters. When the parameters change, the fixed control input gain will restrict the robust performance of such control methods. In addition, the bandwidth of the observer in the hyperlocal model needs to be manually tuned by the designer and kept fixed. The setting of the bandwidth needs to balance the robustness and sensitivity of the control method. Therefore, the selected fixed bandwidth is difficult to adapt to all working conditions, limiting the application scope of such methods and possibly leading to a decrease in control accuracy or system oscillation under complex working conditions. Summary of the Invention

[0004] To solve the deficiencies of the prior art, the purpose of this application is to provide a predictive control method and device for permanent magnet motors based on a self-learning hyperlocal model. This method can effectively solve the problem of insufficient working condition adaptability of the finite set predictive control method for permanent magnet motors in related technologies.

[0005] In a first aspect, this application provides a predictive control method for permanent magnet motors based on a self-learning hyperlocal model. The method includes:

[0006] Construct a self-learning hyperlocal model for the permanent magnet motor. The self-learning hyperlocal model includes a speed hyperlocal model and a current hyperlocal model. The lumped disturbance terms and control input gains in the speed hyperlocal model and the current hyperlocal model are both online self-learning, and the self-learning parts in the speed hyperlocal model and the current hyperlocal model are both approximated by an online neural network or fuzzy logic;

[0007] Construct a historical data queue of the input-output states of the permanent magnet motor. The historical data queue satisfies the first-in-first-out update rule, and the input-output states include speed, current, and voltage;

[0008] Collect the input-output states of the permanent magnet motor at the current moment and update the historical data queue;

[0009] Utilize the data in the updated historical data queue and the first self - learning formula to update the lumped disturbance term and control input gain in the speed super - local model, and utilize the data in the updated historical data queue and the second self - learning formula to update the lumped disturbance term and control input gain in the current super - local model;

[0010] Substitute the set reference speed into the updated speed super - local model to obtain the optimal reference current, or, substitute the set reference speed into the updated speed super - local model to obtain the optimal reference current, and then substitute the optimal reference current into the updated current super - local model to obtain the optimal input voltage;

[0011] Select the optimal switching combination according to the optimal reference current and its corresponding current cost function or according to the optimal input voltage and its corresponding voltage cost function to drive the permanent - magnet motor to operate.

[0012] In one of the embodiments, the speed super - local model is:

[0013] ;

[0014] Wherein, represents the sampling period of the permanent - magnet motor speed, represents the predicted speed of the permanent - magnet motor at time represents the speed of the permanent - magnet motor collected at time represents the q - axis current collected at time and respectively represent the lumped disturbance term and control input gain of the speed super - local model;

[0015] Specifically, and are both approximated by the on - line neural network in the speed super - local model, represents the weight of the neural network corresponding to in the speed super - local model at time represents the weight of the neural network corresponding to in the speed super - local model at time represents the input vector of the neural network in the speed super - local model at time;

[0016] The current super - local model is:

[0017] ;

[0018] Wherein, represents the sampling period of the current, denotes the predicted current on the q-axis at time denotes the predicted current on the d-axis at time denotes the d / q-axis current collected at time denotes the d / q-axis voltage collected at time and respectively denote the lumped disturbance term and the control input gain of the current super-local model;

[0019] Specifically, and are both approximated by the online neural network in the current super-local model, denotes the weight of the neural network corresponding to in the current super-local model at time denotes the weight of the neural network corresponding to in the current super-local model at time denotes the input vector of the neural network in the current super-local model at time;

[0020] The neural networks both have one hidden layer. The input of the neural network is a vector composed of the data in the historical data queue. The queue length of the historical data queue is selected according to the required control accuracy and the computing speed of the hardware;

[0021] The neural networks can all be replaced by fuzzy logic.

[0022] In one embodiment, the first self-learning formula is obtained by minimizing the first error, and the first error is the error between the predicted speed and the actual speed;

[0023] The first self-learning formula is:

[0024] ;

[0025] wherein, denotes the predicted speed of the permanent magnet motor at time denotes the actual speed of the permanent magnet motor at time denotes the learning rate of the first self-learning formula.

[0026] In one embodiment, the second self-learning formula is obtained by minimizing the second error, and the second error is the error between the predicted current and the actual current;

[0027] The second self-learning formula is:

[0028] ;

[0029] Wherein, represents the predicted current of the permanent magnet motor at time represents the actual current of the permanent magnet motor at time represents the learning rate of the second self - learning formula.

[0030] In one embodiment, the stability condition and parameter selection condition of the first self - learning formula are as follows:

[0031] For the speed super - local model, construct the following first Lyapunov function:

[0032] ;

[0033] Wherein, , , , and are respectively and the corresponding optimal weights;

[0034] Substitute the first self - learning formula into the first Lyapunov function, and when is satisfied, we get: ;

[0035] Wherein, represents the upper bound of the activation function in the neural network, is the upper bound of the fitting error of the neural network under the optimal weight, ;

[0036] According to the approximation theorem, the upper bound of the fitting error decreases as the number of neurons in the hidden layer of the neural network increases. When is satisfied and the design of the first learning rate satisfies , we get , that is, the final consistency stability is satisfied.

[0037] In one embodiment, the stability condition and parameter selection condition of the second self - learning formula are as follows:

[0038] For the current super - local model, construct the following second Lyapunov function:

[0039] ;

[0040] Wherein, , ; and respectively are and the corresponding optimal weights;

[0041] Substitute the second self - learning formula into the second Lyapunov function, and when satisfied, we get: ;

[0042] wherein, represents the upper bound of the activation function in the neural network, is the upper bound of the fitting error of the neural network under the optimal weight, ;

[0043] According to the approximation theorem, the upper bound of the fitting error decreases as the number of neurons in the hidden layer of the neural network increases. When is satisfied and the design of the second learning rate satisfies at this time, we get , that is, the final consistency stability is satisfied.

[0044] In one of the embodiments, the calculation method of the optimal reference current is: Substitute and the reference speed into the updated speed super - local model to obtain the optimal reference current;

[0045] The calculation formula of the optimal reference current is:

[0046] ;

[0047] wherein, , , represents the optimal reference current, represents the current amplitude constraint, represents the speed of the permanent - magnet motor collected at time represents the input vector of the neural network in the speed super - local model at time represents the sampling period of the permanent - magnet motor speed;

[0048] The current cost function is:

[0049] ;

[0050] wherein, , , represents the input vector of the neural network in the current super - local model at time represents the sampling period of the current, Indicates the d / q axis voltages corresponding to the i th switching combination, and indicates the d / q axis currents collected at

[0051] In one embodiment, the calculation formula for the optimal input voltage is:

[0052] ;

[0053] where , , represents the sampling period of the current, and indicates the d / q axis currents collected at while represents the input vector of the neural network in the current super-local model at , indicates the optimal reference current at , and

[0054] The voltage cost function is: ;

[0055] where indicates the d / q axis voltages corresponding to the i th switching combination.

[0056] In a second aspect, the present application also provides a permanent magnet motor predictive control device based on a self-learning super-local model, and the device includes:

[0057] A construction module, configured to construct a self-learning super-local model for the permanent magnet motor. The self-learning super-local model includes a speed super-local model and a current super-local model. The lumped disturbance terms and control input gains in the speed super-local model and the current super-local model are both online self-learning, and the self-learning parts in the speed super-local model and the current super-local model are both approximated by an online neural network or fuzzy logic. The construction module is further configured to construct a historical data queue of the input-output states of the permanent magnet motor, and the historical data queue satisfies the first-in first-out update rule. The input-output states include speed, current, and voltage;

[0058] An acquisition module, configured to acquire the input-output states of the permanent magnet motor at the current moment and update the historical data queue;

[0059] An update module, configured to use the data in the updated historical data queue and the first self - learning formula to update the lumped disturbance term and the control input gain in the rotational speed super - local model, and use the data in the updated historical data queue and the second self - learning formula to update the lumped disturbance term and the control input gain in the current super - local model;

[0060] A selection module, configured to substitute a set reference rotational speed into the updated rotational speed super - local model to obtain an optimal reference current, or substitute the set reference rotational speed into the updated rotational speed super - local model to obtain an optimal reference current, and substitute the optimal reference current into the updated current super - local model to obtain an optimal input voltage; the selection module is further configured to select an optimal switch combination according to the optimal reference current and its corresponding current cost function or according to the optimal input voltage and its corresponding voltage cost function to drive the permanent - magnet motor to operate.

[0061] In a third aspect, the present application further provides a computer device, including a memory and a processor, where the memory stores a computer program, and when the processor executes the computer program, it implements the permanent - magnet motor predictive control method based on a self - learning super - local model in the first aspect.

[0062] The above - mentioned permanent - magnet motor predictive control method based on a self - learning super - local model includes: constructing a rotational speed super - local model and a current super - local model, where the lumped disturbance term and the control input gain in the rotational speed super - local model and the current super - local model are dynamically updated through an online self - learning mechanism; constructing a historical data queue following the first - in - first - out principle for storing the input - output state data (including rotational speed, current, and voltage) of the permanent - magnet motor and updating it in real time; using the data in the updated historical data queue and the first self - learning formula to update the lumped disturbance term and the control input gain in the rotational speed super - local model, and using the data in the updated historical data queue and the second self - learning formula to update the lumped disturbance term and the control input gain in the current super - local model; selecting an optimal inverter switch combination according to the optimal reference current and its corresponding current cost function or the optimal input voltage and its corresponding voltage cost function to drive the motor to operate.

[0063] By introducing an online self - learning mechanism into the super - local model, this solution only uses historical data to approximate the actual dynamics of the system, realizes the online update of the control input gain and the lumped disturbance term, and alleviates the negative impact on the control performance caused by parameter changes and unmodeled disturbances. This solution does not need to design the control input gain and the observer bandwidth according to physical information, reduces the dependence on prior motor parameters and manual experience, and can further improve the robustness and applicability of predictive control. Description of the Drawings

[0064] Figure 1Flowchart of the predictive control method for permanent magnet motors based on self - learning hyper - local model in an embodiment;

[0065] Figure 2 Structural diagram of the predictive control method for permanent magnet motors based on self - learning hyper - local model in an embodiment;

[0066] Figures 3(a) and 3(b) are respectively the control effect diagrams of the rotational speed and current of the permanent magnet motor in an embodiment;

[0067] Figure 4 Device diagram of the predictive control device for permanent magnet motors based on self - learning hyper - local model in an embodiment;

[0068] Figure 5 Internal structural diagram of a computer device in an embodiment. Detailed implementation manners

[0069] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application. Unless otherwise defined, the technical terms or scientific terms involved in the present application should have the general meanings understood by those with ordinary skills in the technical field to which the present application belongs.

[0070] In an embodiment, as Figure 1 shown, a predictive control method for permanent magnet motors based on self - learning hyper - local model is provided, and the method includes the following steps:

[0071] Step 101: Construct a self - learning hyper - local model for the permanent magnet motor. The self - learning hyper - local model includes a rotational speed hyper - local model and a current hyper - local model. The lumped disturbance terms and control input gains in the rotational speed hyper - local model and the current hyper - local model are both online self - learning, and the self - learning parts in the rotational speed hyper - local model and the current hyper - local model are both approximated by an online neural network or fuzzy logic.

[0072] A self - learning hyper - local model is constructed for the permanent - magnet motor. The self - learning hyper - local model includes a speed hyper - local model and a current hyper - local model. The speed hyper - local model includes a speed equation. The speed hyper - local model is used to describe the relationship between the speed of the permanent - magnet motor and the q - axis current. The speed hyper - local model contains a lumped disturbance term and a control input gain. The current hyper - local model includes a current equation. The current hyper - local model is used to describe the relationship between the current and the voltage. The current hyper - local model contains a lumped disturbance term and a control input gain. Among them, the control input gain is used to describe the gain of the system input quantity, and the lumped disturbance term is used to describe the combined influence of internal disturbances and external disturbances. Among them, the internal disturbances include changes in the parameters of the permanent - magnet motor, and the external disturbances include sudden changes or fluctuations in the load, changes in environmental conditions, etc.

[0073] The speed hyper - local model is used to predict the change in the motor speed. Through an online self - learning mechanism, the lumped disturbance term and the control input gain can be updated in real time to adapt to the actual operating state. The online self - learning mechanism can utilize historical state data and real - time sampled state data. By minimizing the difference between the predicted value and the actual value, it continuously adjusts the lumped disturbance term and the control input gain, enabling the speed hyper - local model to more accurately reflect the dynamic characteristics of the permanent - magnet motor.

[0074] Similarly, the current hyper - local model is used to predict the change in the motor current. Through an online self - learning mechanism, the lumped disturbance term and the control input gain can be updated in real time to adapt to the actual operating state. The online self - learning mechanism adjusts the lumped disturbance term and the control input gain in real time according to the actual operating state of the permanent - magnet motor, ensuring that the output of the current hyper - local model is consistent with the actual current change of the permanent - magnet motor.

[0075] It should be noted that the self - learning part in the speed hyper - local model and the current hyper - local model is realized by online neural network or fuzzy logic approximation. The neural network approximates complex non - linear dynamic characteristics by learning the input - output relationship in the historical data queue, while fuzzy logic processes uncertainties and ambiguities through preset rules and real - time data adjustment. Both can effectively capture dynamic changes and ensure the prediction accuracy and control performance of the hyper - local model.

[0076] Step 102: Construct a historical data queue of the input - output states of the permanent - magnet motor. The historical data queue follows the update rule of first - in - first - out. The input - output states include speed, current, and voltage.

[0077] The historical data queue is used to store the input - output state data of the permanent - magnet motor. The input - output state data includes speed, current, and voltage. The historical data queue follows the first - in - first - out (FIFO) update rule to ensure the timeliness and relevance of the input - output state data.

[0078] When new input-output status data is collected, the new input-output status data is added to the end of the queue, and the oldest input-output status data in the queue is removed. The first-in, first-out update rule can ensure that the historical data queue always contains the latest and most valuable input-output status data for control.

[0079] Step 103: Collect the input-output status of the permanent magnet motor at the current moment and update the historical data queue.

[0080] Whenever a data sampling is performed, the rotational speed, current, and voltage values at the current moment are recorded as the latest input-output status data. The latest collected input-output status data is added to the end of the historical data queue, and the oldest input-output status data in the queue is removed, thereby keeping the length of the queue unchanged.

[0081] Step 104: Use the data in the updated historical data queue and the first self-learning formula to update the lumped disturbance term and control input gain in the rotational speed super-local model, and use the data in the updated historical data queue and the second self-learning formula to update the lumped disturbance term and control input gain in the current super-local model.

[0082] The first self-learning formula and the second self-learning formula are used to update the lumped disturbance term and control input gain in the rotational speed super-local model and the current super-local model respectively. The core of the first self-learning formula and the second self-learning formula is to dynamically adjust the lumped disturbance term and control input gain in the rotational speed super-local model and the lumped disturbance term and control input gain in the current super-local model by minimizing the difference between the predicted value and the actual value.

[0083] For the rotational speed super-local model, the first self-learning formula can update the lumped disturbance term and control input gain of the rotational speed super-local model by calculating the difference between the predicted rotational speed and the actual rotational speed. The second self-learning formula can update the lumped disturbance term and control input gain of the current super-local model by calculating the difference between the predicted current and the actual current.

[0084] Step 105: Substitute the set reference rotational speed into the updated rotational speed super-local model to obtain the optimal reference current, or, substitute the set reference rotational speed into the updated rotational speed super-local model to obtain the optimal reference current, and substitute the optimal reference current into the updated current super-local model to obtain the optimal input voltage.

[0085] The updated rotational speed super-local model can more accurately reflect the actual operating state of the permanent magnet motor. Input the reference rotational speed into the updated rotational speed super-local model. Furthermore, the rotational speed super-local model can obtain the optimal reference current based on the current input-output status data and the reference rotational speed of the permanent magnet motor.

[0086] It should be noted that the reference speed is usually determined according to the speed requirements and load conditions of the permanent magnet motor. For example, in semiconductor manufacturing equipment, in order to ensure processing accuracy, an accurate reference speed needs to be set. The reference current is the output of the outer loop and the input of the inner loop in the permanent magnet motor control, and is used to guide the actual current adjustment of the permanent magnet motor to achieve the desired torque output.

[0087] The current hyperlocal model is used to describe the relationship between current and voltage. Through the current hyperlocal model, the optimal reference current can be converted into the optimal input voltage.

[0088] Step 106: Select the optimal switching combination according to the optimal reference current and its corresponding current cost function or according to the optimal input voltage and its corresponding voltage cost function to drive the permanent magnet motor to operate.

[0089] There are two ways to select the optimal switching combination to drive the permanent magnet motor to operate:

[0090] One way: For each possible switching combination of the inverter, use the current hyperlocal model to calculate the corresponding predicted current. Using the optimal reference current and its corresponding current cost function, calculate the difference between the predicted current and the optimal reference current under each switching combination. The smaller this difference, the closer the predicted current under this switching combination is to the optimal reference current, and thus the better this switching combination.

[0091] By traversing all the inverter switching combinations and calculating the corresponding current cost function values, the switching combination with the minimum current cost function value can be selected as the optimal switching combination. This optimal switching combination will be used to control the on / off of the inverter, thereby driving the permanent magnet motor to operate at the desired speed.

[0092] Another way: Substitute the optimal reference current into the current hyperlocal model to obtain the optimal input voltage. Each switching combination of the inverter corresponds to an input voltage. Using the optimal input voltage and its corresponding voltage cost function, calculate the difference between the input voltage corresponding to each switching combination and the optimal input voltage. The smaller this difference, the closer the input voltage under this switching combination is to the optimal input voltage, and thus the better this switching combination.

[0093] By traversing all the inverter switching combinations and calculating the corresponding voltage cost function values, select the switching combination with the minimum voltage cost function value as the optimal switching combination. This optimal switching combination will be used to control the on / off of the inverter, and then drive the permanent magnet motor to operate at the desired speed.

[0094] It should be noted that the inverter, as a key component in the permanent magnet motor, its main function is to convert direct current into alternating current to provide the required voltage and current for the permanent magnet motor. The inverter contains multiple power electronic switching devices, and the combined states of these switching devices determine the voltage waveform and amplitude of the inverter output.

[0095] Through the evaluation of the self-learning super-local model and the cost function, the optimal switching combination has been determined. The optimal switching combination is obtained by evaluating all possible combinations of switching states and selecting the combination that can make the actual current or voltage of the permanent magnet motor closest to the optimal reference value. After determining the optimal switching combination, a corresponding control signal is sent to the inverter, and the control signal is used to drive the switching devices inside the inverter to perform on-off operations according to the optimal combination.

[0096] In this embodiment, the method includes: constructing a speed super-local model and a current super-local model, and the lumped disturbance term and control input gain in the speed super-local model and the current super-local model are dynamically updated through an online self-learning mechanism; constructing a historical data queue that follows the first-in-first-out principle to store the input-output state data (including speed, current, and voltage) of the permanent magnet motor and updating it in real time; using the data in the updated historical data queue and the first self-learning formula to update the lumped disturbance term and control input gain in the speed super-local model, and using the data in the updated historical data queue and the second self-learning formula to update the lumped disturbance term and control input gain in the current super-local model; selecting the optimal inverter switching combination to drive the motor to operate according to the optimal reference current and its corresponding current cost function or the optimal input voltage and its corresponding voltage cost function.

[0097] By introducing the online self-learning mechanism into the super-local model, this method only uses historical data to approximate the actual dynamics of the system, realizes the online update of the control input gain and the lumped disturbance term, and alleviates the negative impact on the control performance caused by parameter changes and unmodeled disturbances. At the same time, this method does not need to design the control input gain and the observer bandwidth according to physical information, reduces the dependence on the prior predicted motor parameters and manual experience, and can further improve the robustness and applicability of the predictive control.

[0098] In one embodiment, as Figure 2 shown, this figure shows the structure of the predictive control method for a permanent magnet motor based on a self-learning super-local model. The input-output state data of the permanent magnet motor (PMSM) is collected in real time, and the input-output state data includes: the speed at time ω ( n ) and the current at time i dq ( k ) and the voltage udq ( k ) Update the historical data queue by following the first-in-first-out update rule using the input-output status data collected in real time. The historical data queue can output the input vector of the neural network in the current moment current super-local model and the input vector of the neural network in the current moment speed super-local model . The historical data sequence can also output the input vector of the neural network in the current moment current super-local model and the input vector of the neural network in the current moment speed super-local model . Based on , , the current moment speed ω ( n ) and the current moment current i dq ( k ) and voltage u dq ( k ), use the first and second self-learning formulas to update the lumped disturbance term weights and control input gain weights in the speed super-local model and the current super-local model. Using the updated lumped disturbance term weights and control input gain weights in the speed super-local model and the current super-local model, the input vector of the neural network in the current moment current super-local model and the input vector of the neural network in the current moment speed super-local model , the updated speed super-local model and the current super-local model can be obtained.

[0099] Input the reference speed ω * , the actual speed and the updated lumped disturbance term and the input gain into the system constraint and speed control module to generate the optimal reference current i dq * . According to the actual current i dq ( k ), the optimal reference current i dq * and the candidate voltages corresponding to different switch states and the lumped disturbance term of the updated current super-local model and the control input gain , calculate the current cost function and select the switching state with the minimum current cost function value , and drive the permanent magnet synchronous motor (PMSM) to operate.

[0100] In one embodiment, the initial speed superlocal model is:

[0101] ;

[0102] Wherein, and are respectively the speed of the permanent magnet motor and the current of the q-axis, and are respectively the lumped disturbance term and the control input gain in the speed equation.

[0103] Perform first-order discretization on the initial speed superlocal model and select a suitable speed sampling period T s1 , divide the continuous time into discrete time steps; further, at each sampling moment , use ω ( n ) to represent the speed value at the moment of ; according to the Euler method, approximate the initial speed superlocal model as a difference equation to obtain the speed superlocal model as:

[0104] ;

[0105] Furthermore, and respectively represent the lumped disturbance term and the control input gain in the speed superlocal model at the moment of , = , = . It should be noted that and are both approximated by the on-line neural network in the speed superlocal model, represents the weight of the neural network corresponding to in the speed superlocal model at the moment of , represents the weight of the neural network corresponding to in the speed superlocal model at the moment of ,

[0106] Therefore, the speed superlocal model can be as follows:

[0107] ;

[0108] Among them, represents the sampling period of the permanent magnet motor speed, represents the predicted speed of the permanent magnet motor at time represents the speed of the permanent magnet motor collected at time represents the q-axis current collected at time and respectively represent the lumped disturbance term and the control input gain of the speed super-local model; and are both approximated by an on-line neural network in the speed super-local model, represents the weights of the neural network corresponding to in the speed super-local model at time represents the weights of the neural network corresponding to in the speed super-local model at time represents the input vector of the neural network in the speed super-local model at time

[0109] In one embodiment, the initial current super-local model is:

[0110] ;

[0111] Among them, is the stator d-axis current; , are the q-axis and d-axis voltages of the stator port; , are the lumped disturbance terms of the q-axis and d-axis in the current equation; , are the control input gains of the q-axis and d-axis in the current equation.

[0112] For the initial current super-local model, select a suitable current sampling period T s2 , and divide the continuous time into discrete time steps. At each sampling moment , use i q ( k ) and i d ( k ) to represent the q-axis and d-axis current values at time

[0113]

[0114] ;

[0115] Furthermore, and respectively represent the lumped disturbance terms and control input gains of the q-axis and d-axis in the current hyperlocal model at time = , = , = and = . Among them, and are both approximated by the online neural network in the current hyperlocal model, represents the weights of the neural network corresponding to in the current hyperlocal model at time represents the weights of the neural network corresponding to in the current hyperlocal model at time represents the input vector of the neural network in the current hyperlocal model at time

[0116] Therefore, the current hyperlocal model can be as follows:

[0117] ;

[0118] Among them, represents the sampling period of the current, represents the predicted current of the q-axis at time represents the predicted current of the d-axis at time represents the d / q-axis current collected at time represents the d / q-axis voltage collected at time and respectively represent the lumped disturbance term and control input gain of the current hyperlocal model, represents the weights of the neural network corresponding to in the current hyperlocal model at time represents the weights of the neural network corresponding to in the current hyperlocal model at time represents The input vector of the neural network in the moment current super-local model.

[0119] In one embodiment, the neural networks all have one hidden layer. The input of the neural network is a vector composed of data in the historical data queue, and the queue length of the historical data queue is selected according to the required control accuracy and the computing speed of the hardware. It should be noted that the neural network in the current super-local model and the neural network in the speed super-local model can both be replaced by fuzzy logic.

[0120] In one embodiment, the speed sampling period is the outer-loop sampling period, and the current sampling period is the inner-loop sampling period. Due to the different time constants of the inner and outer loops, the current sampling period is usually 1 / 5 or 1 / 10 of the speed sampling period .

[0121] In one embodiment, the speed super-local model is substituted into the first error and the first error is minimized to obtain the first self-learning formula of the speed super-local model; based on the first self-learning formula of the speed super-local model, the lumped disturbance term of the updated speed super-local model and the weight of the control input gain are obtained.

[0122] Specifically, the first error is: .

[0123] Substituting the speed super-local model into the first error and minimizing the first error, the first self-learning formula of the speed super-local model can be obtained:

[0124] ;

[0125] where represents the sampling period of the permanent magnet motor speed, represents the weight of the lumped disturbance term of the speed super-local model at time ; represents the weight of the lumped disturbance term of the speed super-local model at time ; represents the input vector of the neural network in the speed super-local model at time ; represents the weight of the control input gain of the speed super-local model at time ; represents the weight of the control input gain of the speed super-local model at time ; Predicted speed of the permanent magnet motor at a moment denotes Actual speed of the permanent magnet motor at a moment (i.e., the collected speed) denotes the learning rate of the first self - learning formula.

[0126] In one embodiment, substitute the current hyper - local model into the second error and minimize the second error to obtain the second self - learning formula of the current hyper - local model; based on the second self - learning formula of the current hyper - local model, obtain the updated lumped disturbance term weights and input gain of weights of

[0127] Specifically, the second error is: .

[0128] Substitute the current hyper - local model into the second error and minimize the second error to obtain the second self - learning formula of the current hyper - local model:

[0129] ;

[0130] wherein denotes the sampling period of the permanent magnet motor current, denotes the lumped disturbance term of the current hyper - local model at moment at moment weights of denotes the lumped disturbance term of the current hyper - local model at moment weights of denotes the control input gain of the current hyper - local model at moment weights of denotes predicted current of the permanent magnet motor at moment denotes actual current of the permanent magnet motor collected at moment denotes the learning rate of the second self - learning formula, denotes input vector of the neural network in the current hyper - local model at moment denotes actual voltage of the permanent magnet motor at moment

[0131] In one embodiment, the stability condition and parameter selection condition of the first self - learning formula are as follows:

[0132] For the speed hyper - local model, construct the following first Lyapunov function:

[0133] ;

[0134] Among them, , , , and are respectively and corresponding optimal weights;

[0135] Substitute the first self - learning formula into the first Lyapunov function, and when is satisfied, we get: ;

[0136] Among them, represents the upper bound of the activation function in the neural network, is the upper bound of the fitting error of the neural network under the optimal weight, ;

[0137] According to the approximation theorem, the upper bound of the fitting error decreases as the number of neurons in the hidden layer of the neural network increases. When is satisfied and the first learning rate is designed to satisfy , we get , that is, the final consistency stability is satisfied.

[0138] In one embodiment, the stability condition and parameter selection condition of the second self - learning formula are as follows:

[0139] For the current hyper - local model, construct the following second Lyapunov function:

[0140] ;

[0141] Among them, , ; and are respectively and corresponding optimal weights;

[0142] Substitute the second self - learning formula into the second Lyapunov function, and when is satisfied, we get: ;

[0143] Among them, represents the upper bound of the activation function in the neural network, is the upper bound of the fitting error of the neural network under the optimal weight, ;

[0144] According to the approximation theorem, the upper bound of the fitting error decreases as the number of neurons in the hidden layer of the neural network increases. When Satisfied and the design of the second learning rate satisfies When , it is satisfied that the final consistency is stable.

[0145] It should be noted that is used to constrain the fitting error and the first learning rate parameters, is used to constrain the fitting error , learning rate parameters. And The value range is between 0 and 1. Constraining the fitting error of the rotational speed super-local model The relationship with the rotational speed error and the disturbance term, Constraining the fitting error of the current super-local model The relationship with the current error and the disturbance term.

[0146] In this embodiment, by constructing a Lyapunov function, the final consistency stability is ensured. By dynamically adjusting the learning rate and neural network parameters, it can adapt to internal and external disturbances, reduce errors, and improve control accuracy and robustness. As the number of neurons in the neural network increases, the fitting error gradually decreases, further improving the control performance.

[0147] In one embodiment, the calculation method of the optimal reference current is: Substitute and the reference rotational speed into the updated rotational speed super-local model to obtain the optimal reference current;

[0148] The calculation formula of the optimal reference current is:

[0149] ;

[0150] Wherein, , , represents the sampling period of the permanent magnet motor rotational speed, and respectively represent the lumped disturbance term and the control input gain of the rotational speed super-local model, represents At the moment The weight of the corresponding neural network in the rotational speed super-local model, represents At the moment The weight of the corresponding neural network in the rotational speed super-local model. represents the optimal reference current, represents the current amplitude constraint, represents The rotational speed of the permanent magnet motor collected at a certain moment, denotes the input vector of the neural network in the rotational speed super-local model at a certain moment, i q ∗ denotes the optimal reference current of the q-axis, i d ∗ denotes the optimal reference current of the d-axis.

[0151] It should be noted that the d-axis current is set to zero ( ), that is, under the maximum torque control strategy, the d-axis current is usually set to zero to reduce torque ripple and improve control performance. Further, the reference rotational speed ω ∗ is substituted into the rotational speed super-local model, and this rotational speed super-local model can predict the rotational speed at the next moment according to the current rotational speed, lumped disturbance parameters, control input gain, and q-axis current. Through the calculation of the rotational speed super-local model, the optimal reference currents i q ∗ and i d ∗ are obtained.

[0152] The current cost function is:

[0153] ;

[0154] wherein, , , denotes the sampling period of the current, and respectively denote the lumped disturbance term and the control input gain of the current super-local model at a certain moment, denotes the weight of the neural network corresponding to in the current super-local model at a certain moment, denotes the weight of the neural network corresponding to in the current super-local model at a certain moment. denotes the input vector of the neural network in the current super-local model at a certain moment, denotes the d / q-axis voltage corresponding to the i th switch combination, denotes the d / q-axis current collected at a certain moment, i q ∗ denotes the optimal reference current.

[0155] It should be noted that the current cost function can be used to evaluate the difference between the predicted current and the optimal reference current. By minimizing the current cost function, the predicted current can be made as close as possible to the optimal reference current, thereby achieving high-precision current control. By traversing all possible inverter switch combinations and calculating the corresponding current cost function values, the switch combination with the minimum current cost function value can be selected as the optimal switch combination. This optimal switch combination will be used to control the on / off of the inverter, thereby driving the permanent magnet motor to run at the desired speed.

[0156] In one embodiment, the calculation formula for the optimal input voltage is:

[0157] ;

[0158] Where , , represents the sampling period of the current, represents the d / q axis current collected at time represents the input vector of the neural network in the current hyper-local model at time represents the optimal reference current at time represents the optimal input voltage of the d / q axis at time

[0159] It should be noted that the optimal reference current at time and can be substituted into the current hyper-local model, and the optimal input voltage of the d / q axis at time can be obtained.

[0160] Furthermore, the voltage cost function is: ;

[0161] Where represents the optimal input voltage of the d / q axis at time represents the i d / q axis voltage corresponding to the

[0162] Furthermore, substitute the optimal input voltage of the d / q axis at time into the voltage cost function and calculate the voltage cost function value.

[0163] By traversing all possible inverter switch combinations, calculating the input voltage corresponding to each switch combination, and calculating the corresponding voltage cost function value, the switch combination with the minimum voltage cost function value can be selected as the optimal switch combination. This optimal switch combination will be used to control the on / off of the inverter, thereby driving the permanent magnet motor to operate at the desired speed.

[0164] In one embodiment, the control effect of the proposed predictive control method based on the self-learning hyperlocal model on the permanent magnet motor is tested, including operating conditions such as starting, constant speed, speed mutation, and torque mutation. Among them, Fig. 3(a) is the speed control effect diagram of the permanent magnet motor, and Fig. 3(b) is the current control effect diagram of the permanent magnet motor. In the starting stage, the permanent magnet motor outputs the maximum torque to accelerate to the given 800 RPM, and at the same time, the q-axis current saturates; in the next stage, the permanent magnet motor operates at a constant speed, and the q-axis current decreases, only providing the load torque; in the next stage, the given speed suddenly changes to 1200 RPM, the q-axis current of the permanent magnet motor immediately saturates, provides the maximum torque to accelerate the permanent magnet motor to the given value, and enters the constant speed operation again; in the next stage, the load of the permanent magnet motor mutates, the q-axis current immediately responds, provides a larger torque, and maintains the constant speed operation of the permanent magnet motor.

[0165] Based on the same concept as the predictive control method of the permanent magnet motor based on the self-learning hyperlocal model, as Figure 4 shown, the present application also provides a predictive control device for a permanent magnet motor based on a self-learning hyperlocal model, and the device includes:

[0166] A construction module 401, configured to construct a self-learning hyperlocal model for the permanent magnet motor. The self-learning hyperlocal model includes a speed hyperlocal model and a current hyperlocal model. The lumped disturbance term and the control input gain in the speed hyperlocal model and the current hyperlocal model are both online self-learning, and the self-learning parts in the speed hyperlocal model and the current hyperlocal model are both approximated by an online neural network or fuzzy logic; the construction module is also configured to construct a historical data queue of the input-output states of the permanent magnet motor, and the historical data queue satisfies the update rule of first-in first-out. The input-output states include speed, current, and voltage;

[0167] An acquisition module 402, configured to acquire the input-output states of the permanent magnet motor at the current moment and update the historical data queue;

[0168] An update module 403, configured to use the data in the updated historical data queue and the first self-learning formula to update the lumped disturbance term and the control input gain in the speed hyperlocal model, and use the data in the updated historical data queue and the second self-learning formula to update the lumped disturbance term and the control input gain in the current hyperlocal model;

[0169] A selection module 404 is configured to substitute a set reference speed into the updated speed super-local model to obtain an optimal reference current, or substitute the set reference speed into the updated speed super-local model to obtain an optimal reference current, and then substitute the optimal reference current into the updated current super-local model to obtain an optimal input voltage. The selection module is further configured to select an optimal switching combination according to the optimal reference current and its corresponding current cost function or according to the optimal input voltage and its corresponding voltage cost function to drive the permanent magnet motor to operate.

[0170] In one embodiment, the speed super-local model obtained by the acquisition module 401 is:

[0171] ;

[0172] Wherein, represents the sampling period of the permanent magnet motor speed, represents the predicted speed of the permanent magnet motor at time represents the speed of the permanent magnet motor collected at time represents the q-axis current collected at time and respectively represent the lumped disturbance term and the control input gain of the speed super-local model; and are both approximated by an online neural network in the speed super-local model, represents the weight of the neural network corresponding to in the speed super-local model at time represents the weight of the neural network corresponding to in the speed super-local model at time represents the input vector of the neural network in the speed super-local model at time;

[0173] The current super-local model obtained by the acquisition module 401 is:

[0174] ;

[0175] Wherein, represents the sampling period of the current, represents the predicted q-axis current at time represents the predicted d-axis current at time represents the d / q-axis current collected at time represents The d / q-axis voltages collected at the moment, and respectively represent the lumped disturbance term and the control input gain of the current hyper-local model;

[0176] Specifically, and are both approximated by the on-line neural network in the current hyper-local model, represents the weights of the neural network corresponding to in the current hyper-local model at the moment, represents the weights of the neural network corresponding to in the current hyper-local model at the moment, represents the input vector of the neural network in the current hyper-local model at the moment;

[0177] The neural networks both have a hidden layer. The input of the neural network is a vector composed of the data in the historical data queue, and the queue length of the historical data queue is selected according to the required control accuracy and the computing speed of the hardware;

[0178] The neural networks can all be replaced by fuzzy logic.

[0179] In one embodiment, the update module 403 calculates a first self-learning formula, which is obtained by minimizing a first error, and the first error is the error between the predicted speed and the actual speed;

[0180] The first self-learning formula is:

[0181] ;

[0182] Wherein, represents the predicted speed of the permanent magnet motor at the moment, represents the actual speed of the permanent magnet motor at the moment, represents the learning rate of the first self-learning formula.

[0183] In one embodiment, the update module 403 calculates a second self-learning formula, which is obtained by minimizing a second error, and the second error is the error between the predicted current and the actual current;

[0184] The second self-learning formula is:

[0185] ;

[0186] Wherein, represents the predicted current of the permanent magnet motor at the moment, represents The actual current of the permanent magnet motor at each moment represents the learning rate of the second self-learning formula.

[0187] In one embodiment, the updating module 403 constructs the following first Lyapunov function for the speed super-local model:

[0188] ;

[0189] where , , , and are respectively and corresponding optimal weights;

[0190] Substitute the first self-learning formula into the first Lyapunov function, and when is satisfied, we get: ;

[0191] where represents the upper bound of the activation function in the neural network, is the upper bound of the fitting error of the neural network under the optimal weights, ;

[0192] According to the approximation theorem, the upper bound of the fitting error decreases as the number of neurons in the hidden layer of the neural network increases. When is satisfied and the design of the first learning rate meets , we get , that is, the final consistency stability is satisfied.

[0193] In one embodiment, the updating module 403 constructs the following second Lyapunov function for the current super-local model:

[0194] ;

[0195] where , ; and are respectively and corresponding optimal weights;

[0196] Substitute the second self-learning formula into the second Lyapunov function, and when is satisfied, we get: ;

[0197] where represents the upper bound of the activation function in the neural network is the upper bound of the fitting error of the neural network under the optimal weights, ;

[0198] According to the approximation theorem, the upper bound of the fitting error decreases as the number of neurons in the hidden layer of the neural network increases. When is satisfied and the design of the second learning rate satisfies , we get , that is, the final consistency stability is satisfied.

[0199] In one embodiment, the selection module 404 will and the reference speed are substituted into the updated speed super-local model to obtain the optimal reference current;

[0200] The calculation formula for the optimal reference current is:

[0201] ;

[0202] where, , , represents the optimal reference current, represents the current amplitude constraint, represents the speed of the permanent magnet motor collected at time represents the input vector of the neural network in the speed super-local model at time represents the sampling period of the permanent magnet motor speed;

[0203] The current cost function is:

[0204] ;

[0205] where, , , represents the input vector of the neural network in the current super-local model at time represents the sampling period of the current, represents the i d / q axis voltage corresponding to the represents d / q axis current collected at time

[0206] In one embodiment, the formula for the selection module 404 to calculate the optimal input voltage is:

[0207] ;

[0208] where, , , represents the sampling period of the current, represents the d / q-axis current sampled at the represents the input vector of the neural network in the current super-local model at the represents the optimal reference current at the represents the optimal input voltage of the d / q-axis at the

[0209] The voltage cost function is: ;

[0210] wherein, represents i the d / q-axis voltage corresponding to the

[0211] Based on the same concept as the permanent magnet motor predictive control method based on the self-learning super-local model, the present application also provides a computer device, which includes a processor and a memory. The memory stores a computer program, and when the processor executes the computer program, it implements the permanent magnet motor predictive control method based on the self-learning super-local model.

[0212] In one embodiment, a computer device is provided, and its internal structure diagram can be as Figure 5 shown. The computer device includes a processor, a memory, a communication interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be implemented through WIFI, a mobile cellular network, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a permanent magnet motor predictive control method based on the self-learning super-local model. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer covered on the display screen, or a button, a trackball, or a touchpad provided on the outer shell of the computer device, or an external keyboard, a touchpad, or a mouse, etc.

[0213] Those skilled in the art can understand, Figure 5The structure shown is only a block diagram of some of the structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. Specifically, the computer device may include more or fewer components than those shown in the figure, or combine some components, or have a different component layout.

[0214] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, 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, it should be considered as the scope described in this specification.

[0215] The above embodiments only represent several implementation manners of this application, and their descriptions are relatively specific and detailed, but they should not be construed as a limitation on the patent scope of this application. It should be noted that for those of ordinary skill in the art, without departing from the concept of this application, several modifications and improvements can still be made, and these all belong to the protection scope of this application. Therefore, the protection scope of this application should be subject to the appended claims.

Claims

1. A predictive control method for a permanent magnet motor based on a self-learning ultra-local model, characterized in that The method includes the following steps: Construct a self - learning hyper - local model for the permanent magnet motor. The self - learning hyper - local model includes a speed hyper - local model and a current hyper - local model. The lumped disturbance terms and control input gains in the speed hyper - local model and the current hyper - local model are both online self - learning, and the self - learning parts in the speed hyper - local model and the current hyper - local model are approximated by online neural networks or fuzzy logic; Construct a historical data queue of the input - output states of the permanent magnet motor. The historical data queue satisfies the first - in - first - out update rule, and the input - output states include speed, current, and voltage; Collect the input - output states of the permanent magnet motor at the current moment and update the historical data queue; Use the data in the updated historical data queue and the first self - learning formula to update the lumped disturbance term and control input gain in the speed hyper - local model, and use the data in the updated historical data queue and the second self - learning formula to update the lumped disturbance term and control input gain in the current hyper - local model; Substitute the set reference speed into the updated speed hyper - local model to obtain the optimal reference current, or substitute the set reference speed into the updated speed hyper - local model to obtain the optimal reference current, and then substitute the optimal reference current into the updated current hyper - local model to obtain the optimal input voltage; Select the optimal switching combination according to the optimal reference current and its corresponding current cost function or according to the optimal input voltage and its corresponding voltage cost function to drive the permanent magnet motor to operate; The speed hyper - local model is: ; represents the sampling period of the rotational speed of the permanent magnet motor, represents the predicted rotational speed of the permanent magnet motor at time represents the rotational speed of the permanent magnet motor collected at time represents the q-axis current collected at time and respectively represent the lumped disturbance term and the control input gain of the rotational speed super-local model; the current super-local model is: ; represents the sampling period of the current, represents the predicted current of the q-axis at time represents the predicted current of the d-axis at time represents the collected d / q-axis current at time represents the collected d / q-axis voltage at time and respectively represent the lumped disturbance term and the control input gain of the current super-local model; the first self-learning formula is: ; represents the predicted speed of the permanent magnet motor at the moment represents the actual speed of the permanent magnet motor at the moment represents the learning rate of the first self - learning formula; the second self - learning formula is: ; represents the predicted current of the permanent magnet motor at the moment represents the actual current of the permanent magnet motor at the moment represents the learning rate of the second self - learning formula.

2. The predictive control method for a permanent magnet motor based on a self-learning ultra-local model according to claim 1, wherein and are both approximated by the online neural network in the rotational speed super-local model, denotes the weights of the neural network corresponding to the rotational speed super-local model at time ; denotes the weights of the neural network corresponding to the rotational speed super-local model at time ; denotes the input vector of the neural network in the rotational speed super-local model at time; and both are approximated by the on-line neural network in the current hyper-local model, denotes the weights of the neural network in the current hyper-local model at time corresponding thereto, denotes the weights of the neural network in the current hyper-local model at time corresponding thereto, denotes the input vector of the neural network in the current hyper-local model at time; All the neural networks have one hidden layer. The input of the neural network is a vector composed of the data in the historical data queue. The queue length of the historical data queue is selected according to the required control accuracy and the computing speed of the hardware; The neural networks can all be replaced by fuzzy logic.

3. The predictive control method of the permanent magnet motor based on the self-learning hyperlocal model according to claim 2, characterized in that, The first self - learning formula is obtained by minimizing the first error, and the first error is the error between the predicted speed and the actual speed.

4. The permanent magnet motor predictive control method based on a self-learning ultra-local model according to claim 2, wherein The second self - learning formula is obtained by minimizing the second error, and the second error is the error between the predicted current and the actual current.

5. The predictive control method of the permanent magnet motor based on the self-learning ultra-local model according to claim 3, characterized in that The stability conditions and parameter selection conditions of the first self - learning formula are as follows: For the speed hyper - local model, construct the following first Lyapunov function: ； Among them, , , , and are respectively and corresponding optimal weights; Substitute the first self-learning formula into the first Lyapunov function, and when satisfied, we get: ; Among them, represents the upper bound of the activation function in the neural network, is the upper bound of the fitting error of the neural network under the optimal weights, ; According to the approximation theorem, the upper bound of the fitting error decreases as the number of neurons in the hidden layer of the neural network increases. When is satisfied and the first learning rate is designed to satisfy when, we get , that is, the final consistency stability is satisfied.

6. The predictive control method of the permanent magnet motor based on the self-learning hyper-local model according to claim 4, characterized in that The stability conditions and parameter selection conditions of the second self - learning formula are as follows: For the current hyper - local model, construct the following second Lyapunov function: ； Among them, , ; and are respectively and the corresponding optimal weights; Substitute the second self-learning formula into the second Lyapunov function, and when it is satisfied, we get: ; Among them, represents the upper bound of the activation function in the neural network, is the upper bound of the fitting error of the neural network under the optimal weights, ; According to the approximation theorem, the upper bound of the fitting error decreases as the number of neurons in the hidden layer of the neural network increases. When is satisfied and the design of the second learning rate satisfies , we get , that is, the final consistency stability is satisfied.

7. The predictive control method for a permanent magnet motor based on a self-learning hyperlocal model according to claim 2, wherein The calculation method of the optimal reference current is as follows: Substitute and the reference speed into the updated speed super-local model to obtain the optimal reference current. The calculation formula of the optimal reference current is: ； Among them, , , represents the optimal reference current, represents the current amplitude constraint, represents the rotational speed of the permanent magnet motor collected at the moment of represents the input vector of the neural network in the rotational speed super local model at the moment of represents the sampling period of the rotational speed of the permanent magnet motor; The current cost function is: ； Among them, , , represents the input vector of the neural network in the current super-local model at the moment, represents the sampling period of the current, represents the d / q axis voltage corresponding to the i th switch combination, represents the d / q axis current collected at the moment.

8. The predictive control method of the permanent magnet motor based on the self-learning hyperlocal model according to claim 2, characterized in that, The calculation formula of the optimal input voltage is: ； Among them, , , represents the sampling period of the current, represents the d / q-axis current collected at the represents the input vector of the neural network in the current super-local model at the represents the optimal reference current at the represents the optimal input voltage of the d / q axis at the The voltage cost function is as follows: ; Among them, represents i the d / q axis voltages corresponding to the 9. A predictive control device for a permanent magnet motor based on a self-learning hyperlocal model, characterized in that, The device executes the permanent magnet motor predictive control method based on the self - learning hyper - local model according to any one of claims 1 to 8. The device includes: A building block for constructing a self-learning hyperlocal model for the permanent magnet motor. The self-learning hyperlocal model includes a speed hyperlocal model and a current hyperlocal model. The lumped disturbance terms and control input gains in the speed hyperlocal model and the current hyperlocal model are both online self-learned. The self-learning parts in the speed hyperlocal model and the current hyperlocal model are both approximated by an online neural network or fuzzy logic. The building block is also used to construct a historical data queue of the input-output states of the permanent magnet motor. The historical data queue satisfies the first-in, first-out update rule. The input-output states include speed, current, and voltage. An acquisition module for acquiring the input-output states of the permanent magnet motor at the current moment and updating the historical data queue. An update module for using the data in the updated historical data queue and a first self-learning formula to update the lumped disturbance term and control input gain in the speed hyperlocal model, and using the data in the updated historical data queue and a second self-learning formula to update the lumped disturbance term and control input gain in the current hyperlocal model. A selection module for substituting a set reference speed into the updated speed hyperlocal model to obtain an optimal reference current, or substituting the set reference speed into the updated speed hyperlocal model to obtain an optimal reference current and substituting the optimal reference current into the updated current hyperlocal model to obtain an optimal input voltage. The selection module is also used to select an optimal switch combination according to the optimal reference current and its corresponding current cost function or according to the optimal input voltage and its corresponding voltage cost function to drive the permanent magnet motor to operate.

10. A computer device, comprising a memory and a processor, the memory storing a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the permanent magnet motor predictive control method based on the self-learning hyperlocal model according to any one of claims 1 to 8.

Images