Permanent magnet motor load adaptive energy-saving optimization method and system based on artificial intelligence

Abstract

The present invention provides a permanent magnet motor load adaptive energy-saving optimization method and system based on artificial intelligence, which relates to the field of motor technology, including extracting features and analyzing working conditions of collected real-time motor load data through a deep neural network model to obtain a load feature vector. Then, a reinforcement learning algorithm based on a deep Q network and a policy gradient network is used to output optimal control parameters including excitation current, switching frequency, and power factor according to the load feature vector, and adjust the motor operating state in real time. Finally, the online learning system updates the model and algorithm parameters in real time through a self-correction module according to the comparison results of the operating data and the energy consumption optimization target to achieve closed-loop control. The present invention can adaptively adjust the motor control parameters according to load changes to achieve the purpose of energy-saving optimization.

Description

Technical Field

[0001] The present invention relates to motor technology, and in particular to an artificial intelligence-based permanent magnet motor load adaptive energy-saving optimization method and system. Background Art

[0002] Permanent magnet motors are widely used in industries such as industry, transportation, and household appliances due to their advantages such as high efficiency, high power density, and reliability. With the continuous rise in energy costs and the enhancement of environmental awareness, energy-saving optimization of permanent magnet motors has become an important research direction. Traditional energy-saving control methods for permanent magnet motors are mainly based on pre-set rules or models, such as PID control, vector control, etc. However, these methods are difficult to adapt to complex working conditions and cannot achieve optimal energy-saving effects. In recent years, artificial intelligence technology, especially deep learning and reinforcement learning, has provided new ideas for load-adaptive energy-saving optimization of permanent magnet motors.

[0003] The prior art has the following defects and deficiencies:

[0004] 1. Traditional control methods are difficult to adapt to complex working conditions: Permanent magnet motors face various load conditions in actual operation, such as sudden load changes, speed fluctuations, etc. Traditional control methods are usually based on pre-set parameters and are difficult to dynamically adjust according to real-time load changes, resulting in increased energy consumption.

[0005] 2. Unable to achieve global optimal energy-saving control: Traditional control methods usually only focus on local optimization, such as current or speed optimization, while ignoring the overall energy consumption state of the motor. This makes it impossible to achieve global optimal energy-saving control, limiting the improvement of energy-saving effects.

[0006] 3. Model parameters are difficult to adjust online: Traditional control methods usually require parameter adjustment based on experience or experimental data, and these parameters are difficult to update in real time during motor operation. This leads to a decrease in control performance and is unable to adapt to motor aging and environmental changes. Summary of the invention

[0007] The embodiments of the present invention provide a permanent magnet motor load adaptive energy-saving optimization method and system based on artificial intelligence, which can solve the problems in the prior art.

[0008] According to a first aspect of the embodiments of the present invention,

[0009] Provides an AI-based permanent magnet motor load adaptive energy-saving optimization method, including:

[0010] The real-time load data of the permanent magnet motor under different working conditions are collected by multiple load sensors, wherein the real-time load data includes speed, torque, current, voltage, temperature, and vibration frequency, and the real-time load data is input into a pre-trained deep neural network model, wherein the deep neural network model adopts a hybrid architecture of a convolutional neural network and a long short-term memory network, and the real-time load data is subjected to feature extraction and working condition analysis by the deep neural network model to obtain a load feature vector of the permanent magnet motor;

[0011] Based on the load characteristic vector, a reinforcement learning algorithm is used to construct a dynamic optimization strategy for the permanent magnet motor, wherein the reinforcement learning algorithm includes a deep Q network and a policy gradient network, wherein the deep Q network is used to evaluate the energy consumption state under different control parameter combinations, and the policy gradient network is used to output the optimal control parameters, wherein the optimal control parameters include the excitation current, switching frequency, and power factor of the permanent magnet motor, and the operating state of the permanent magnet motor is adjusted in real time according to the optimal control parameters;

[0012] An online learning system including a self-correction module is constructed. The online learning system continuously collects the operating data of the permanent magnet motor under the optimal control parameters, and compares the operating data with a preset energy consumption optimization target. When the operating data deviates from the energy consumption optimization target, the network parameters of the deep neural network model and the strategy parameters of the reinforcement learning algorithm are updated in real time through the self-correction module to realize closed-loop control of adaptive energy-saving optimization of the permanent magnet motor load.

[0013] The real-time load data is input into a pre-trained deep neural network model, wherein the deep neural network model adopts a hybrid architecture of a convolutional neural network and a long short-term memory network, and the real-time load data is subjected to feature extraction and working condition analysis through the deep neural network model to obtain a load feature vector of the permanent magnet motor, including:

[0014] The real-time load data is mapped to the interval of zero to one by the maximum and minimum value standardization method to obtain standardized load data; the standardized load data is input into the deep neural network model for feature extraction, the deep neural network model includes a convolutional neural network module and a long short-term memory network module, wherein the convolutional neural network module includes three convolutional layers connected in series, the convolution kernel sizes of the three convolutional layers connected in series are five times five, three times three, and three times three, respectively, the output of each convolutional layer is connected to a batch normalization layer and a rectified linear unit activation function layer, the long short-term memory network module includes a two-layer bidirectional long short-term memory network structure, and the number of hidden layer neurons of the two-layer bidirectional long short-term memory network structure is one hundred and twenty-eight;

[0015] The spatial features output by the convolutional neural network module and the temporal features output by the long short-term memory network module are input into an attention fusion module, the attention weights of the spatial features and the temporal features are calculated by the attention fusion module, and the spatial features and the temporal features are weightedly fused according to the attention weights to obtain a 256-dimensional load feature vector, wherein the load feature vector includes load fluctuation features, temporal correlation features, and operating state features;

[0016] The load feature vector is input into a fully connected mapping layer, and a working condition classification probability distribution is output through the fully connected mapping layer. The load working condition type of the permanent magnet motor is determined according to the working condition classification probability distribution. The load working condition type includes a normal load working condition, a light load working condition, a heavy load working condition and a fluctuating load working condition.

[0017] Based on the load characteristic vector, a dynamic optimization strategy for the permanent magnet motor is constructed by using a reinforcement learning algorithm. The reinforcement learning algorithm includes a deep Q network and a policy gradient network. The energy consumption state under different control parameter combinations is evaluated by the deep Q network. The optimal control parameters output by the policy gradient network include:

[0018] A Markov decision environment for permanent magnet motor control is constructed, the load characteristic vector of the permanent magnet motor is used as the state space, the control parameter combination is used as the action space, and a reward function is constructed based on energy consumption evaluation. The reward function is obtained by weighted combination of current energy consumption, deviation between actual output power and target power, and control quantity change rate;

[0019] A control strategy of a permanent magnet motor is evaluated by using a deep Q network structure, wherein the deep Q network structure comprises four fully connected layers, and a rectified linear unit activation function layer is arranged between the four fully connected layers. Data of a state space and an action space are input into the deep Q network structure, an objective function is constructed based on a temporal difference error, and network parameters of the deep Q network structure are optimized by minimizing the objective function.

[0020] Constructing a policy gradient network based on the Actor-Cr itic architecture, inputting the load feature vector of the state space into the policy gradient network, outputting a probability distribution of control parameters that obeys a Gaussian distribution through the policy gradient network, wherein the probability distribution includes a mean parameter and a variance parameter, calculating the ratio of the probability of a new policy to the probability of an old policy, estimating and constructing a clipping objective function based on the ratio and the advantage function, and optimizing the policy parameters of the policy gradient network using a proximal policy optimization algorithm;

[0021] Establish a dual-network collaborative optimization mechanism, use the network parameters of the deep Q network structure as the target network parameters, use the policy parameters of the policy gradient network as the online network parameters, use a soft update method to dynamically adjust the target network parameters and the online network parameters, and control the rate at which the target network parameters are transferred to the online network parameters through a soft update coefficient;

[0022] A priority experience replay pool is constructed, and the temporal difference error, the state space data, the action space data and the reward function data are stored in the priority experience replay pool; a priority sampling probability of the experience data is set based on the temporal difference error; a power function relationship between the priority sampling probability and the priority value is used as a sampling basis for the experience data; training data is selected from the priority experience replay pool according to the sampling basis; and the deep Q network structure and the policy gradient network are iteratively trained using the training data.

[0023] Constructing a policy gradient network based on the Actor-Cr itic architecture, inputting the load feature vector of the state space into the policy gradient network, outputting a control parameter probability distribution that obeys a Gaussian distribution through the policy gradient network, wherein the probability distribution includes a mean parameter and a variance parameter, calculating the ratio of the new policy probability to the old policy probability, constructing a clipping objective function based on the ratio and the advantage function estimation, and optimizing the policy parameters of the policy gradient network using a proximal policy optimization algorithm, including:

[0024] Construct an Actor network and a Cr it ic network, wherein the Actor network adopts a four-layer feedforward neural network structure, the output layer of the Actor network generates the mean of the control parameters through a hyperbolic tangent function, generates the variance of the control parameters through a soft positive function, takes the mean as the center position parameter of the Gaussian distribution, takes the variance as the discrete degree parameter of the Gaussian distribution, and constructs a probability distribution of the control parameters that obeys the Gaussian distribution;

[0025] Based on the control parameter probability distribution, the control parameter at the current moment is obtained, the control parameter is applied to the control object to obtain an immediate reward signal, an advantage function is calculated using a generalized advantage estimation method, the immediate reward signal, the discount factor, the state value function and the generalized advantage estimation parameter are substituted into a temporal difference error calculation formula to obtain a temporal difference error, and an advantage function estimation value is calculated according to the temporal difference error;

[0026] Calculating the new strategy probability by using the Actor network, calculating the old strategy probability by using the Actor network of the previous iteration, calculating the ratio of the new strategy probability to the old strategy probability, multiplying the ratio by the advantage function estimate to obtain a strategy objective function, and limiting the strategy objective function within a preset clipping range to obtain a clipping objective function;

[0027] An adaptive learning rate adjustment mechanism is constructed, and a first-order momentum estimate and a second-order momentum estimate of the policy gradient are calculated based on the clipping objective function. The first-order momentum estimate and the second-order momentum estimate are substituted into an adaptive learning rate calculation formula to obtain an adaptive learning rate, and the policy parameters of the Actor network are iteratively optimized using the adaptive learning rate.

[0028] The energy consumption state under different control parameter combinations is evaluated by the deep Q network, and the optimal control parameters are output by the policy gradient network. The optimal control parameters include the excitation current, switching frequency, and power factor of the permanent magnet motor. The operation state of the permanent magnet motor is adjusted in real time according to the optimal control parameters, including:

[0029] The speed, current and torque data of the permanent magnet motor are input into the deep Q network, the copper loss, iron loss and switching loss of the permanent magnet motor are calculated based on the deep Q network, and the copper loss, iron loss and switching loss are superimposed to obtain a total energy consumption state;

[0030] The speed, current and torque data of the permanent magnet motor are input into the policy gradient network, and the policy gradient network outputs a control parameter probability distribution that obeys a Gaussian distribution, wherein the control parameter probability distribution includes a mean vector and a variance vector, and the mean vector includes the excitation current, switching frequency and power factor of the permanent magnet motor;

[0031] Constructing an energy consumption evaluation function based on the total energy consumption state, combining the energy consumption evaluation function with a discount factor to construct a Q value function, evaluating the energy consumption state of different control parameter combinations through the Q value function, and inputting the evaluation result into the policy gradient network;

[0032] The system energy consumption index, dynamic response index and stability index are weighted and combined to obtain the comprehensive performance index, and based on the change trend of the comprehensive performance index, the weighting coefficient is adaptively adjusted by exponential decay;

[0033] The adjustment result of the weighted coefficient is fed back to the policy gradient network, and the control parameter probability distribution is recalculated through the policy gradient network, and the updated excitation current, switching frequency and power factor are output, and the operating state of the permanent magnet motor is adjusted in real time according to the updated control parameters.

[0034] The network parameters of the deep neural network model and the strategy parameters of the reinforcement learning algorithm are updated in real time by the self-correction module to realize closed-loop control of adaptive energy-saving optimization of the permanent magnet motor load, including:

[0035] Calculate the standard deviation of each characteristic quantity in the load characteristic vector, calculate the sensitivity of each characteristic quantity to the control output based on the standard deviation, and use the sensitivity as the characteristic weight coefficient; multiply the load characteristic vector by the characteristic weight coefficient to obtain a weighted characteristic vector, input the weighted characteristic vector into a deep neural network model, predict the operating state of the permanent magnet motor through the deep neural network model, and construct a loss function by combining the error between the predicted state and the actual state, the regularization term of the network parameter, and the smoothing term of the control quantity;

[0036] Calculate the gradient of the loss function with respect to the network weight and the network bias, update the network weight and the network bias by back propagation, use the updated network output as the state input of the reinforcement learning strategy network, and control the action through the reinforcement learning strategy network output;

[0037] Applying the control action to the permanent magnet motor system to obtain an immediate reward, constructing a performance objective function including system energy consumption, dynamic response and control stability, calculating the gradient of the performance objective function to the policy parameters, and adaptively adjusting the learning step size according to the ratio of the current policy gradient to the historical policy gradient;

[0038] The strategy parameters are optimized by using an adaptive learning step size, a state feedback controller is designed based on the optimized strategy parameters, a state feedback gain matrix is ​​multiplied by the system state to obtain a feedback control quantity, and the feedback control quantity is superimposed with a pre-acquired feedforward compensation term to obtain a control output;

[0039] The system state error caused by the control output is calculated, a Lyapunov function is constructed and its derivative is obtained, an adaptive law is designed based on the Lyapunov stability theory, the state feedback gain matrix is ​​adjusted online, and an integral constraint is imposed on the adjustment amount to ensure parameter convergence.

[0040] The strategy parameters are optimized by using an adaptive learning step size, a state feedback controller is designed based on the optimized strategy parameters, the state feedback gain matrix is ​​multiplied by the system state to obtain a feedback control quantity, and the feedback control quantity is superimposed with the pre-acquired feedforward compensation term to obtain a control output including:

[0041] Calculate the ratio of the current policy gradient to the previous policy gradient, multiply the ratio by a preset attenuation coefficient and take the negative exponent, multiply the negative exponent by the initial learning step length to obtain the current adaptive learning step length, and update the policy parameters using the product of the current adaptive learning step length and the current policy gradient;

[0042] Constructing a state feedback controller, calculating the Rickati equation to obtain a state feedback matrix, multiplying the inverse matrix of the control weight matrix with the transpose of the input matrix and the state feedback matrix to obtain a state feedback gain matrix, and multiplying the state feedback gain matrix with the system state vector to obtain a feedback control component;

[0043] Constructing a feedforward compensation controller, calculating the inverse matrix of the product of the output matrix and the input matrix, multiplying the inverse matrix with the product of the output matrix and the system state matrix to obtain a feedforward compensation matrix, and multiplying the feedforward compensation matrix with a reference input to obtain a feedforward control component;

[0044] The feedback control component and the feedforward control component are superimposed to obtain a total control output, a quadratic function is constructed by multiplying the total control output and a control weight matrix, and the quadratic function is integrated within a time interval to obtain a control performance evaluation index;

[0045] Calculate a new policy gradient based on the control performance evaluation index, substitute the new policy gradient into the adaptive learning step calculation formula to update the learning step, and use the updated learning step to optimize and adjust the state feedback gain matrix and the feedforward compensation matrix;

[0046] The optimized state feedback gain matrix and feedforward compensation matrix are re-substituted into the controller to generate a new control output to act on the system, forming a closed-loop adaptive control.

[0047] A second aspect of an embodiment of the present invention provides an artificial intelligence-based permanent magnet motor load adaptive energy-saving optimization system, comprising:

[0048] The first unit is used to collect real-time load data of the permanent magnet motor under different working conditions through multiple load sensors, wherein the real-time load data includes speed, torque, current, voltage, temperature, and vibration frequency, and input the real-time load data into a pre-trained deep neural network model, wherein the deep neural network model adopts a hybrid architecture of a convolutional neural network and a long short-term memory network, and performs feature extraction and working condition analysis on the real-time load data through the deep neural network model to obtain a load feature vector of the permanent magnet motor;

[0049] The second unit is used to construct a dynamic optimization strategy of the permanent magnet motor based on the load feature vector by using a reinforcement learning algorithm, wherein the reinforcement learning algorithm includes a deep Q network and a policy gradient network, and the energy consumption state under different control parameter combinations is evaluated by the deep Q network, and the optimal control parameters are output by the policy gradient network, wherein the optimal control parameters include the excitation current, switching frequency, and power factor of the permanent magnet motor, and the operating state of the permanent magnet motor is adjusted in real time according to the optimal control parameters;

[0050] The third unit is used to construct an online learning system including a self-correction module. The online learning system continuously collects the operating data of the permanent magnet motor under the optimal control parameters, and compares the operating data with the preset energy consumption optimization target. When the operating data deviates from the energy consumption optimization target, the network parameters of the deep neural network model and the strategy parameters of the reinforcement learning algorithm are updated in real time through the self-correction module to realize closed-loop control of adaptive energy-saving optimization of the permanent magnet motor load.

[0051] A third aspect of the embodiments of the present invention

[0052] An electronic device is provided, comprising:

[0053] processor;

[0054] a memory for storing processor-executable instructions;

[0055] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0056] A fourth aspect of the embodiments of the present invention is:

[0057] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0058] The beneficial effects of this application are as follows:

[0059] 1. Improve the operating efficiency of permanent magnet motors: Through the deep neural network model, the load conditions and characteristics are accurately analyzed, and the control parameters are dynamically optimized in combination with the reinforcement learning algorithm, which can realize the efficient operation of permanent magnet motors under different loads, thereby reducing energy consumption.

[0060] 2. Adapt to different working conditions: Based on real-time load data and online learning system, this method can adaptively adjust the control strategy to cope with complex and changing working conditions, ensuring energy-saving optimization in various situations.

[0061] 3. Achieve closed-loop control and continuous optimization: The introduction of online learning system and self-correction module enables the system to self-adjust and optimize according to actual operation data, continuously improve energy-saving effects, and maintain long-term stable operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 A schematic diagram of a flow chart of a method for adaptive energy-saving optimization of a permanent magnet motor load based on artificial intelligence according to an embodiment of the present invention;

[0063] Figure 2It is a structural schematic diagram of an artificial intelligence-based permanent magnet motor load adaptive energy-saving optimization system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0065] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0066] Figure 1 FIG. 1 is a flow chart of a method for adaptive energy-saving optimization of permanent magnet motor load based on artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0067] S101. Collect real-time load data of the permanent magnet motor under different working conditions through multiple load sensors, the real-time load data including speed, torque, current, voltage, temperature, and vibration frequency, input the real-time load data into a pre-trained deep neural network model, the deep neural network model adopts a hybrid architecture of a convolutional neural network and a long short-term memory network, and perform feature extraction and working condition analysis on the real-time load data through the deep neural network model to obtain a load feature vector of the permanent magnet motor;

[0068] S102. Based on the load characteristic vector, a dynamic optimization strategy of the permanent magnet motor is constructed by using a reinforcement learning algorithm, wherein the reinforcement learning algorithm includes a deep Q network and a policy gradient network, wherein the energy consumption state under different control parameter combinations is evaluated by the deep Q network, and the optimal control parameters are output by the policy gradient network, wherein the optimal control parameters include the excitation current, switching frequency, and power factor of the permanent magnet motor, and the operating state of the permanent magnet motor is adjusted in real time according to the optimal control parameters;

[0069] S103. Construct an online learning system including a self-correction module, wherein the online learning system continuously collects the operating data of the permanent magnet motor under the optimal control parameters, and compares the operating data with the preset energy consumption optimization target. When the operating data deviates from the energy consumption optimization target, the network parameters of the deep neural network model and the strategy parameters of the reinforcement learning algorithm are updated in real time through the self-correction module to realize closed-loop control of adaptive energy-saving optimization of the permanent magnet motor load.

[0070] In an optional implementation, the real-time load data is input into a pre-trained deep neural network model, the deep neural network model adopts a hybrid architecture of a convolutional neural network and a long short-term memory network, and the real-time load data is subjected to feature extraction and operating condition analysis by the deep neural network model to obtain a load feature vector of the permanent magnet motor, including:

[0071] The real-time load data is mapped to the interval of zero to one by the maximum and minimum value standardization method to obtain standardized load data; the standardized load data is input into the deep neural network model for feature extraction, the deep neural network model includes a convolutional neural network module and a long short-term memory network module, wherein the convolutional neural network module includes three convolutional layers connected in series, the convolution kernel sizes of the three convolutional layers connected in series are five times five, three times three, and three times three, respectively, the output of each convolutional layer is connected to a batch normalization layer and a rectified linear unit activation function layer, the long short-term memory network module includes a two-layer bidirectional long short-term memory network structure, and the number of hidden layer neurons of the two-layer bidirectional long short-term memory network structure is one hundred and twenty-eight;

[0072] The spatial features output by the convolutional neural network module and the temporal features output by the long short-term memory network module are input into an attention fusion module, the attention weights of the spatial features and the temporal features are calculated by the attention fusion module, and the spatial features and the temporal features are weightedly fused according to the attention weights to obtain a 256-dimensional load feature vector, wherein the load feature vector includes load fluctuation features, temporal correlation features, and operating state features;

[0073] The load feature vector is input into a fully connected mapping layer, and a working condition classification probability distribution is output through the fully connected mapping layer. The load working condition type of the permanent magnet motor is determined according to the working condition classification probability distribution. The load working condition type includes a normal load working condition, a light load working condition, a heavy load working condition and a fluctuating load working condition.

[0074] The invention discloses an intelligent identification method for load conditions of a permanent magnet motor, and relates to the field of motor control technology. The invention aims to provide a method for identifying load conditions of a permanent magnet motor based on a deep neural network, which can accurately identify different load condition types of the permanent magnet motor.

[0075] Collect the load data of the permanent magnet motor in real time. The load data can be parameters such as the motor's current, voltage, speed, torque, etc., or a combination of multiple parameters. For example, the motor's current, voltage, and speed data are collected every 0.01 seconds to form a set of load data.

[0076] Preprocess the collected real-time load data. Use the maximum and minimum value normalization method to map the load data to the range of zero to one to eliminate the influence of different dimensions and improve the training efficiency and generalization ability of the model. For example, assuming that the collected current data ranges from 10A to 50A, 10A is mapped to 0, 50A is mapped to 1, and other current values ​​are linearly mapped to between 0 and 1 in proportion.

[0077] The preprocessed normalized load data is input into the pre-trained deep neural network model for feature extraction. The deep neural network model adopts a hybrid architecture of convolutional neural network and long short-term memory network.

[0078] The convolutional neural network module is used to extract the spatial features of the load data. The module contains three convolutional layers in series, with convolution kernel sizes of five by five, three by three, and three by three, respectively. The output of each convolutional layer is connected to a batch normalization layer and a rectified linear unit activation function layer to speed up the convergence of the model and improve the nonlinear expression ability of the model. For example, the first convolutional layer uses a five by five convolution kernel to perform convolution operations on the input data to extract local features; then the convolution result is input to the batch normalization layer for normalization, and then input to the rectified linear unit activation function layer for nonlinear transformation.

[0079] The LSTM module is used to extract the time series features of the load data. The module contains a two-layer bidirectional LSTM network structure with one hundred and twenty-eight neurons in the hidden layer. The bidirectional LSTM network can consider both past and future information and better capture the time series dependencies of the load data. For example, the first layer of the bidirectional LSTM network receives the output of the convolutional neural network module and processes it in the time dimension to extract the time series features of the load data.

[0080] The spatial features output by the convolutional neural network module and the temporal features output by the long short-term memory network module are input into the attention fusion module. The attention fusion module calculates the attention weights of the spatial features and the temporal features, and weightedly fuses the spatial features and the temporal features according to the attention weights to obtain a 256-dimensional load feature vector. The load feature vector contains load fluctuation features, temporal correlation features, and operating state features. For example, if the spatial features are more important for judging the load condition at the current moment, the attention mechanism will give the spatial features a greater weight.

[0081] The load feature vector is input to the fully connected mapping layer. The fully connected mapping layer outputs the operating condition classification probability distribution. The load condition type of the permanent magnet motor is determined according to the operating condition classification probability distribution. The load condition types include normal load condition, light load condition, heavy load condition and fluctuating load condition. For example, if the probability distribution output by the fully connected mapping layer is [0.1, 0.2, 0.6, 0.1], the permanent magnet motor is considered to be in a heavy load condition.

[0082] The beneficial effects of this method are as follows:

[0083] 1. Improve recognition accuracy: The deep neural network model can effectively extract the deep features of load data. Combined with the attention mechanism, it can better integrate spatial features and temporal features, thereby improving the accuracy of load condition recognition.

[0084] 2. Enhanced robustness: The maximum and minimum normalization method can eliminate the influence of different dimensions and improve the robustness of the model. The batch normalization layer can reduce the internal covariate shift and further enhance the robustness of the model.

[0085] 3. Intelligentization: This method can automatically learn the characteristics of load data and classify the working conditions according to the learned characteristics to achieve intelligent identification of permanent magnet motor load conditions.

[0086] In an optional implementation, based on the load feature vector, a reinforcement learning algorithm is used to construct a dynamic optimization strategy for the permanent magnet motor, the reinforcement learning algorithm includes a deep Q network and a policy gradient network, the deep Q network is used to evaluate the energy consumption state under different control parameter combinations, and the policy gradient network outputs the optimal control parameters including:

[0087] A Markov decision environment for permanent magnet motor control is constructed, the load characteristic vector of the permanent magnet motor is used as the state space, the control parameter combination is used as the action space, and a reward function is constructed based on energy consumption evaluation. The reward function is obtained by weighted combination of current energy consumption, deviation between actual output power and target power, and control quantity change rate;

[0088] A control strategy of a permanent magnet motor is evaluated by using a deep Q network structure, wherein the deep Q network structure comprises four fully connected layers, and a rectified linear unit activation function layer is arranged between the four fully connected layers. Data of a state space and an action space are input into the deep Q network structure, an objective function is constructed based on a temporal difference error, and network parameters of the deep Q network structure are optimized by minimizing the objective function.

[0089] Constructing a policy gradient network based on the Actor-Cr it ic architecture, inputting the load feature vector of the state space into the policy gradient network, outputting a control parameter probability distribution that obeys a Gaussian distribution through the policy gradient network, wherein the probability distribution includes a mean parameter and a variance parameter, calculating the ratio of the new policy probability to the old policy probability, constructing a clipping objective function based on the ratio and the advantage function estimation, and optimizing the policy parameters of the policy gradient network using a proximal policy optimization algorithm;

[0090] Establish a dual-network collaborative optimization mechanism, use the network parameters of the deep Q network structure as the target network parameters, use the policy parameters of the policy gradient network as the online network parameters, use a soft update method to dynamically adjust the target network parameters and the online network parameters, and control the rate at which the target network parameters are transferred to the online network parameters through a soft update coefficient;

[0091] A priority experience replay pool is constructed, and the temporal difference error, the state space data, the action space data and the reward function data are stored in the priority experience replay pool; a priority sampling probability of the experience data is set based on the temporal difference error; a power function relationship between the priority sampling probability and the priority value is used as a sampling basis for the experience data; training data is selected from the priority experience replay pool according to the sampling basis; and the deep Q network structure and the policy gradient network are iteratively trained using the training data.

[0092] The dynamic optimization control method of permanent magnet motor is based on the load characteristic vector and uses deep reinforcement learning algorithm to build a dynamic optimization strategy for permanent magnet motor to improve the motor operation efficiency. The method includes the following steps:

[0093] First, a Markov decision environment for permanent magnet motor control is constructed. The multidimensional feature vector that characterizes the motor load condition, such as load torque, speed, current, etc., is used as the state space. The control parameters of the motor, such as the amplitude and frequency of voltage and current, are combined into an action space. In order to evaluate the pros and cons of the control strategy, a reward function is constructed. The reward function comprehensively considers the current energy consumption, the deviation between the actual output power and the target power, and the rate of change of the control quantity, and is obtained by weighted combination of these three factors. For example, a larger weight can be set for energy consumption to prioritize reducing energy consumption; a medium weight can be set for power deviation to ensure that the output power is as close to the target value as possible; a smaller weight can be set for the rate of change of the control quantity to avoid excessive control action.

[0094] Secondly, a deep Q network is constructed to evaluate the control strategy. The network contains four fully connected layers, and the rectified linear unit activation function is used between layers to enhance the nonlinear expression ability of the network. The load feature vector of the state space and the control parameter combination of the action space are input into the deep Q network, and the network outputs the value estimate of the corresponding state-action pair. During the training process, the time difference error is used as the objective function, and the network parameters are optimized by minimizing the objective function to continuously improve the evaluation accuracy of the network. For example, the back propagation algorithm and the Adam optimizer can be used to update the network parameters.

[0095] Next, a policy gradient network based on the Actor-Critic architecture is constructed. The load feature vector of the state space is input into the policy gradient network, and the network outputs the probability distribution of the control parameters, which obeys the Gaussian distribution and contains two parameters: mean and variance. By calculating the ratio of the probability of the new policy to the probability of the old policy and combining it with the advantage function estimation, a clipping objective function is constructed. The proximal policy optimization algorithm is used to optimize the policy parameters of the network so that the network can output a better control policy. For example, the clipping range can be set to 0.2 to limit the policy update amplitude.

[0096] Then, a dual-network collaborative optimization mechanism is established. The network parameters of the deep Q network are used as the target network parameters, and the policy parameters of the policy gradient network are used as the online network parameters. The target network parameters are slowly moved closer to the online network parameters using a soft update method to maintain the stability of the target network. The soft update coefficient controls the rate at which the target network parameters are transferred to the online network parameters. For example, the soft update coefficient can be set to 0.005, which means that each time the target network parameters are updated, they will be moved 0.5% toward the online network parameters.

[0097] Finally, a priority experience replay pool is constructed. The temporal difference error, state space data, action space data, and reward function data are stored in the experience replay pool. The priority sampling probability of the experience data is set according to the size of the temporal difference error. The larger the priority value, the higher the sampling probability. The power function relationship between the priority sampling probability and the priority value is used as the basis for sampling the experience data. According to the sampling basis, training data is selected from the experience replay pool, and the deep Q network and policy gradient network are iteratively trained using the training data. For example, the power function exponent can be set to 0.6 to balance exploration and exploitation.

[0098] Example: Assume that the load feature vector of a permanent magnet motor contains two dimensions: load torque and speed, and the control parameter is voltage amplitude. In the initial state, the load torque is 10Nm and the speed is 1000rpm. After the policy gradient network calculation, the output voltage amplitude is 100V. After the motor runs for a period of time, the load torque becomes 12Nm and the speed becomes 1100rpm. At this time, the energy consumption is 100J, the actual output power is 2W away from the target power, and the control amount change rate is 1V / s. According to the preset weights, the reward value is calculated to be -102.2. These data are stored in the experience replay pool for subsequent network training.

[0099] Beneficial effects:

[0100] 1. Improve control efficiency: This method can dynamically adjust the control parameters according to the load characteristic vector to achieve refined control of the permanent magnet motor, thereby improving the motor's operating efficiency and reducing energy consumption.

[0101] 2. Enhanced robustness: Through the training of reinforcement learning algorithms, the control strategy can adapt to different load conditions, enhancing the robustness and stability of the control system.

[0102] 3. Simplified control design: This method does not rely on complex mathematical models, but only needs to learn control strategies in a data-driven manner, which simplifies the design and debugging process of the control system.

[0103] In an optional implementation, a policy gradient network based on an Actor-Cr it ic architecture is constructed, the load feature vector of the state space is input into the policy gradient network, the policy gradient network outputs a control parameter probability distribution that obeys a Gaussian distribution, the probability distribution includes a mean parameter and a variance parameter, the ratio of the new policy probability to the old policy probability is calculated, a clipping objective function is constructed based on the ratio and the advantage function estimation, and the policy parameters of the policy gradient network are optimized using a proximal policy optimization algorithm, including:

[0104] Construct an Actor network and a Cr it ic network, wherein the Actor network adopts a four-layer feedforward neural network structure, the output layer of the Actor network generates the mean of the control parameters through a hyperbolic tangent function, generates the variance of the control parameters through a soft positive function, takes the mean as the center position parameter of the Gaussian distribution, takes the variance as the discrete degree parameter of the Gaussian distribution, and constructs a probability distribution of the control parameters that obeys the Gaussian distribution;

[0105] Based on the control parameter probability distribution, the control parameter at the current moment is obtained, the control parameter is applied to the control object to obtain an immediate reward signal, an advantage function is calculated using a generalized advantage estimation method, the immediate reward signal, the discount factor, the state value function and the generalized advantage estimation parameter are substituted into a temporal difference error calculation formula to obtain a temporal difference error, and an advantage function estimation value is calculated according to the temporal difference error;

[0106] Calculating the new strategy probability by using the Actor network, calculating the old strategy probability by using the Actor network of the previous iteration, calculating the ratio of the new strategy probability to the old strategy probability, multiplying the ratio by the advantage function estimate to obtain a strategy objective function, and limiting the strategy objective function within a preset clipping range to obtain a clipping objective function;

[0107] An adaptive learning rate adjustment mechanism is constructed, and a first-order momentum estimate and a second-order momentum estimate of the policy gradient are calculated based on the clipping objective function. The first-order momentum estimate and the second-order momentum estimate are substituted into an adaptive learning rate calculation formula to obtain an adaptive learning rate, and the policy parameters of the Actor network are iteratively optimized using the adaptive learning rate.

[0108] A control parameter optimization method based on the Actor-Cr it ic architecture and the proximal policy optimization algorithm learns the optimal control strategy of the control object through the policy gradient network. The core of this method is to use the state feature vector as input, output the probability distribution of the control parameters through the policy gradient network, and optimize the network parameters according to the reward signal and advantage function, and finally obtain the optimal control strategy.

[0109] First, construct the Actor network and the Critic network. The Actor network adopts a four-layer feedforward neural network structure, and the input is the load feature vector of the state space. The output layer is divided into two parts, one part generates the mean of the control parameter through the hyperbolic tangent function, and the other part generates the variance of the control parameter through the soft positive function. The mean and variance together constitute the probability distribution of the control parameter that obeys the Gaussian distribution, where the mean is the center position parameter of the Gaussian distribution and the variance is the discrete degree parameter of the Gaussian distribution. For example, if the control parameter is the motor speed, the mean of the speed output by the Actor network is 1000 rpm and the variance is 100, which means that the speed obeys a normal distribution with a mean of 1000 and a standard deviation of 10. The Critic network is used to evaluate the state value function. Its structure can be similar to that of the Actor network, but the output is a scalar value, indicating the value of the current state.

[0110] Then, the control parameters at the current moment are sampled according to the probability distribution output by the Actor network. The control parameters are applied to the control object, such as setting them as the target speed of the motor, observing the state changes of the control object and obtaining an immediate reward signal. For example, if the goal is to stabilize the motor at 1000 rpm, the closer the actual speed is to 1000, the greater the reward signal.

[0111] Next, calculate the advantage function. The generalized advantage estimation method requires the use of immediate reward signals, discount factors, state value functions, and generalized advantage estimation parameters. The temporal difference error is calculated by the difference between the immediate reward, discount factor, and the state value function at the next moment and the state value function at the current moment. The advantage function estimate is calculated by the weighted sum of the temporal difference error and the generalized advantage estimation parameters. For example, assuming that the state value function at the current moment is 10, the state value function at the next moment is 12, the immediate reward is 2, the discount factor is 0.9, and the generalized advantage estimation parameter is 0.8, then the temporal difference error is 2+0.9*12-10=2.8, and the advantage function estimate is 2.8*0.8=2.24.

[0112] Subsequently, the new strategy probability is calculated using the Actor network of the current iteration, and the old strategy probability is calculated using the Actor network of the previous iteration. The ratio of the new strategy probability to the old strategy probability is multiplied by the advantage function estimate to obtain the strategy objective function. In order to avoid excessive strategy updates, the strategy objective function is limited to a preset clipping range, such as [0.8, 1.2], to obtain the clipped objective function. Assuming that the ratio of the new and old strategy probabilities is 1.1 and the advantage function estimate is 2.24, the strategy objective function is 2.464. After clipping, the clipped objective function is 2.24*1.2=2.688.

[0113] Finally, an adaptive learning rate adjustment mechanism is constructed. The first-order momentum estimate and the second-order momentum estimate of the policy gradient are calculated based on the clipping objective function. The first-order momentum estimate and the second-order momentum estimate are substituted into the adaptive learning rate calculation formula to obtain the adaptive learning rate. The policy parameters of the Actor network are iteratively optimized using the adaptive learning rate, for example, using the Adam optimizer. The Critic network parameters are also updated using a similar method.

[0114] This method has three beneficial effects:

[0115] 1. Improve control effect: Learning control strategies through policy gradient networks can better adapt to complex dynamic environments and obtain better control effects, such as making the motor speed more stable and reaching the target speed more quickly.

[0116] 2. Enhanced robustness: Clipping the objective function can limit the amplitude of policy updates, avoid violent fluctuations, improve the robustness of the algorithm, and enable stable operation even in the presence of noise or interference.

[0117] 3. Accelerate learning convergence: The adaptive learning rate adjustment mechanism can dynamically adjust the learning rate according to the learning situation, accelerate learning convergence, reduce training time, and learn the optimal control strategy more efficiently.

[0118] In an optional implementation, the energy consumption state under different control parameter combinations is evaluated by the deep Q network, and the optimal control parameters are output by the policy gradient network, wherein the optimal control parameters include the excitation current, switching frequency, and power factor of the permanent magnet motor. The operation state of the permanent magnet motor is adjusted in real time according to the optimal control parameters, including:

[0119] The speed, current and torque data of the permanent magnet motor are input into the deep Q network, the copper loss, iron loss and switching loss of the permanent magnet motor are calculated based on the deep Q network, and the copper loss, iron loss and switching loss are superimposed to obtain a total energy consumption state;

[0120] The speed, current and torque data of the permanent magnet motor are input into the policy gradient network, and the policy gradient network outputs a control parameter probability distribution that obeys a Gaussian distribution, wherein the control parameter probability distribution includes a mean vector and a variance vector, and the mean vector includes the excitation current, switching frequency and power factor of the permanent magnet motor;

[0121] Constructing an energy consumption evaluation function based on the total energy consumption state, combining the energy consumption evaluation function with a discount factor to construct a Q value function, evaluating the energy consumption state of different control parameter combinations through the Q value function, and inputting the evaluation result into the policy gradient network;

[0122] The system energy consumption index, dynamic response index and stability index are weighted and combined to obtain the comprehensive performance index, and based on the change trend of the comprehensive performance index, the weighting coefficient is adaptively adjusted by exponential decay;

[0123] The adjustment result of the weighted coefficient is fed back to the policy gradient network, and the control parameter probability distribution is recalculated through the policy gradient network, and the updated excitation current, switching frequency and power factor are output, and the operating state of the permanent magnet motor is adjusted in real time according to the updated control parameters.

[0124] First, build a permanent magnet motor experimental platform. The platform includes a permanent magnet motor, a frequency converter, a power analyzer and a data acquisition system. The permanent magnet motor is used to simulate the actual motor operating conditions; the frequency converter is used to adjust the motor's excitation current and switching frequency; the power analyzer is used to measure the motor's input power, current, torque and other parameters; the data acquisition system is used to collect the motor's operating data and transmit it to the computer for processing. For example, a permanent magnet synchronous motor with a rated power of 10kW and a rated speed of 1450rpm is selected as the experimental object.

[0125] Then, a deep Q network and a policy gradient network are constructed. The deep Q network is used to evaluate the energy consumption state of the motor under different control parameter combinations. Its input is the speed, current and torque of the motor, and its output is the corresponding Q value. The policy gradient network is used to output the optimal control parameters. Its input is the speed, current, torque of the motor and the evaluation result of the deep Q network, and its output is the probability distribution of the control parameters that obey the Gaussian distribution, including the mean vector and the variance vector. The mean vector includes the excitation current, switching frequency and power factor. For example, the deep Q network can adopt a three-layer fully connected neural network structure, with 3 neurons in the input layer, corresponding to the speed, current and torque respectively; the number of neurons in the two hidden layers is set to 64 and 32 respectively; the number of neurons in the output layer is 1, representing the Q value. The policy gradient network can adopt a two-layer fully connected neural network structure, with 3 neurons in the input layer, corresponding to the speed, current and torque respectively, the number of neurons in the hidden layer is set to 64, and the number of neurons in the output layer is 6, representing the mean and variance of the excitation current, switching frequency and power factor respectively.

[0126] Next, calculate the energy consumption state of the permanent magnet motor. According to the collected motor speed, current and torque data, calculate the copper loss, iron loss and switching loss of the permanent magnet motor. Copper loss is proportional to the square of the current, iron loss is proportional to the square of the speed, and switching loss is proportional to the switching frequency. Superimpose the copper loss, iron loss and switching loss to obtain the total energy consumption state. For example, when the motor speed is 1000rpm and the current is 10A, the copper loss is measured to be 100W, the iron loss is 50W, and the switching frequency is 10kHz. The measured switching loss is 20W, then the total energy consumption state is 170W.

[0127] Then, the energy consumption status is evaluated. An energy consumption evaluation function is constructed based on the total energy consumption status, and the energy consumption evaluation function is combined with the discount factor to construct the Q value function. The energy consumption status of different control parameter combinations is evaluated through the Q value function, and the evaluation results are input into the policy gradient network. For example, the energy consumption evaluation function can be defined as the negative value of the total energy consumption, and the discount factor is set to 0.9.

[0128] After that, adjust the weighting coefficient. The system energy consumption index, dynamic response index and stability index are weighted and combined to obtain the comprehensive performance index. Based on the change trend of the comprehensive performance index, the weighting coefficient is adaptively adjusted by exponential decay. The adjustment result of the weighting coefficient is fed back to the policy gradient network. For example, initially, the weighting coefficients of the energy consumption index, dynamic response index and stability index are set to 0.5, 0.3 and 0.2 respectively. If the comprehensive performance index continues to decline, the weight of the energy consumption index is increased, and the weights of the dynamic response index and the stability index are reduced.

[0129] Finally, the optimal control parameters are output and applied. The policy gradient network recalculates the control parameter probability distribution based on the received evaluation results and weighted coefficients, and outputs the updated excitation current, switching frequency, and power factor. According to the updated control parameters, the operating state of the permanent magnet motor is adjusted in real time through the inverter. For example, if the updated excitation current output by the policy gradient network is 8A, the switching frequency is 8kHz, and the power factor is 0.95, the motor excitation current is adjusted to 8A and the switching frequency is adjusted to 8kHz through the inverter.

[0130] The beneficial effects of this method can be summarized in the following three aspects:

[0131] Reduce energy consumption: By optimizing control parameters, the copper loss, iron loss and switching loss of the permanent magnet motor can be effectively reduced, thereby reducing the overall energy consumption.

[0132] Improve the overall performance of the system: This method comprehensively considers energy consumption indicators, dynamic response indicators and stability indicators, and achieves the improvement of the overall performance of the system by adaptively adjusting the weighting coefficients.

[0133] Intelligent control: By adopting deep reinforcement learning algorithm, intelligent control of permanent magnet motor can be realized without relying on complex mathematical models, which improves the robustness and adaptability of the control system.

[0134] In an optional implementation, the network parameters of the deep neural network model and the strategy parameters of the reinforcement learning algorithm are updated in real time by the self-correction module to achieve closed-loop control of adaptive energy-saving optimization of the permanent magnet motor load, including:

[0135] Calculate the standard deviation of each characteristic quantity in the load characteristic vector, calculate the sensitivity of each characteristic quantity to the control output based on the standard deviation, and use the sensitivity as the characteristic weight coefficient; multiply the load characteristic vector by the characteristic weight coefficient to obtain a weighted characteristic vector, input the weighted characteristic vector into a deep neural network model, predict the operating state of the permanent magnet motor through the deep neural network model, and construct a loss function by combining the error between the predicted state and the actual state, the regularization term of the network parameter, and the smoothing term of the control quantity;

[0136] Calculate the gradient of the loss function with respect to the network weight and the network bias, update the network weight and the network bias by back propagation, use the updated network output as the state input of the reinforcement learning strategy network, and control the action through the reinforcement learning strategy network output;

[0137] Applying the control action to the permanent magnet motor system to obtain an immediate reward, constructing a performance objective function including system energy consumption, dynamic response and control stability, calculating the gradient of the performance objective function to the policy parameters, and adaptively adjusting the learning step size according to the ratio of the current policy gradient to the historical policy gradient;

[0138] The strategy parameters are optimized by using an adaptive learning step size, a state feedback controller is designed based on the optimized strategy parameters, a state feedback gain matrix is ​​multiplied by the system state to obtain a feedback control quantity, and the feedback control quantity is superimposed with a pre-acquired feedforward compensation term to obtain a control output;

[0139] The system state error caused by the control output is calculated, a Lyapunov function is constructed and its derivative is obtained, an adaptive law is designed based on the Lyapunov stability theory, the state feedback gain matrix is ​​adjusted online, and an integral constraint is imposed on the adjustment amount to ensure parameter convergence.

[0140] The load adaptive energy-saving optimization control method for permanent magnet motors aims to achieve efficient operation of the motor under different load conditions. The core idea of ​​this method is to predict the motor operating state through a deep neural network model, and adjust the control strategy in real time in combination with a reinforcement learning algorithm, ultimately achieving the energy-saving optimization goal.

[0141] First, it is necessary to extract load characteristics, such as load torque, speed, current, etc. Then, the standard deviation of these characteristic quantities is calculated to evaluate the degree of influence of each characteristic quantity on the control output. For example, if the standard deviation of the load torque is large, it means that the torque change has a greater impact on the control output, and the torque characteristic is given a higher weight. The standard deviation of each characteristic quantity is combined with its influence on the control output to obtain the characteristic weight coefficient. The original load characteristic vector is multiplied by the corresponding characteristic weight coefficient to obtain a weighted characteristic vector.

[0142] Next, the weighted feature vector is input into a pre-trained deep neural network model. The model can predict the motor's operating state under the current load conditions, such as speed, current, efficiency, etc. Assume that the input weighted feature vector is [0.8, 0.2, 0.5], and the model predicts that the motor speed is 1500 rpm and the current is 10 amps.

[0143] The predicted motor running state is compared with the actual measured state, and the error between them is calculated. At the same time, in order to prevent the model from overfitting, the regularization term of the network parameters is introduced. In addition, in order to ensure the smoothness of the control output, the smoothing term of the control amount is added. These three terms are combined to construct the loss function. For example, if the predicted speed is 1500 rpm and the actual speed is 1550 rpm, the speed error is 50 rpm. Assuming the regularization term is 0.1 and the control amount smoothing term is 0.05, the loss function value is 50+0.1+0.05=50.15.

[0144] Based on the value of the loss function, the back propagation algorithm is used to calculate the gradient of the loss function to the network weights and biases, and the gradient is used to update the network parameters. The updated network output will be used as the input of the reinforcement learning policy network. The reinforcement learning policy network outputs control actions based on the input state, such as the adjustment amount of voltage or current. Assume that the updated network output is [0.7, 0.3], and the control action output by the policy network is to increase the voltage by 2 volts.

[0145] Apply control actions to the permanent magnet motor system and observe the system response. Based on the system response, calculate the immediate reward. The design of the reward function needs to consider multiple performance indicators such as system energy consumption, dynamic response, and control stability. For example, if the control action improves the motor efficiency, a positive reward is given; if the control action causes the system to oscillate, a negative reward is given. Assuming that the current system energy consumption is 100 watts, the dynamic response time is 0.1 seconds, and the control stability is good, the immediate reward is +1.

[0146] In order to optimize the control strategy, it is necessary to construct a performance objective function that includes system energy consumption, dynamic response, and control stability. Calculate the gradient of the performance objective function to the policy parameters. In order to improve learning efficiency, the learning step size is adaptively adjusted according to the ratio of the current policy gradient to the historical policy gradient. For example, if the current policy gradient is larger than the historical policy gradient, it means that the learning speed is faster, and the learning step size can be appropriately increased.

[0147] The policy parameters are optimized using adaptive learning step size. A state feedback controller is designed based on the optimized policy parameters. The feedback control amount is obtained by multiplying the state feedback gain matrix with the system state. The feedback control amount is superimposed with the pre-acquired feedforward compensation term to obtain the final control output. For example, assuming that the state feedback gain matrix is ​​[0.5, 0.2] and the system state is [10, 5], the feedback control amount is 0.5*10+0.2*5=6. Assuming that the feedforward compensation term is 2, the final control output is 6+2=8.

[0148] Finally, the system state error caused by the control output is calculated and the Lyapunov function is constructed. The Lyapunov function is used to evaluate the stability of the system. The derivative of the Lyapunov function is obtained, and the adaptive law is designed based on the Lyapunov stability theory to adjust the state feedback gain matrix online. In order to ensure parameter convergence, an integral constraint is imposed on the adjustment amount.

[0149] Beneficial effects:

[0150] 1. Improve motor efficiency: By adjusting the control strategy in real time, this method can keep the motor running efficiently under different load conditions, thereby reducing energy consumption.

[0151] 2. Improve dynamic response: This method can quickly adjust the control output according to load changes, thereby improving the dynamic response speed and accuracy of the system.

[0152] 3. Enhance system stability: The adaptive law designed based on Lyapunov stability theory can ensure the stable operation of the system and avoid oscillation or loss of control.

[0153] In an optional implementation, the strategy parameters are optimized using an adaptive learning step size, a state feedback controller is designed based on the optimized strategy parameters, a state feedback gain matrix is ​​multiplied by the system state to obtain a feedback control amount, and the feedback control amount is superimposed with a pre-acquired feedforward compensation term to obtain a control output, including:

[0154] Calculate the ratio of the current policy gradient to the previous policy gradient, multiply the ratio by a preset attenuation coefficient and take the negative exponent, multiply the negative exponent by the initial learning step length to obtain the current adaptive learning step length, and update the policy parameters using the product of the current adaptive learning step length and the current policy gradient;

[0155] Constructing a state feedback controller, calculating the Rickati equation to obtain a state feedback matrix, multiplying the inverse matrix of the control weight matrix with the transpose of the input matrix and the state feedback matrix to obtain a state feedback gain matrix, and multiplying the state feedback gain matrix with the system state vector to obtain a feedback control component;

[0156] Constructing a feedforward compensation controller, calculating the inverse matrix of the product of the output matrix and the input matrix, multiplying the inverse matrix with the product of the output matrix and the system state matrix to obtain a feedforward compensation matrix, and multiplying the feedforward compensation matrix with a reference input to obtain a feedforward control component;

[0157] The feedback control component and the feedforward control component are superimposed to obtain a total control output, a quadratic function is constructed by multiplying the total control output and a control weight matrix, and the quadratic function is integrated within a time interval to obtain a control performance evaluation index;

[0158] Calculate a new policy gradient based on the control performance evaluation index, substitute the new policy gradient into the adaptive learning step calculation formula to update the learning step, and use the updated learning step to optimize and adjust the state feedback gain matrix and the feedforward compensation matrix;

[0159] The optimized state feedback gain matrix and feedforward compensation matrix are re-substituted into the controller to generate a new control output to act on the system, forming a closed-loop adaptive control.

[0160] First, initialize the policy parameters. Policy parameters can be understood as a set of values ​​used to adjust the behavior of the controller. For example, the initial policy parameters can be set to [0.1, 0.2, 0.3]. At the same time, set the initial learning step and decay coefficient. The learning step determines the amplitude of each parameter update, for example, set it to 0.01. The decay coefficient controls the decay speed of the learning step, for example, set it to 0.95.

[0161] Next, calculate the policy gradient at the current moment. The policy gradient indicates the sensitivity of the control performance index to the policy parameters. It can be estimated by making a small perturbation to the control performance index and observing the change in the index. For example, assuming that the performance index improves by 0.05 after the perturbation, the policy gradient can be approximated as 0.05 divided by the perturbation amount. At the same time, record the policy gradient at the previous moment.

[0162] Then, calculate the adaptive learning step. Multiply the ratio of the current policy gradient to the previous policy gradient by the attenuation coefficient, then take the negative exponential operation, and then multiply by the initial learning step to get the current adaptive learning step. For example, assuming that the current policy gradient is 0.1 and the previous policy gradient is 0.08, the ratio is 1.25, which is 1.1875 after multiplying by the attenuation coefficient of 0.95, which is about 0.84 after taking the negative exponent, and then multiplying by the initial learning step of 0.01, the current learning step is about 0.0084.

[0163] Use the calculated adaptive learning step size to update the policy parameters. Multiply the current learning step size by the current policy gradient, and then use this product to update the policy parameters. For example, if the current learning step size is 0.0084 and the current policy gradient is 0.1, the parameter update amount is 0.00084. Assuming the original policy parameters are [0.1, 0.2, 0.3], the updated policy parameters are [0.10084, 0.20084, 0.30084].

[0164] Design a state feedback controller based on the updated policy parameters. Obtain the state feedback matrix by solving the Rickati equation. Then, multiply the inverse matrix of the control weight matrix by the transpose of the input matrix and the state feedback matrix to obtain the state feedback gain matrix. For example, assuming that the state feedback matrix is ​​[[1,2],[3,4]], the inverse matrix of the control weight matrix is ​​[[0.5,0],[0,0.5]], and the transpose of the input matrix is ​​[[0.1,0.2]], then the state feedback gain matrix is ​​[[0.05+0.3,0.1+0.4]], that is, [[0.35,0.5]]. Multiply the state feedback gain matrix by the system state vector to obtain the feedback control component.

[0165] Construct a feedforward compensation controller. Calculate the inverse matrix of the product of the output matrix and the input matrix. Then multiply this inverse matrix by the product of the output matrix and the system state matrix to obtain the feedforward compensation matrix. For example, assuming the output matrix is ​​[[5,6]] and the input matrix is ​​[[0.1],[0.2]], then their product is 0.5+1.2=1.7, and its inverse matrix is ​​1 / 1.7. Assuming the system state matrix is ​​[[0.7],[0.8]], then the feedforward compensation matrix is ​​(1 / 1.7)*[[5,6]]*[[0.7],[0.8]], which is approximately [[3.65,4.18]]. Multiply the feedforward compensation matrix by the reference input to obtain the feedforward control component.

[0166] The feedback control component and the feedforward control component are superimposed to obtain the total control output. The product of the total control output and the control weight matrix is ​​used to construct a quadratic function, which is integrated over a time interval to obtain a control performance evaluation index.

[0167] The new policy gradient is calculated based on the control performance evaluation index, and it is substituted into the adaptive learning step calculation formula to update the learning step. The updated learning step is used to optimize the state feedback gain matrix and the feedforward compensation matrix.

[0168] Substitute the optimized state feedback gain matrix and feedforward compensation matrix back into the controller to generate a new control output to act on the system to form a closed-loop adaptive control. Repeat the above steps until the control performance index meets the requirements.

[0169] The beneficial effects of this method are reflected in three aspects:

[0170] 1. Improve control accuracy: By optimizing the strategy parameters through adaptive learning step size, the controller can better adapt to the dynamic characteristics of the system, thereby improving control accuracy.

[0171] 2. Enhance system robustness: This method combines state feedback and feedforward compensation, which can effectively suppress the influence of system disturbances and uncertainties and enhance the robustness of the system.

[0172] 3. Optimize control performance: By continuously iterating and optimizing strategy parameters, state feedback gain matrix and feedforward compensation matrix, the control performance indicators can be optimized and effective control of the system can be achieved.

[0173] Figure 2 FIG. 1 is a schematic diagram of a structure of a permanent magnet motor load adaptive energy-saving optimization system based on artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0174] The first unit is used to collect real-time load data of the permanent magnet motor under different working conditions through multiple load sensors, wherein the real-time load data includes speed, torque, current, voltage, temperature, and vibration frequency, and input the real-time load data into a pre-trained deep neural network model, wherein the deep neural network model adopts a hybrid architecture of a convolutional neural network and a long short-term memory network, and performs feature extraction and working condition analysis on the real-time load data through the deep neural network model to obtain a load feature vector of the permanent magnet motor;

[0175] The second unit is used to construct a dynamic optimization strategy of the permanent magnet motor based on the load feature vector by using a reinforcement learning algorithm, wherein the reinforcement learning algorithm includes a deep Q network and a policy gradient network, and the energy consumption state under different control parameter combinations is evaluated by the deep Q network, and the optimal control parameters are output by the policy gradient network, wherein the optimal control parameters include the excitation current, switching frequency, and power factor of the permanent magnet motor, and the operating state of the permanent magnet motor is adjusted in real time according to the optimal control parameters;

[0176] The third unit is used to construct an online learning system including a self-correction module. The online learning system continuously collects the operating data of the permanent magnet motor under the optimal control parameters, and compares the operating data with the preset energy consumption optimization target. When the operating data deviates from the energy consumption optimization target, the network parameters of the deep neural network model and the strategy parameters of the reinforcement learning algorithm are updated in real time through the self-correction module to realize closed-loop control of adaptive energy-saving optimization of the permanent magnet motor load.

[0177] According to a third aspect of the embodiments of the present invention,

[0178] An electronic device is provided, comprising:

[0179] processor;

[0180] a memory for storing processor-executable instructions;

[0181] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0182] A fourth aspect of the embodiments of the present invention is:

[0183] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0184] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0185] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An artificial intelligence-based permanent magnet motor load adaptive energy-saving optimization method, characterized in that: include: The real-time load data of the permanent magnet motor under different working conditions are collected by multiple load sensors, wherein the real-time load data includes speed, torque, current, voltage, temperature, and vibration frequency, and the real-time load data is input into a pre-trained deep neural network model, wherein the deep neural network model adopts a hybrid architecture of a convolutional neural network and a long short-term memory network, and the real-time load data is subjected to feature extraction and working condition analysis by the deep neural network model to obtain a load feature vector of the permanent magnet motor; Based on the load characteristic vector, a reinforcement learning algorithm is used to construct a dynamic optimization strategy for the permanent magnet motor, wherein the reinforcement learning algorithm includes a deep Q network and a policy gradient network, wherein the deep Q network is used to evaluate the energy consumption state under different control parameter combinations, and the policy gradient network is used to output the optimal control parameters, wherein the optimal control parameters include the excitation current, switching frequency, and power factor of the permanent magnet motor, and the operating state of the permanent magnet motor is adjusted in real time according to the optimal control parameters; An online learning system including a self-correction module is constructed. The online learning system continuously collects the operating data of the permanent magnet motor under the optimal control parameters, and compares the operating data with a preset energy consumption optimization target. When the operating data deviates from the energy consumption optimization target, the network parameters of the deep neural network model and the strategy parameters of the reinforcement learning algorithm are updated in real time through the self-correction module to realize closed-loop control of adaptive energy-saving optimization of the permanent magnet motor load.

2. The method according to claim 1, characterized in that The real-time load data is input into a pre-trained deep neural network model, wherein the deep neural network model adopts a hybrid architecture of a convolutional neural network and a long short-term memory network, and the real-time load data is subjected to feature extraction and working condition analysis through the deep neural network model to obtain a load feature vector of the permanent magnet motor, including: The real-time load data is mapped to the interval of zero to one by the maximum and minimum value standardization method to obtain standardized load data; the standardized load data is input into the deep neural network model for feature extraction, the deep neural network model includes a convolutional neural network module and a long short-term memory network module, wherein the convolutional neural network module includes three convolutional layers connected in series, the convolution kernel sizes of the three convolutional layers connected in series are five times five, three times three, and three times three, respectively, the output of each convolutional layer is connected to a batch normalization layer and a rectified linear unit activation function layer, the long short-term memory network module includes a two-layer bidirectional long short-term memory network structure, and the number of hidden layer neurons of the two-layer bidirectional long short-term memory network structure is one hundred and twenty-eight; The spatial features output by the convolutional neural network module and the temporal features output by the long short-term memory network module are input into an attention fusion module, the attention weights of the spatial features and the temporal features are calculated by the attention fusion module, and the spatial features and the temporal features are weightedly fused according to the attention weights to obtain a 256-dimensional load feature vector, wherein the load feature vector includes load fluctuation features, temporal correlation features, and operating state features; The load feature vector is input into a fully connected mapping layer, and a working condition classification probability distribution is output through the fully connected mapping layer. The load working condition type of the permanent magnet motor is determined according to the working condition classification probability distribution. The load working condition type includes a normal load working condition, a light load working condition, a heavy load working condition and a fluctuating load working condition.

3. The method according to claim 1, characterized in that: Based on the load characteristic vector, a dynamic optimization strategy for the permanent magnet motor is constructed by using a reinforcement learning algorithm. The reinforcement learning algorithm includes a deep Q network and a policy gradient network. The energy consumption state under different control parameter combinations is evaluated by the deep Q network. The optimal control parameters output by the policy gradient network include: A Markov decision environment for permanent magnet motor control is constructed, the load characteristic vector of the permanent magnet motor is used as the state space, the control parameter combination is used as the action space, and a reward function is constructed based on energy consumption evaluation. The reward function is obtained by weighted combination of current energy consumption, deviation between actual output power and target power, and control quantity change rate; A control strategy of a permanent magnet motor is evaluated by using a deep Q network structure, wherein the deep Q network structure comprises four fully connected layers, and a rectified linear unit activation function layer is arranged between the four fully connected layers. Data of a state space and an action space are input into the deep Q network structure, an objective function is constructed based on a temporal difference error, and network parameters of the deep Q network structure are optimized by minimizing the objective function. Constructing a policy gradient network based on the Actor-Critic architecture, inputting the load feature vector of the state space into the policy gradient network, outputting a probability distribution of control parameters that obeys a Gaussian distribution through the policy gradient network, wherein the probability distribution includes a mean parameter and a variance parameter, calculating the ratio of the probability of a new policy to the probability of an old policy, estimating and constructing a clipping objective function based on the ratio and the advantage function, and optimizing the policy parameters of the policy gradient network using a proximal policy optimization algorithm; Establish a dual-network collaborative optimization mechanism, use the network parameters of the deep Q network structure as the target network parameters, use the policy parameters of the policy gradient network as the online network parameters, use a soft update method to dynamically adjust the target network parameters and the online network parameters, and control the rate at which the target network parameters are transferred to the online network parameters through a soft update coefficient; A priority experience replay pool is constructed, and the temporal difference error, the state space data, the action space data and the reward function data are stored in the priority experience replay pool; a priority sampling probability of the experience data is set based on the temporal difference error; a power function relationship between the priority sampling probability and the priority value is used as a sampling basis for the experience data; training data is selected from the priority experience replay pool according to the sampling basis; and the deep Q network structure and the policy gradient network are iteratively trained using the training data.

4. The method according to claim 3, characterized in that Constructing a policy gradient network based on the Actor-Critic architecture, inputting the load feature vector of the state space into the policy gradient network, outputting a control parameter probability distribution that obeys a Gaussian distribution through the policy gradient network, wherein the probability distribution includes a mean parameter and a variance parameter, calculating the ratio of the new policy probability to the old policy probability, estimating and constructing a clipping objective function based on the ratio and the advantage function, and optimizing the policy parameters of the policy gradient network using a proximal policy optimization algorithm, including: Constructing an Actor network and a Critic network, wherein the Actor network adopts a four-layer feedforward neural network structure, wherein the output layer of the Actor network generates the mean of the control parameters through a hyperbolic tangent function, and generates the variance of the control parameters through a soft positive function, and uses the mean as the center position parameter of the Gaussian distribution, and uses the variance as the discrete degree parameter of the Gaussian distribution, to construct a probability distribution of the control parameters that obeys the Gaussian distribution; Based on the control parameter probability distribution, the control parameter at the current moment is obtained, the control parameter is applied to the control object to obtain an immediate reward signal, an advantage function is calculated using a generalized advantage estimation method, the immediate reward signal, the discount factor, the state value function and the generalized advantage estimation parameter are substituted into a temporal difference error calculation formula to obtain a temporal difference error, and an advantage function estimation value is calculated according to the temporal difference error; Calculating the new strategy probability by using the Actor network, calculating the old strategy probability by using the Actor network of the previous iteration, calculating the ratio of the new strategy probability to the old strategy probability, multiplying the ratio by the advantage function estimate to obtain a strategy objective function, and limiting the strategy objective function within a preset clipping range to obtain a clipping objective function; An adaptive learning rate adjustment mechanism is constructed, and a first-order momentum estimate and a second-order momentum estimate of the policy gradient are calculated based on the clipping objective function. The first-order momentum estimate and the second-order momentum estimate are substituted into an adaptive learning rate calculation formula to obtain an adaptive learning rate, and the policy parameters of the Actor network are iteratively optimized using the adaptive learning rate.

5. The method according to claim 3, characterized in that: The energy consumption state under different control parameter combinations is evaluated by the deep Q network, and the optimal control parameters are output by the policy gradient network. The optimal control parameters include the excitation current, switching frequency, and power factor of the permanent magnet motor. The operation state of the permanent magnet motor is adjusted in real time according to the optimal control parameters, including: The speed, current and torque data of the permanent magnet motor are input into the deep Q network, the copper loss, iron loss and switching loss of the permanent magnet motor are calculated based on the deep Q network, and the copper loss, iron loss and switching loss are superimposed to obtain a total energy consumption state; The speed, current and torque data of the permanent magnet motor are input into the policy gradient network, and the policy gradient network outputs a control parameter probability distribution that obeys a Gaussian distribution, wherein the control parameter probability distribution includes a mean vector and a variance vector, and the mean vector includes the excitation current, switching frequency and power factor of the permanent magnet motor; Constructing an energy consumption evaluation function based on the total energy consumption state, combining the energy consumption evaluation function with a discount factor to construct a Q value function, evaluating the energy consumption state of different control parameter combinations through the Q value function, and inputting the evaluation result into the policy gradient network; The system energy consumption index, dynamic response index and stability index are weighted and combined to obtain the comprehensive performance index, and based on the change trend of the comprehensive performance index, the weighting coefficient is adaptively adjusted by exponential decay; The adjustment result of the weighted coefficient is fed back to the policy gradient network, and the control parameter probability distribution is recalculated through the policy gradient network, and the updated excitation current, switching frequency and power factor are output, and the operating state of the permanent magnet motor is adjusted in real time according to the updated control parameters.

6. The method according to claim 1, characterized in that The network parameters of the deep neural network model and the strategy parameters of the reinforcement learning algorithm are updated in real time by the self-correction module to realize closed-loop control of adaptive energy-saving optimization of the permanent magnet motor load, including: Calculate the standard deviation of each characteristic quantity in the load characteristic vector, calculate the sensitivity of each characteristic quantity to the control output based on the standard deviation, and use the sensitivity as the characteristic weight coefficient; multiply the load characteristic vector by the characteristic weight coefficient to obtain a weighted characteristic vector, input the weighted characteristic vector into a deep neural network model, predict the operating state of the permanent magnet motor through the deep neural network model, and construct a loss function by combining the error between the predicted state and the actual state, the regularization term of the network parameter, and the smoothing term of the control quantity; Calculate the gradient of the loss function with respect to the network weight and the network bias, update the network weight and the network bias by back propagation, use the updated network output as the state input of the reinforcement learning strategy network, and control the action through the reinforcement learning strategy network output; Applying the control action to the permanent magnet motor system to obtain an immediate reward, constructing a performance objective function including system energy consumption, dynamic response and control stability, calculating the gradient of the performance objective function to the policy parameters, and adaptively adjusting the learning step size according to the ratio of the current policy gradient to the historical policy gradient; The strategy parameters are optimized by using an adaptive learning step size, a state feedback controller is designed based on the optimized strategy parameters, a state feedback gain matrix is ​​multiplied by the system state to obtain a feedback control quantity, and the feedback control quantity is superimposed with a pre-acquired feedforward compensation term to obtain a control output; The system state error caused by the control output is calculated, a Lyapunov function is constructed and its derivative is obtained, an adaptive law is designed based on the Lyapunov stability theory, the state feedback gain matrix is ​​adjusted online, and an integral constraint is imposed on the adjustment amount to ensure parameter convergence.

7. The method according to claim 6, characterized in that The strategy parameters are optimized by using an adaptive learning step size, a state feedback controller is designed based on the optimized strategy parameters, the state feedback gain matrix is ​​multiplied by the system state to obtain a feedback control quantity, and the feedback control quantity is superimposed with the pre-acquired feedforward compensation term to obtain a control output including: Calculate the ratio of the current policy gradient to the previous policy gradient, multiply the ratio by a preset attenuation coefficient and take the negative exponent, multiply the negative exponent by the initial learning step length to obtain the current adaptive learning step length, and update the policy parameters using the product of the current adaptive learning step length and the current policy gradient; Constructing a state feedback controller, calculating the Riccati equation to obtain a state feedback matrix, multiplying the inverse matrix of the control weight matrix with the transpose of the input matrix and the state feedback matrix to obtain a state feedback gain matrix, and multiplying the state feedback gain matrix with the system state vector to obtain a feedback control component; Constructing a feedforward compensation controller, calculating the inverse matrix of the product of the output matrix and the input matrix, multiplying the inverse matrix with the product of the output matrix and the system state matrix to obtain a feedforward compensation matrix, and multiplying the feedforward compensation matrix with a reference input to obtain a feedforward control component; The feedback control component and the feedforward control component are superimposed to obtain a total control output, a quadratic function is constructed by multiplying the total control output and a control weight matrix, and the quadratic function is integrated within a time interval to obtain a control performance evaluation index; Calculate a new policy gradient based on the control performance evaluation index, substitute the new policy gradient into the adaptive learning step calculation formula to update the learning step, and use the updated learning step to optimize and adjust the state feedback gain matrix and the feedforward compensation matrix; The optimized state feedback gain matrix and feedforward compensation matrix are re-substituted into the controller to generate a new control output to act on the system, forming a closed-loop adaptive control.

8. An artificial intelligence-based permanent magnet motor load adaptive energy-saving optimization system, used to implement the method as described in any one of claims 1 to 7, characterized in that: include: The first unit is used to collect real-time load data of the permanent magnet motor under different working conditions through multiple load sensors, wherein the real-time load data includes speed, torque, current, voltage, temperature, and vibration frequency, and input the real-time load data into a pre-trained deep neural network model, wherein the deep neural network model adopts a hybrid architecture of a convolutional neural network and a long short-term memory network, and performs feature extraction and working condition analysis on the real-time load data through the deep neural network model to obtain a load feature vector of the permanent magnet motor; The second unit is used to construct a dynamic optimization strategy of the permanent magnet motor based on the load feature vector by using a reinforcement learning algorithm, wherein the reinforcement learning algorithm includes a deep Q network and a policy gradient network, and the energy consumption state under different control parameter combinations is evaluated by the deep Q network, and the optimal control parameters are output by the policy gradient network, wherein the optimal control parameters include the excitation current, switching frequency, and power factor of the permanent magnet motor, and the operating state of the permanent magnet motor is adjusted in real time according to the optimal control parameters; The third unit is used to construct an online learning system including a self-correction module. The online learning system continuously collects the operating data of the permanent magnet motor under the optimal control parameters, and compares the operating data with the preset energy consumption optimization target. When the operating data deviates from the energy consumption optimization target, the network parameters of the deep neural network model and the strategy parameters of the reinforcement learning algorithm are updated in real time through the self-correction module to realize closed-loop control of adaptive energy-saving optimization of the permanent magnet motor load.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images