Electric vehicle permanent magnet motor control method and control system based on temperature rise prediction

Abstract

The application discloses a kind of based on temperature rise prediction electric vehicle permanent magnet motor control method and control system, belong to motor control field.The method method includes: based on meshless method generates motor temperature rise sample data;Using chaos mapping-dream optimization algorithm to optimize multivariate variational modal decomposition parameters, reconstruct high-precision temperature rise data;GNN-BILSTM-Attention prediction model is constructed, and space-time double Attention mechanism is fused to predict future temperature rise;Establish temperature rise-electromagnetic parameter mapping model, design hierarchical early warning control strategy, form prediction-control closed loop.The application improves the speed and accuracy of temperature rise prediction, realizes prediction-control deep coupling, significantly enhances motor operation reliability and control accuracy.

Description

Technical Field

[0001] This invention relates to a motor control method and control system, and more particularly to a permanent magnet motor control method and control system for electric vehicles based on temperature rise prediction. Background Technology

[0002] As a core power device that converts electrical energy into mechanical energy, the electric motor has been deeply integrated into industrial production, transportation, energy and power, and consumer goods. However, traditional motor research still faces several technical bottlenecks: in electromagnetic design, the accuracy of loss calculations under the coupling effects of multiple physical fields (electromagnetic, thermal, and mechanical) is insufficient, leading to significant discrepancies between predicted and actual motor efficiency. In operation control, there is a lack of effective solutions to the drift of core parameters and performance degradation caused by temperature rise, further restricting the reliability of motor operation. This invention focuses on predicting the motor temperature field and optimizing control based on the prediction results.

[0003] Current methods for predicting motor temperature mainly rely on traditional numerical methods (such as the finite element method) and empirical formulas, but these methods suffer from significant technical bottlenecks: First, the finite element method requires the construction of complex mesh models, and mesh reconstruction is time-consuming when facing dynamic conditions such as motor startup and sudden load changes, making it difficult to meet real-time prediction requirements. Second, empirical formulas rely on fitting a large amount of experimental data, neglecting the complex physical field effects such as electromagnetic and thermal coupling within the motor, resulting in large prediction errors under various operating conditions and limiting applicability. Third, these methods only focus on temperature rise prediction and do not deeply couple the prediction results with the motor control system, failing to address issues such as motor resistance and inductance parameter drift and output power attenuation caused by temperature rise, making it difficult to guarantee the accuracy and stability of motor operation. Therefore, this invention mainly focuses on the prediction and control of motor temperature rise.

[0004] More specifically: 1. In the current process of motor temperature rise management analysis and modeling, the commonly used finite element method (FEA) requires high-precision mesh generation, which requires high computing power resources and cannot quickly provide a large amount of data for artificial intelligence training. It has problems of insufficient calculation accuracy and slow calculation efficiency.

[0005] 2. Traditional decomposition methods are prone to mode aliasing and endpoint effects, resulting in distortion of thermophysical mode decomposition and low feature fidelity. Key parameters of mode decomposition rely on manual or grid search for debugging, which is inefficient and prone to getting trapped in local optima. It cannot adapt to the needs of modeling massive amounts of data and directly affects the quality of subsequent prediction inputs.

[0006] 3. Traditional prediction models cannot adapt to the three-dimensional high-dimensional coupled data of "space-time-modality" in gridless methods. High-dimensional data processing is prone to overfitting or underfitting, making it difficult to extract key information. Furthermore, it cannot capture multi-dimensional dependencies at the same time, pays insufficient attention to core nodes and key time steps, and has poor feature fusion effect, which limits the accuracy and generalization ability of temperature rise prediction.

[0007] 4. Existing technologies lack a closed-loop design of "temperature rise prediction - control optimization". Even if the temperature rise prediction results are obtained, it is impossible to specifically correct the parameter drift caused by temperature rise and optimize the control performance. As a result, the prediction results are difficult to be transformed into actual operation guarantee capabilities, and the control accuracy and stability of the motor under dynamic temperature rise conditions are insufficient. Summary of the Invention

[0008] Purpose of the invention: In view of the above-mentioned prior art, a control method and control system for permanent magnet motors of electric vehicles based on temperature rise prediction is proposed to realize the prediction of motor temperature field and control optimization based on the prediction results.

[0009] Technical solution: A control method for permanent magnet motors in electric vehicles based on temperature rise prediction, comprising:

[0010] Step 1: Generate permanent magnet motor temperature rise data using a meshless method;

[0011] Step 2: Preprocess the temperature rise data, then adaptively optimize the parameters of the multivariate variational mode decomposition using the dream optimization algorithm improved by chaotic mapping. After completing the MVMD decomposition based on the optimal parameters, select the effective modes and reconstruct the temperature rise data based on correlation analysis.

[0012] Step 3: Construct a GNN-BILSTM-Attention temperature rise prediction model: The effective modes are reorganized into three-dimensional feature tensors, a GNN graph structure is constructed based on node coordinates, spatial features are extracted through GCN convolutional layers, the core heat-generating node features are enhanced by a spatial attention module, temporal features are extracted through BILSTM and combined with a temporal attention module to focus on key periods of temperature rise, and finally the temperature rise prediction results within a preset time window are output by a fully connected layer; the temperature rise prediction model is deployed to the edge computing unit of the motor control system;

[0013] Step 4: Temperature rise prediction-driven motor control optimization: Based on the temperature rise prediction results, the edge computing unit dynamically corrects the motor electromagnetic parameters through the temperature rise-electromagnetic parameter mapping model, and then adjusts the motor control parameters based on the updated electromagnetic parameters.

[0014] A control system for implementing a permanent magnet motor control method for electric vehicles based on temperature rise prediction includes:

[0015] The data generation module is used to generate sample data of temperature rise of permanent magnet motors based on the meshless method;

[0016] The data processing module is used for preprocessing and optimizing the temperature rise data.

[0017] The prediction model module is used to run the GNN-BILSTM-Attention temperature rise prediction model;

[0018] The control optimization module is used to perform control optimization driven by temperature rise prediction based on the output of the temperature rise prediction model.

[0019] The modules are connected via a data bus to form a predictive-control closed-loop system.

[0020] Beneficial effects: This invention first obtains high-dimensional spatiotemporal temperature rise data using a gridless method, decomposes the temperature rise into subsequences, and then uses artificial intelligence to extract features, fully leveraging the speed advantage of the gridless method. A trained temperature rise prediction model is then used to obtain future temperature rise data for the motor. This model is deployed to the edge computing unit of the motor control system, effectively improving the speed and accuracy of prediction. The temperature rise prediction results are transformed into control optimization actions. By constructing a prediction-driven control optimization closed loop, the electromagnetic parameters such as motor resistance and inductance are dynamically corrected based on the high-precision temperature rise prediction results, compensating for parameter drift errors and ensuring control accuracy, thereby improving the reliability of motor operation. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating the implementation of the method of the present invention. Detailed Implementation

[0022] The invention will now be further explained with reference to the accompanying drawings.

[0023] like Figure 1 As shown, a control method for permanent magnet motors in electric vehicles based on temperature rise prediction is described, with the following specific steps:

[0024] Step 1: Generate motor temperature rise data based on the meshless method. The sample data includes the temperature time series data of each node under different operating conditions.

[0025] Meshless methods eliminate the need for mesh construction, allowing for flexible adaptation to the complex geometry and physical field distribution of motors. They can accurately acquire data on the temperature changes over time in key components such as the stator and rotor under different loads and ambient temperatures. By rapidly obtaining sufficient temperature rise data, high-quality raw data is provided for subsequent temperature rise prediction model training.

[0026] 1.1: Input the geometric and physical parameters of the motor, and arrange N discrete points in the motor temperature field calculation domain.

[0027] The algorithm inputs the geometric descriptions, material properties, heat source distribution, boundary conditions, and initial conditions of each component of the motor. Within the motor's internal space where temperature analysis is required, including key heat-generating / heat-dissipating components such as the stator core, stator windings, rotor permanent magnets, casing, and cooling channels, a series of discrete points are randomly or arranged according to a critical region densification principle. These points are called computational nodes, and no pre-defined mesh connection relationships are required between nodes. The number of nodes is designed based on the motor size and accuracy requirements, with the total number of nodes N typically ranging from several hundred to several thousand.

[0028] 1.2: Using the moving least squares (MLS) method to construct spatial shape functions, the continuous real temperature field is approximated as a linear combination of discrete nodal temperatures.

[0029] The mathematical representation of this process is: ;in, Indicates position place, time Temperature; It is an approximate temperature field based on the meshless method; Represents a node Spatial shape functions; Represents a node At any moment Temperature value. , For the basis function vector, For matrix functions, The weight function is used; the basis functions can be linear or quadratic basis functions, and the weight function can be a Gaussian weight function.

[0030] 1.3: Approximate temperature field Substituting the partial differential control equations describing the internal temperature field of the motor, and combining them with the boundary conditions, the continuous partial differential equations are transformed into a set of equations concerning the nodal temperatures using the weighted residual method. The system of ordinary differential equations. The applied boundary conditions include convective boundaries, radiative boundaries, and thermal conduction boundaries.

[0031] 1.4: Solve the equations obtained in step 1.3 using the time integration method to obtain the transient thermal analysis results, that is, the temperature sequence of each node under different operating conditions as a function of time.

[0032] Since the motor temperature rise is a slow dynamic process, this embodiment selects the implicit time integration method. To ensure the generalization ability of the subsequent model, temperature sequences under multiple operating conditions are generated. Typical operating conditions include: different load conditions, different ambient temperature conditions, load sudden changes, start-stop cycles, and other dynamic conditions.

[0033] The final output data format is as follows: each node corresponds to a time-temperature series, and the overall data is three-dimensional spatiotemporal data of "node's three-dimensional coordinates + time t + temperature X". Data example: Node 1, coordinates (0.01, 0.02, 0.03), under rated load and 25℃ environment, X=25℃ at t=0s, X=32℃ at t=10s, X=38℃ at t=20s...X=85℃ at t=300s.

[0034] Step 2: Data preprocessing and optimization decomposition.

[0035] 2.1: Outlier handling for sample data generated by the gridless method.

[0036] In the raw data generated by the meshless method, outliers caused by sudden load changes exceeding the boundary or temperature changes due to numerical iteration errors during the simulation process are identified. Single-point outliers are filled using linear interpolation, while continuous outlier segments are directly removed from the entire data under the corresponding working condition.

[0037] 2.2: The original gridless data contains multi-dimensional features with extremely large differences in scale. Therefore, the min-max normalization method is used to map the multi-dimensional feature data to the [0,1] interval to avoid abnormal gradient descent during model training caused by the difference in scale.

[0038] 2.3: Improved Dream Optimization Algorithm (DOA) based on Chaotic Mapping to optimize Multivariate Variational Mode Decomposition (MVMD).

[0039] Because the meshless method for solving temperature rise data exhibits nonlinearity, non-stationarity, and multi-scale coupling characteristics, and includes numerical discretization errors and disturbances introduced by boundary conditions, this invention introduces a chaotic mapping-improved Dream Optimization (DOA) algorithm to adaptively find the optimal number of modes in MVMD in order to perform MVMD decomposition on the preprocessed data. With penalty factor This avoids the inefficiency and local optima problems of traditional manual debugging or grid search.

[0040] This step employs the Chaotic Mapping-Dream Optimization (DOA) algorithm to optimize the number of modes in the Multivariate Variational Mode Decomposition (MVMD). and penalty factor The specific steps are as follows.

[0041] 2.3.1: Initializing the Population with Chaotic Mapping: N chaotic sequences are generated using a Logistic mapping and normalized and mapped to the MVMD parameter space to generate the initial population. , and They are the first The individual's modality number and penalty factor; among which , , The first in the chaotic sequence A chaotic value, , These represent the upper and lower search boundaries for the modality number, , These represent the upper and lower search boundaries of the penalty factor, respectively.

[0042] 2.3.2: DOA Iterative Optimization: Design a fitness function that uses the following steps: a memory phase to select the current optimal parameters, a forgetting phase to apply chaotic perturbation to individuals with low fitness, and an association phase to expand the search range based on the current optimal parameters to generate new candidate individuals. When the fitness value change is less than [a certain value] over 10 consecutive generations, the optimization is successful. If the number of iterations exceeds the preset maximum value, output the optimal number of modes. and penalty factor .

[0043] Wherein, fitness function Where L is the fitness value, For reconstruction error, Information entropy; , yes The temperature rise data is represented by a time series matrix of spatial sampling points, where m is the number of spatial sampling points and T is the number of time sampling points. This represents the temperature rise data reconstructed after MVMD decomposition; , The energy percentage of the k-th mode in the MVMD decomposition. , This represents the k-th mode.

[0044] During the memorization phase, the fitness value of each parameter individual in the initial population is calculated, and the individual with the smallest value of L is selected as the current optimal parameter. .

[0045] During the forgetting phase, chaotic perturbations are applied to individuals with low fitness, mathematically represented as: , ,in and These are individuals after the loss and mutation. Chaotic random numbers generated for the Logistic mapping. The variable asynchronous length of the modal number, , For the variable length of the penalty factor, By subjecting low-fit individuals to chaotic perturbations, the algorithm can avoid getting trapped in local optima. This preserves the basic parameter characteristics of the individuals while breaking the fixed state of the difference parameters through the randomness of chaos, guiding the algorithm to explore new parameter regions, maintaining population diversity, and avoiding local convergence.

[0046] During the association phase, the search scope is expanded based on the current optimal parameters to generate new candidate individuals. and Specifically: , ,in This indicates the generation of pseudo-random numbers. .

[0047] After iterative evaluation, the globally optimal parameters of MVMD, i.e., the optimal number of modes, are output. and penalty factor .

[0048] Step 2.3 Improves the global search capability of the Dream Optimization Algorithm (DOA) through chaotic mapping, and adaptively solves for the optimal number of modes in the Multivariate Variational Mode Decomposition (MVMD). With penalty factor The Dream Optimization Algorithm (DOA) with chaotic mapping significantly enhances the algorithm's exploration ability and ability to escape local optima, thereby improving the algorithm's global optimization performance.

[0049] 2.4: MVMD Decomposition and Mode Selection: An MVMD model is constructed based on optimal parameters. The Alternating Direction Method (ADMM) is used iteratively to obtain multi-scale modal components. The Pearson correlation coefficient between each mode and the original temperature rise data is calculated. The modes constitute the effective mode set. After removing the noise modes, high-precision temperature rise data is reconstructed. The specific steps are as follows.

[0050] 2.4.1: MVMD temperature rise data processing flow with integrated improved algorithm.

[0051] Based on optimal parameters and A multi-scale modal component (MVMD) model is constructed to decompose the preprocessed temperature rise data and obtain the multi-scale modal components. The input... , , Through iterative and efficient solutions, accurate modal components can be obtained quickly with limited computing resources. With center frequency This provides high-quality feature input for subsequent modality selection.

[0052] 2.4.2: Modal screening and data reconstruction.

[0053] By eliminating noisy and invalid modes, high-precision temperature rise data is reconstructed, restoring the true temperature rise trend.

[0054] Effective mode selection: Input the k-th mode and temperature rise data The Pearson correlation coefficient between the k-th mode and the temperature rise data is calculated according to the following formula. .

[0055]

[0056] in, , They are and The average value. Modes containing effective temperature rise information are identified based on the Pearson correlation coefficient, and the correlation coefficient is retained. The corresponding modes constitute the effective mode set. .

[0057] Data Reconstruction: The effective modes after filtering are scattered multi-scale features and cannot be directly used for thermal analysis or prediction. Therefore, the effective modes are summed and reconstructed into complete temperature rise data. This yields high-precision temperature rise data after noise reduction.

[0058] Step 2.4 uses MVMD to adaptively decompose the gridless temperature rise data into several intrinsic modes with clear physical meaning, thereby separating the influence of different heat sources. Moreover, MVMD can maintain the frequency consistency of multi-spatial measurement data during the decomposition process, thus collaboratively mining the intrinsic frequency domain characteristics of the spatiotemporal evolution of the temperature field, laying the foundation for accurate prediction of the full-field temperature rise based on gridless data.

[0059] Step 2 employs an improved multivariate variational mode decomposition (MVMD) method to decompose the preprocessed data. By introducing the chaotic mapping-dream optimization algorithm, the optimal number of modes and bandwidth constraint parameters of the MVMD method are adaptively optimized, avoiding the inefficiency and local optima problems of traditional manual debugging or grid search. Modes containing effective temperature rise information are then identified, and the Pearson correlation coefficient between each decomposed mode and the original temperature rise data is calculated. A set S of effective modes with high correlation is selected, while noisy invalid modes are removed. High-precision temperature rise data is obtained through reconstruction of the effective modes, restoring the true temperature rise trend.

[0060] Step 3: Construct a GNN-BILSTM-Attention temperature rise prediction model. By introducing a spatial-temporal dual attention mechanism on the basis of the GNN-BILSTM architecture, the spatial and temporal features of the temperature rise data are accurately extracted. The specific steps are as follows.

[0061] 3.1: Feature Reorganization and Sample Division: The selected... Each effective mode is recombined into a three-dimensional feature tensor of "node-feature-time". The samples were divided into training samples and test samples according to time steps.

[0062] 3.2: Convert the data into a graph neural network (GNN) compatible format.

[0063] GNN graph structure construction: Adjacency relationships are defined based on the spatial coordinates of the meshless computation domain. The spatial sampling points corresponding to the nodes are calculated using the node coordinate matrix of the meshless computation domain, and the edges correspond to the spatial relationships between the nodes. Node features need to be multi-dimensional attributes of each node. The m spatial sampling points of the MVMD mode can be directly used as the node set of the GNN, and the spatial adjacency relationships between nodes can be constructed based on the geometric coordinates of the meshless computation domain.

[0064] Graph structure quantization: Transforms the physical laws of heat conduction into quantized relational logic that can be recognized by GNN.

[0065] Traditional GNNs often use binary adjacency matrices, but this approach fails to capture the thermal conduction gradient characteristics of gridless temperature rise data. The degree of thermal conduction influence varies between core nodes and edge nodes at different distances; the binary matrix treats strongly correlated and weakly correlated nodes equally, leading to distortion in subsequent feature aggregation. Therefore, this invention achieves graph structure quantization through three steps: distance calculation, weighted assignment, and weight normalization. The adjacency matrix weights are calculated based on the gridless node coordinates. .

[0066] Specifically, based on the 3D coordinates of the nodes using the meshless method, Euclidean distance is used to accurately quantify the spatial interval between nodes, ensuring that the distance value is negatively correlated with the heat conduction intensity.

[0067] Introducing the exponential decay function Make the support domain radius of the meshless basis functions The node weights within the range decrease smoothly with increasing distance; the distance is greater than the radius of the support domain of the meshless basis function. The time weight is set to 0 to avoid interference from irrelevant nodes.

[0068] The number of connections at sampling points in the gridless method varies. Without normalization, the features of nodes with high degrees will be over-amplified. By normalizing rows, we ensure that the sum of the adjacency weights of each node is 1, achieving fairness in feature aggregation for nodes with different degrees and providing a balanced spatial correlation foundation for subsequent GCN convolutional layers.

[0069] 3.3: Feature transfer in GNN convolutional layers.

[0070] This invention selects Graph Convolutional Networks (GCN) to achieve feature transfer because GCN, combined with a weighted adjacency matrix, can accurately capture spatial relationships and has higher computational efficiency.

[0071] The modal data after MVMD decomposition has a three-dimensional structure of "spatial nodes × modalities × time steps". GCN needs to be processed by time step. Although the spatial correlation between different time steps is stable, the modal feature values ​​change with time, so spatial features need to be aggregated with time steps. Specifically, S modal features are mapped to D1 hidden spaces. The ReLU activation function is selected based on the non-negativity of the temperature rise feature. ReLU can set negative feature values ​​to 0, avoiding meaningless negative features from participating in subsequent modeling, and at the same time solving the gradient vanishing problem of deep GCN.

[0072] The ReLU activation function is as follows: ,in, It is a time step The output features (spatial features) of the GCN convolutional layer. Indicates the activation function; It is a time step The input features are the effective modal features of spatial nodes in gridless data. ,in To predict the step size; It is the weight matrix of the GCN convolution kernel. ; It is a bias vector. .

[0073] 3.5: GNN multi-timestep feature concatenation.

[0074] The core purpose of multi-timestep feature concatenation is to preserve the integrity of spatial features in the temporal dimension. BILSTM then needs to process the two-dimensional sequence of "timestep × spatial encoding." If the features are not stacked in the correct timestep order, temporal correlations will break, affecting the temporal modeling accuracy of BILSTM. Finally, the final encoded features of the GNN are obtained. .

[0075] Feature validity verification is a crucial step in avoiding invalid spatial encoding. The threshold setting for the Pearson correlation coefficient is based on extensive experiments with meshless temperature rise data. If the correlation coefficient is < 0.5, it indicates that the spatial features encoded by the GCN are out of sync with the actual temperature rise trend. In this case, the parameters need to be adjusted and recalculated until the correlation coefficient meets the requirements, ensuring that the GNN output can effectively represent the correspondence between "spatial correlation" and "temperature rise trend," providing high-quality input for BILSTM time series modeling.

[0076] 3.5: Spatial Attention Module Processing.

[0077] Based on the characteristic that core nodes have a greater impact in temperature rise data, a spatial attention module and weighted processing are used to strengthen the features of core nodes and weaken the interference of edge nodes. Specifically, the GNN features and gridless node coordinates are first input; then, a learnable parameter matrix is ​​used... and bias Calculate the feature score for each node Then, combining the distance between the node and the core node of the device, the position correction term is calculated using an exponential decay function. This ensures that the closer the physical location, the higher the weight; then, Softmax normalization is used to transform "feature score × location correction term" into node weights. Finally, the weights The spatially weighted features are obtained by weighting the features of the corresponding nodes with the GNN features. .

[0078] Among them, feature scoring , This represents the GNN features of the i-th node at all time steps.

[0079] Position correction item , Indicates the core node, This indicates the calculation of Euclidean distance.

[0080] Normalized representation is: , where j represents the index of the node being traversed.

[0081] The feature weighting is represented as: , This represents the GNN-encoded feature tensor of the i-th node at all time steps after spatial attention weighting.

[0082] 3.6: Feature Dimension Reorganization:

[0083] For the output Concatenate the hidden features of all nodes in the GNN at each time step: ,in, This represents the BILSTM encoded feature at time step t, which is the recombined feature of the data structure as "time step × (node ​​× GNN dimension)". .

[0084] 3.7: BILSTM and Temporal Attention Processing.

[0085] The GNN and BILSTM are concatenated. The main function of BiLSTM is to perform temporal dynamic modeling of the features output by GNN.

[0086] Recombination characteristics The input is fed into a BILSTM. The forward BILSTM processes the data step-by-step from the first time step to the last, capturing the cumulative effect of historical temperature rises. The backward BILSTM processes the data in reverse from the last time step, capturing information about future trends. The forward and backward hidden states of each time step are concatenated and fused to output a temporal feature with a dimension of "time step × BILSTM hidden layer dimension". It provides the Attention module with input containing bidirectional temporal information.

[0087] Then, after processing by the temporal attention module and feature weighting, key time steps are strengthened through weight allocation. Specifically, the input features... With the predicted target First, learnable parameters and Calculate the mean of the feature at each time step and the predicted target. Relevance score A higher score indicates a greater contribution of that time step to the prediction; then, Softmax normalization is used to obtain the weights of each time step. This ensures that the weight of the inflection point time step is significantly higher than that of the stationary time step; finally, the time series features are... The features at each time step are weighted and summed to obtain a one-dimensional global time series feature. This feature focuses on key time-series information, providing core support for subsequent temperature rise prediction of the fully connected layer.

[0088] Among them, the correlation score , Temporal characteristics The feature at time step t.

[0089] The normalized weights are represented as: , This is for iterating through time steps.

[0090] The weighted summation of the BILSTM features at each time step is expressed as follows: , express The feature at the t-th time step.

[0091] 3.8 Fully Connected Layer Output: Temperature rise prediction results are obtained through mapping using the fully connected layer. ;in, This represents the learnable weight matrix of the output layer; This indicates the learnable bias term of the output layer.

[0092] Step 4: Model training.

[0093] An end-to-end model is formed, consisting of GCN → Spatial Attention → BILSTM → Temporal Attention → Fully Connected Layer. Training parameters are then configured: the Adam optimizer is selected, and iterative training is performed using mean squared error (MSE) as the loss function. Gradient clipping is enabled during training to prevent gradient explosion due to excessive parameter updates, ensuring training stability. The mean absolute error (MAE) of the training set is calculated after each training round. Training stops when the MAE no longer decreases for 10 consecutive rounds or reaches a preset threshold. Finally, the optimal model parameter set is output, including GCN convolutional kernel weights, Attention weights, BILSTM hidden layer parameters, and fully connected layer weights.

[0094] Step 5: Model Validation.

[0095] The global temporal features corresponding to the test samples are input into the trained model, and the temperature rise prediction results of the test set are output by the fully connected layer. The dimensions are consistent with the real labels of the test set, which are "spatial nodes × effective modes × prediction step size". Then, the prediction results and the real set are restored to temperature values ​​through inverse normalization. The prediction accuracy is evaluated by MAE to reflect the average deviation of the prediction. Finally, a verification report containing the actual temperature rise prediction results, the actual temperature rise of the test set, and the MAE error value is output.

[0096] Step 6: Temperature rise prediction-driven control optimization.

[0097] After model training and performance verification, the trained model is deployed to the edge computing unit of the motor control system. Based on the long-term temperature rise prediction results of multiple spatial nodes (stator, rotor, housing, etc.) output by the model, the following two control optimization steps are implemented to specifically solve the problems of parameter drift and control performance degradation caused by temperature rise, and realize the transformation of prediction capability into control effect.

[0098] 6.1: Parameter drift dynamic correction.

[0099] The core electromagnetic parameters of the motor, such as stator resistance and rotor inductance, drift linearly with temperature rise, causing deviations in excitation current calculation and torque output from the set value, thereby reducing control accuracy and failing to meet the requirements of high-precision operation.

[0100] First, a "temperature rise-electromagnetic parameter" mapping model is established: based on the temperature dependence law of the motor's electromagnetic characteristics, a quantitative mapping relationship between temperature rise and stator resistance and rotor inductance is constructed:

[0101]

[0102]

[0103] in, , These represent the values ​​of stator resistance and rotor inductance as a function of temperature. , These are a room-temperature reference resistor and an inductor, respectively. , These are the temperature coefficients of the resistance and inductance pairs, respectively. , These are the predicted temperature rise values ​​for the stator winding and rotor core, respectively, output by the temperature rise prediction model.

[0104] Real-time data interaction and parameter update: The edge computing unit extracts the temperature rise prediction data corresponding to the stator and rotor from the trained model according to the control cycle (e.g., 20ms) and transmits it to the motor controller through the industrial communication protocol; the motor controller synchronously calls the above mapping formula to dynamically update the stator resistance value and rotor inductance value in the electromagnetic model.

[0105] Control command adaptive correction: Based on the updated electromagnetic parameters, the calculation logic of the excitation current setpoint and torque setpoint is adjusted synchronously to ensure that the excitation characteristics and torque output are consistent with the set values, and to compensate for the accuracy loss caused by parameter drift.

[0106] Step 6.1 By dynamically tracking the temperature rise prediction results and correcting the electromagnetic parameters, the deviation between the motor's operating position, speed and control commands can be stably controlled to a very small extent, significantly restoring and ensuring high-precision control performance.

[0107] 6.2: Control performance optimization.

[0108] Excessive motor temperature rise can lead to decreased efficiency and reduced output power, which in turn increases the response delay of the control system, deteriorates dynamic performance, and makes it unable to quickly track control commands.

[0109] Temperature rise warning level determination: based on the motor's rated temperature rise threshold Combined with the predicted temperature rise of the motor housing output by the model Divide the warning zones into two levels:

[0110] Level 1 Warning: The stage has been determined to be at risk of performance degradation.

[0111] Level 2 warning: It has been determined to be in the critical stage of performance degradation.

[0112] By dividing the preload interval into two levels, hierarchical adaptive control is used for adjustment:

[0113] Level 1 Early Warning Response: Initiate "Power-Heat Dissipation" Coordinated Optimization, optimize the current waveform and reduce copper loss by adjusting the PWM carrier frequency, limiting the duty cycle upper limit; simultaneously adjust the current loop PI parameters to improve response speed, and link the cooling system to enhance heat dissipation efficiency, balancing power output and temperature rise suppression.

[0114] Level II Early Warning Response: Implement a closed-loop strategy of "flexible load reduction - temperature rise decline - load recovery": gradually reduce the load rate in fixed steps to avoid sudden load changes that exacerbate the temperature rise; maintain the accuracy of control command tracking during the load reduction process, and gradually restore the load rate in reverse steps after the model predicts that the temperature rise has fallen back to the safe range.

[0115] Step 6.2 Improve motor performance by using human temperature rise early warning, maintain high motor efficiency, reduce control system response delay, effectively avoid dynamic performance deterioration caused by temperature rise, and ensure stable and reliable motor operation.

[0116] Step 6 constructs a closed-loop control method for temperature rise prediction, parameter correction, and performance optimization. This method does not rely on complex experimental calibration and directly implements adaptive control based on the output of the trained model. It achieves deep coupling between the model's predictive capability and the motor's control requirements, significantly improving the motor's control accuracy and operational stability across the entire temperature rise range.

[0117] This embodiment uses a permanent magnet synchronous motor for a certain type of electric vehicle as the controlled object. The motor has a rated power of 150kW and a rated speed of 10000rpm. The input parameters are the geometric parameters of each component of the motor: stator outer diameter 300mm, rotor inner diameter 180mm, etc.; material properties: thermal conductivity of silicon steel sheets for stator core 40W / (m·K), thermal conductivity of copper for stator winding 398W / (m·K), etc.; heat source distribution: copper loss and iron loss calculation model; boundary conditions: ambient temperature 25℃, convective heat transfer coefficient 15W / (m²·K), and initial temperature 25℃. 1200 discrete computational nodes were arranged in key parts such as stator core, stator winding, rotor permanent magnet, and casing. Spatial shape functions were constructed using the moving least squares method. The partial differential equations of the temperature field were transformed into a system of ordinary differential equations using the weighted residual method. The implicit Euler method was used to solve the equations. Time series data of node temperature under different loads (50%, 75%, 100%, 120% of rated load), different ambient temperatures (-20℃, 0℃, 25℃, 45℃), load abrupt changes, and start-stop cycle conditions were generated. The data duration was 300s and the time step was 1s. Finally, a three-dimensional spatiotemporal dataset of "three-dimensional coordinates + 300 time steps + temperature values" for 1200 nodes was obtained.

[0118] Then, the generated raw data is preprocessed: outliers are identified using the 3σ criterion, single-point outliers are filled by linear interpolation, and outlier segments with more than 3 consecutive time steps are removed from the corresponding working condition data; and features such as coordinates and temperature are mapped to the [0,1] interval by min-max normalization.

[0119] Optimizing MVMD parameters using DOA improved by chaotic mapping: setting the search range for the number of modes. , Penalty factor search range , The chaotic sequence length is 100; the fitness function has a reconstruction error weight of 0.7 and an information entropy weight of 0.3, and the number of iterations is set to 200 generations. When the fitness change is less than 1 / 2 for 10 consecutive generations... Stop iteration when the optimal number of modes is reached, and output the optimal number of modes. Punishment factor Based on the optimal parameters, MVMD decomposition was performed, the Pearson correlation coefficient between each mode and the original data was calculated, and four effective modes with a correlation coefficient ≥ 0.5 were selected to reconstruct high-precision temperature rise data after denoising.

[0120] The reconstructed effective modal data is reorganized into a 1200×4×300 three-dimensional feature tensor, which is then divided into training and testing sets in a 7:3 ratio. The prediction step size is then determined. A GNN graph structure is constructed based on the 3D coordinates of nodes. The node spacing is calculated using Euclidean distance, and the support domain radius of the basis functions is determined using a meshless method. Exponential decay coefficient The weighted adjacency matrix is ​​obtained after normalization. The hidden dimension of the GCN convolutional layer... ReLU activation function; learnable parameter matrix in the spatial attention module. Dimensions are 64×1, with bias The vector is 1×1, with the core node selected as the center point of the stator winding; the BILSTM hidden layer has a dimension of 128, with 64 bidirectional neurons each, and a dropout coefficient of 0.2; the learnable parameters in the temporal attention module are... The dimension is 128×1. The dimension is 1×1; the output dimension of the fully connected layer is 1200×10 (1200 nodes × 10 prediction steps).

[0121] The model was trained using the Adam optimizer with a learning rate of 0.001, the MSE loss function, a gradient clipping threshold of 0.5, and a batch size of 32 per training round. Training was stopped when the MAE stopped decreasing for 10 consecutive rounds. The final training set MAE was 1.2℃ and the test set MAE was 1.5℃.

[0122] The trained model is deployed to the edge computing unit (computing power ≥ 2 TOPS) of the motor controller to acquire temperature rise prediction data of the stator winding and rotor core at 20ms control cycles. Based on the "temperature rise-electromagnetic parameter" mapping model, the stator resistance temperature coefficient is... Rotor inductance temperature coefficient The electromagnetic model parameters are dynamically updated to correct the excitation current and torque setpoints.

[0123] The motor's rated temperature rise threshold is set to 120℃, the first-level warning threshold to 80℃, and the second-level warning threshold to 100℃. When the predicted temperature rise of the casing reaches 80℃, the PWM carrier frequency is adjusted from 10kHz to 12kHz, the duty cycle is limited to an upper limit of 0.9, the proportional coefficient of the current loop PI parameter is adjusted from 0.8 to 1.0, the integral coefficient is adjusted from 0.05 to 0.08, and the cooling system is linked to increase the fan speed by 30%. When the predicted temperature rise of the casing reaches 100℃, the load rate is reduced in steps of 5% every 100ms. After the predicted temperature rise drops below 70℃, the load rate is restored in reverse steps of 5% every 200ms.

[0124] Through the above implementation process, the present invention achieves rapid and accurate prediction of temperature rise of permanent magnet motor for electric vehicles, with prediction error controlled within 2℃. At the same time, through predictive drive control optimization, the deviation of motor position and speed control is reduced by more than 30%, effectively suppressing the performance degradation caused by temperature rise and improving the reliability and control accuracy of motor operation.

[0125] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A temperature rise prediction-based control method for a permanent magnet motor of an electric vehicle, characterized by, The application relates to a permanent magnet motor temperature rise prediction and control optimization method, which comprises the following steps: Step 1: generating permanent magnet motor temperature rise data by a meshless method; Step 2: pre-processing the temperature rise data, adaptively optimizing the parameters of multivariate variational mode decomposition through a dream optimization algorithm improved by a chaotic mapping, screening effective modes based on correlation analysis after MVMD decomposition based on the optimal parameters, and reconstructing the temperature rise data; Step 3: constructing a GNN-BILSTM-Attention temperature rise prediction model: recombining the effective modes into a three-dimensional feature tensor, constructing a GNN graph structure based on node coordinates, extracting spatial features through a GCN convolution layer, strengthening the core heating node features through a spatial Attention module, extracting time sequence features through a BILSTM and combining a time Attention module to focus on the key time period of the temperature rise, and finally outputting the temperature rise prediction results in a preset time window in the future by a full connection layer; the temperature rise prediction model is deployed to an edge computing unit of a motor control system; Step 4: motor control optimization driven by temperature rise prediction: based on the temperature rise prediction results, the edge computing unit dynamically corrects motor electromagnetic parameters through a temperature rise-electromagnetic parameter mapping model, and adjusts motor control parameters based on the updated electromagnetic parameters; In step 2, the parameters of the multi-variate variational modal decomposition include a modal number and a penalty factor The dream optimization algorithm improved by the chaotic mapping adaptively optimizes the parameters of the multi-variate variational modal decomposition, and specifically includes: initializing a population by using a Logistic chaotic mapping, screening a current optimal parameter through a memory stage, chaotically disturbing low-adaptability individuals through a forgetting stage, generating a new candidate individual by expanding a search range based on the current optimal parameter through an association stage, and outputting an optimal modal number and a penalty factor when a change amount of an adaptability value is less than for 10 continuous generations or an iteration number exceeds a preset maximum value.

2. The temperature rise prediction based electric vehicle permanent magnet motor control method according to claim 1, characterized in that, In step 2, the population is initialized by using the Logistic chaotic mapping, specifically, a chaotic sequence is generated by using the Logistic mapping, and the initial population is generated by normalizing the mapping to the parameter space of the multi-element variational modal decomposition wherein , , is the mth chaotic value, , , are the upper and lower search boundaries of the modal number, respectively, , are the upper and lower search boundaries of the penalty factor, respectively. A fitness function for calculating the fitness value is: where L is the fitness value, is the reconstruction error, is the information entropy; , is dimensional temperature rise data, m is the number of spatial sampling points, T is the number of time sampling points, represents the reconstructed temperature rise data after MVMD decomposition; , is the proportion of the kth modal energy of MVMD decomposition.

3. The temperature-rise prediction based electric vehicle permanent magnet motor control method according to claim 2, characterized in that, In step 2, the screening of effective modes based on the correlation analysis and the reconstruction of the temperature rise data specifically include: obtaining the multi-scale mode components by iterative solution with the alternating direction method, calculating the Pearson correlation coefficients of each mode and the original temperature rise data, and the modes with correlation coefficients constitute the effective mode set , and summing the effective modes to obtain the temperature rise data .

4. The temperature-rise prediction based electric vehicle permanent magnet motor control method according to claim 1, characterized in that, In step 3, the Euclidean distance is used to quantify the node space interval in the construction of the GNN graph structure, and the weight is set by an exponential decay function, and the distance is greater than the support domain radius of the meshless method base function The weight is set to 0, and the weighted adjacency matrix is obtained through row normalization.

5. The temperature-rise prediction based electric vehicle permanent magnet motor control method according to claim 4, characterized in that, In step 3, the core heating node feature is enhanced by the spatial Attention module, which specifically includes: inputting the spatial feature and the node coordinates of the meshless method and bias to calculate the feature score of each node , , where represents the spatial feature of the i-th node at all time steps; in combination with the distance between the node and the core node, a position correction term is calculated using the exponential decay function ; the node weight is obtained by Softmax normalization , where j represents the index of the traversed node, and m represents the number of spatial sampling points; finally, the node weight is weighted with the spatial feature of the corresponding node to obtain the spatial weighted feature ; the output is sequentially spliced with the GNN hidden layer features of all nodes at each time step to obtain the reorganized feature .

6. The temperature-rise prediction based electric vehicle permanent magnet motor control method according to claim 5, characterized in that, In step 3, the BILSTM is used to extract time sequence features and the time Attention module is used to focus on the key time period of the temperature rise. Including: recombination features Output timing features after BILSTM processing In the Attention module: first, through the learnable parameters And Calculate the correlation score of each time step feature of the timing feature With the predicted target mean , , Indicates the t-th time step feature of ; Then use Softmax normalization to get the time step weight ; Weighted sum of each time step feature of the timing feature , get the single-dimensional global timing feature ; Finally, through the full connection layer mapping, get the temperature rise prediction result , wherein Indicates the output layer learnable weight matrix, Indicates the output layer learnable bias term.

7. The temperature-rise prediction based electric vehicle permanent magnet motor control method of claim 1, wherein, In step 4, the edge computing unit extracts the temperature rise prediction data corresponding to the stator and rotor according to the control period, and the motor controller dynamically updates the stator resistance value and rotor inductance value in the electromagnetic model according to the temperature rise-electromagnetic parameter mapping model; based on the updated electromagnetic parameters, the calculation logic of the excitation current given value and the torque given value is adjusted synchronously to ensure that the excitation characteristics and torque output are consistent with the set values, and the precision loss caused by parameter drift is compensated; wherein the temperature rise-electromagnetic parameter mapping model includes a stator resistance-temperature mapping function , and a rotor inductance-temperature mapping function ; wherein , respectively represent the values of the stator resistance and the rotor inductance changing with the temperature, , respectively are the normal-temperature reference resistance and inductance; , respectively are the temperature coefficients of the resistance and inductance pairs; , respectively are the stator winding and rotor core temperature rise prediction values output by the temperature rise prediction model.

8. The temperature-rise prediction based electric vehicle permanent magnet motor control method of claim 1, wherein, In step 4, two-level early warning intervals are divided according to the comparison results of the predicted temperature rise and a threshold value, and corresponding control optimization strategies are implemented; wherein, the first-level early warning is that when the predicted temperature rise reaches 80% of the rated maximum value, the "power-heat dissipation" collaborative optimization is started, including adjusting PWM parameters, current loop PI parameters and a linked cooling system, for balancing power output and temperature rise suppression; the second-level early warning is that when the predicted temperature rise reaches 90% of the rated maximum value, the load rate is lowered in a fixed ladder, and the load rate is restored in a reverse ladder after the temperature rise falls back to a safe interval.

9. A control system implementing the method of any one of claims 1-8, characterized by The application relates to a permanent magnet motor temperature rise prediction and control optimization method, which comprises the following steps: A data generation module is used for generating permanent magnet motor temperature rise sample data based on a meshless method; A data processing module is used for pre-processing and optimizing decomposition of the temperature rise data; A prediction model module is used for running a GNN-BILSTM-Attention temperature rise prediction model; A control optimization module is used for performing temperature rise prediction driven control optimization according to the output of the temperature rise prediction model; Wherein, the modules are connected through a data bus to form a prediction-control closed loop system.

Images