Control method and system for a vehicle thermal management system based on multi-dimensional data

Abstract

The present invention discloses a control method and system for a vehicle thermal management system based on multi-dimensional data, relating to the technical field of vehicles. The method includes: collecting multi-dimensional data related to the control of the vehicle's thermal management system in real time; establishing a neural network architecture composed of a TCN encoder, a Transformer encoder, and a Cross attention decoder, and obtaining a hybrid deep learning model through model training; inputting the multi-dimensional data into the hybrid deep learning model to obtain corresponding control information, and controlling the vehicle's thermal management system according to the control information. The present invention solves the problems in the prior art that the data processing ability of the control model of the automotive thermal management system is insufficient, and the poor multi-source data fusion effect leads to inaccurate control.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicles, and particularly to a control method and system for a vehicle thermal management system based on multi-dimensional data. Background Art

[0002] In today's world, as clean energy transportation vehicles, the market share of new energy vehicles (NEVs) is growing rapidly and is gradually becoming the first choice for the public's daily travel. Among the core technologies of new energy vehicles, the performance of the power battery system is crucial. Lithium-ion power batteries have become the mainstream choice for new energy vehicle power batteries due to their advantages such as high energy density, large specific power, light weight, and long cycle life. However, the performance of lithium-ion power batteries is closely related to their operating temperature. Research shows that when the battery operating temperature exceeds the optimal range, the charge and discharge efficiency will be significantly reduced, the cycle life will decay rapidly, and even serious safety problems may occur. Therefore, an efficient and stable thermal management system plays a decisive role in ensuring battery performance and vehicle safety.

[0003] Currently, the thermal management systems for lithium-ion power batteries on the market mainly adopt three modes: air cooling, liquid cooling, and direct cooling. The air cooling mode is widely used due to its simple structure and low cost, but its cooling capacity is limited, it is difficult to meet the heat dissipation requirements of high-power batteries, and the temperature uniformity is poor. Although the liquid cooling mode has significantly better cooling effect than air cooling and more precise temperature control, its system structure is complex, which not only increases the weight and cost of the vehicle, but also places higher requirements on the vehicle layout space, and the maintenance cost is relatively high. Although the direct cooling mode has advantages in terms of cooling efficiency and response speed, there are still problems such as poor stability of the control method, large temperature fluctuations, and room for improvement in system reliability.

[0004] In view of the problems existing in the above various cooling modes, the industry has begun to try to optimize the performance of the thermal management system through intelligent control algorithms. Among them, the control method based on deep learning has received attention because it can handle complex non-linear relationships. For example, in the prior art, the thermal management system is regulated by obtaining and analyzing navigation system information in real time and combining vehicle internal sensor data. However, this method still has the following problems in practical applications: First, the LSTM model used is prone to the problem of gradient disappearance, resulting in limited long-sequence data processing ability; second, in terms of data fusion, this method does not adequately consider the interaction relationship between features, the multi-source data fusion method is too simple, and there is a lack of in-depth mining of temporal features, ultimately leading to inaccurate control of the thermal management system. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a control method and system for a vehicle thermal management system based on multi-dimensional data, aiming to solve the problems of insufficient data processing ability of the control model of the existing vehicle thermal management system and inaccurate control caused by poor multi-source data fusion effect.

[0006] The embodiments of the present invention are implemented as follows:

[0007] A control method for a vehicle thermal management system based on multi-dimensional data, the method comprising:

[0008] During the driving of the vehicle, multi-dimensional data related to the control of the vehicle's thermal management system is collected in real time;

[0009] A neural network architecture composed of a TCN encoder, a Transformer encoder, and a Cross attention decoder is established, and a hybrid deep learning model obtained by training the model with the neural network architecture is obtained;

[0010] The multi-dimensional data is input into the hybrid deep learning model to obtain corresponding control information, and the vehicle's thermal management system is controlled according to the control information. The control information includes the opening degree of the electronic expansion valve of the battery pack cooling system, the opening degree of the electronic expansion valve of the passenger compartment air conditioning system, and the compressor speed;

[0011] Among them, the TCN encoder is used to extract multi-scale time series features through multi-layer causal convolution and different dilation coefficients of multi-layer causal convolution. The Transformer encoder is used to capture long-term dependencies in the features through the multi-head self-attention mechanism. The Cross attention decoder is used to calculate the attention weights between the features output by the TCN encoder and the Transformer encoder to achieve adaptive fusion of multi-source information.

[0012] Further, in the above control method for a vehicle thermal management system based on multi-dimensional data, the TCN encoder includes a first-layer causal convolution layer and a second-layer causal convolution layer. The first-layer causal convolution layer uses 16 convolutional kernels, the dilation coefficient is 1, and the convolutional kernel size is 3×1; the second-layer causal convolution layer uses 32 convolutional kernels, the dilation coefficient is 2, and the convolutional kernel size is 3×1;

[0013] The TCN encoder receives the input features, performs causal convolution processing on the input features using the first-layer causal convolution layer, and performs residual connection with the input features to obtain preliminary input features;

[0014] The initial input features are sequentially processed by Layer Normalization and the ReLU activation function. The obtained target input features are subjected to causal convolution processing using the second-layer causal convolution layer and residual connection with the target input features to obtain the final input features.

[0015] The final input features are sequentially processed by Layer Normalization and the ReLU activation function to obtain the output features of the TCN encoder, so as to achieve multi-scale time series feature extraction.

[0016] Furthermore, for the above control method of the vehicle thermal management system based on multi-dimensional data, the processing formula for causal convolution processing is:

[0017] ;

[0018] Where is the output function, is the input data at time point t, is the time window size, is the weight system, used to adjust the importance of data at different time points, is the step size, used to control the sampling interval, means pushing forward time units, represents historical data points, is the parameterized variation function, used to perform non-linear transformation on the input data, is the sigmoid function, is the regularization coefficient, used to balance the model complexity, is the regularization function set, represents the number of regularization terms, represents the l-th regularization function.

[0019] Furthermore, for the above control method of the vehicle thermal management system based on multi-dimensional data, the Transformer encoder includes positional encoding, multi-head attention mechanism, and feed-forward neural network;

[0020] The Transformer encoder receives the input features. After positional encoding, the features generated by the multi-head attention mechanism are subjected to residual connection with the features after positional encoding to obtain the preliminary input features;

[0021] The preliminary input features are processed by Layer Normalization and the activation function, and then the features generated by the feed-forward neural network are subjected to residual connection with the features after being processed by Layer Normalization and the activation function to obtain the target input features;

[0022] The output features of the Transformer encoder are obtained by processing the target input features through Layer Normalization and activation functions.

[0023] Further, in the above control method of the vehicle thermal management system based on multi-dimensional data, the expression of the multi-head attention mechanism of the Transformer encoder is:

[0024] ;

[0025] where Q, K, and V represent the query matrix, key matrix, and value matrix respectively, d is the dimension of the attention head, R is the number of attention heads, is the weight coefficient of the r-th attention head; is the query transformation matrix of the r-th attention head; is the bias matrix of the r-th attention head; is the output transformation matrix of the r-th attention head, and T is the matrix transpose operation;

[0026] The calculation expression of the feed-forward neural network is:

[0027] ;

[0028] where W1 = [64, 256], W2 = [256, 64], b1 =

[256] , and b2 =

[64] .

[0029] Further, in the above control method of the vehicle thermal management system based on multi-dimensional data, the Cross-attention decoder includes a cross-attention layer, a feed-forward neural network, and an activation function layer;

[0030] The Cross-attention decoder receives the output features of the TCN encoder and the Transformer encoder,

[0031] and sequentially uses the cross-attention layer, the feed-forward neural network, and the activation function layer for feature extraction to obtain the output features of the Cross-attention decoder;

[0032] Among them, residual connections are set in both the cross-attention layer, the feed-forward neural network, and the activation function layer.

[0033] Further, in the above control method of the vehicle thermal management system based on multi-dimensional data, the calculation expression of the cross-attention layer is:

[0034] ;

[0035] ;

[0036] The computational expression of the feedforward neural network is as follows:

[0037] ;

[0038] The computational expression of the activation function layer is as follows:

[0039] ;

[0040] Wherein, W1 = [64, 256], W2 = [256, 64], b1 =

[256] , b2 =

[64] , Q comes from the Cross attention decoder, K and V come from the outputs of the TCN encoder and the Transformer encoder, X is the input feature, Multi Head represents the multi-head attention operation, and Layer Norm is the layer normalization.

[0041] Another object of the present invention is to provide a control system for a vehicle thermal management system based on multi-dimensional data, characterized in that the system includes:

[0042] An acquisition module for real-time acquisition of multi-dimensional data related to the control of the vehicle's thermal management system during vehicle driving;

[0043] A training module for establishing a neural network architecture composed of a TCN encoder, a Transformer encoder, and a Cross attention decoder, and obtaining a hybrid deep learning model obtained by model training with the neural network architecture;

[0044] A control module for inputting the multi-dimensional data into the hybrid deep learning model to obtain corresponding control information, and controlling the vehicle's thermal management system according to the control information, where the control information includes the opening degree of the electronic expansion valve of the battery pack cooling system, the opening degree of the electronic expansion valve of the occupant compartment air conditioning system, the rotational speed of the electric fan, and the rotational speed of the compressor;

[0045] Wherein, the TCN encoder is used to extract multi-scale time series features through multi-layer causal convolution and different dilation coefficients of multi-layer causal convolution, the Transformer encoder is used to capture long-term dependencies in the features through the multi-head self-attention mechanism, and the Cross attention decoder is used to calculate the attention weights between the features output by the TCN encoder and the Transformer encoder to achieve adaptive fusion of multi-source information.

[0046] Another object of the embodiments of the present invention is to provide a readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the above method are implemented.

[0047] Another object of the embodiments of the present invention is to provide an electronic device, including a memory, a processor, and a computer program stored on the memory and running on the processor, wherein when the processor executes the program, the steps of the above method are implemented.

[0048] The present invention designs a hybrid architecture of a TCN encoder, a Transformer encoder, and a Cross attention decoder. This architecture realizes multi-scale temporal feature extraction through multi-layer causal convolutions and different dilation coefficients of multi-layer causal convolutions, uses the multi-head self-attention mechanism of the Transformer encoder to capture the long-term dependencies of features, and combines the Cross attention decoder to calculate the attention weights between the features output by the TCN encoder and the Transformer encoder, realizing the adaptive fusion of multi-source information, effectively solving the limitations of traditional methods in processing long-sequence data and the defects such as the lack of in-depth mining of temporal features, improving the accuracy of model prediction, and thus significantly improving the control accuracy of the new energy vehicle thermal management system. It solves the problems of insufficient data processing ability of the control model of the vehicle thermal management system and inaccurate control caused by poor multi-source data fusion effect in the prior art.

[0049] In addition, the present invention at least also has the following beneficial effects:

[0050] 1. By improving the causal convolution and dilation coefficient of the TCN encoder, multi-scale temporal feature extraction is realized, effectively capturing local time dependencies, and at the same time avoiding the problem of gradient disappearance existing in traditional RNNs; adopting a multi-layer causal convolution structure, realizing multi-scale receptive fields through different dilation coefficients, the first-layer causal convolution performs preliminary temporal feature extraction, and the second-layer causal convolution further extracts higher-level temporal patterns; at the same time, residual connections and LayerNormalization are introduced to optimize gradient flow and numerical stability;

[0051] 2. The improved multi-head self-attention mechanism of the Transformer encoder can establish direct associations at any position in the sequence, supports parallel computing and has higher training efficiency, and can effectively capture long-term dependencies;

[0052] 3. The improved Cross attention decoder realizes the adaptive fusion of multi-source information by calculating the attention weights between the features output by the TCN encoder and the Transformer encoder, improves the model's comprehensive understanding ability of spatio-temporal features, enhances the model's processing ability for multi-source heterogeneous data, and finally maps the fused features to the target output space through a fully connected layer, improving the feature utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0053] Figure 1Flow chart of the control of a vehicle thermal management system based on multi-dimensional data provided by the first embodiment of the present invention;

[0054] Figure 2 Schematic diagram of the neural network architecture in a control method of a vehicle thermal management system based on multi-dimensional data provided by an embodiment of the present invention;

[0055] Figure 3 Schematic diagram of the TCN encoder architecture in a control method of a vehicle thermal management system based on multi-dimensional data provided by an embodiment of the present invention;

[0056] Figure 4 Schematic diagram of the Transformer encoder architecture in a control method of a vehicle thermal management system based on multi-dimensional data provided by an embodiment of the present invention;

[0057] Figure 5 Schematic diagram of the Cross attention decoder architecture in a control method of a vehicle thermal management system based on multi-dimensional data provided by an embodiment of the present invention

[0058] Figure 6 Schematic diagram of the change of the loss function of the training set and the validation set in a control method of a vehicle thermal management system based on multi-dimensional data provided by an embodiment of the present invention;

[0059] Figure 7 Comparison diagram of the predicted value and the actual value of the opening of the battery pack electronic expansion valve in a control method of a vehicle thermal management system based on multi-dimensional data provided by an embodiment of the present invention;

[0060] Figure 8 Comparison diagram of the predicted value and the actual value of the opening of the passenger compartment air conditioner electronic expansion valve in a control method of a vehicle thermal management system based on multi-dimensional data provided by an embodiment of the present invention;

[0061] Figure 9 Comparison diagram of the predicted value and the actual value of the compressor speed in a control method of a vehicle thermal management system based on multi-dimensional data provided by an embodiment of the present invention;

[0062] Figure 10 Comparison diagram of the control accuracy of the passenger compartment temperature under various typical working conditions between the embodiment of the present invention and the prior art;

[0063] Figure 11 Comparison diagram of the COP between the embodiment of the present invention and the prior art;

[0064] Figure 12 Structure block diagram of the control system of a vehicle thermal management system based on multi-dimensional data in the third embodiment of the present invention.

[0065] The following specific embodiments will further illustrate the present invention in conjunction with the above-mentioned drawings. Specific Embodiments

[0066] To facilitate the understanding of the present invention, the present invention will be described more comprehensively below with reference to the relevant drawings. Several embodiments of the present invention are shown in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. On the contrary, these embodiments are provided to make the disclosure of the present invention more thorough and comprehensive.

[0067] It should be noted that when an element is referred to as being "fixedly provided on" another element, it can be directly on the other element or there may also be an intermediate element. When an element is considered to be "connected" to another element, it can be directly connected to the other element or there may be an intermediate element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only.

[0068] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the technical field to which the present invention belongs. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more of the related listed types.

[0069] Example 1

[0070] Please refer to Figure 1 , which shows a control method for a vehicle thermal management system based on multi-dimensional data proposed in the first embodiment of the present invention. The method includes steps S10 to S12.

[0071] Step S10, during the vehicle driving process, multi-dimensional data related to the control of the vehicle's thermal management system is collected in real time.

[0072] Through the distributed sensor network of the vehicle thermal management system, multi-dimensional data under the vehicle operating state is collected in real time. Among them, the multi-dimensional data includes the average temperature of the battery pack cells, the output current of the battery pack, and the output voltage of the battery pack related to the power battery system; the light intensity, the opening degree of the external circulation air damper, and the ambient temperature related to the environment; and parameters such as the average temperature of the passenger compartment. During the data collection process, the sampling time series length and the sampling frequency can be set according to the actual situation, and a high-precision time stamp is used to ensure data synchronization.

[0073] Step S11, establish a neural network architecture composed of a TCN encoder, a Transformer encoder, and a Cross attention decoder, and obtain a hybrid deep learning model obtained by training the model with the neural network architecture.

[0074] Among them, the TCN encoder is used to extract multi-scale time series features through multi-layer causal convolution and different dilation coefficients of multi-layer causal convolution. The Transformer encoder is used to capture long-term dependencies in the features through the multi-head self-attention mechanism. The Cross attention decoder is used to calculate the attention weights between the features output by the TCN encoder and the Transformer encoder to achieve adaptive fusion of multi-source information.

[0075] Specifically, as Figure 2 shown, the neural network architecture includes a TCN encoder, a Transformer encoder, and a Cross attention decoder. The features output after the TCN encoder and the Transformer encoder are adaptively fused in the Cross attention decoder and finally the predicted data is output.

[0076] More specifically, as Figures 3 to 5 shown, the TCN encoder includes a first-layer causal convolution layer and a second-layer causal convolution layer. The first-layer causal convolution layer uses 16 convolutional kernels, the dilation coefficient is 1, and the convolutional kernel size is 3×1; the second-layer causal convolution layer uses 32 convolutional kernels, the dilation coefficient is 2, and the convolutional kernel size is 3×1;

[0077] The TCN encoder receives the input features, performs causal convolution processing on the input features using the first-layer causal convolution layer, and performs residual connection with the input features to obtain preliminary input features;

[0078] The preliminary input features are sequentially processed by Layer Normalization and the ReLU activation function, and the obtained target input features are subjected to causal convolution processing using the second-layer causal convolution layer and residual connection with the target input features to obtain the final input features.

[0079] The final input features are sequentially processed by Layer Normalization and the ReLU activation function to obtain the output features of the TCN encoder to achieve multi-scale time series feature extraction.

[0080] The TCN encoder adopts a multi-layer causal convolution structure, and each layer contains a residual connection to alleviate the problem of gradient disappearance. Since various different-dimensional parameters need to be considered to enhance the data feature extraction ability, a double non-linear transformation is added to the TCN encoder. ; To achieve automatic feature selection, noise suppression, and temporal dependence modeling, and improve the adaptability of the model, learnable parameters in the adaptive weight mechanism are introduced. ; To achieve dynamic balance of model complexity, effectively prevent overfitting, improve the generalization ability of the model, and enhance the sparse representation ability of features, a regularization strength control mechanism is introduced. .

[0081] Therefore, the processing formula for causal convolution processing in the TCN encoder is:

[0082] ;

[0083] Among them, is the output function, is the input data at time point t, is the time window size, is the weight system, used to adjust the importance of data at different time points, is the step size, used to control the sampling interval, means pushing forward time units, represents historical data points, is the parameterized transformation function, used to perform non-linear transformation on the input data, is the sigmoid function, is the regularization coefficient, used to balance the model complexity, is the set of regularization functions, represents the number of regularization terms, represents the l-th regularization function.

[0084] In the Transformer encoder part, the Transformer encoder includes positional encoding, multi-head attention mechanism, and feed-forward neural network;

[0085] The Transformer encoder receives the input features. After positional encoding, the features generated by the multi-head attention mechanism are connected in residual with the features after positional encoding to obtain the preliminary input features;

[0086] The preliminary input features are processed by Layer Normalization and activation function, and the features generated by the feed-forward neural network are connected in residual with the features after being processed by Layer Normalization and activation function to obtain the target input features;

[0087] The target input features are processed by Layer Normalization and activation function to obtain the output features of the Transformer encoder.

[0088] Among them, considering that the influence degrees of different input variables on the output vary, in order to achieve dynamic weight allocation for different attention heads, enabling the model to automatically adjust the focus of attention according to the input data and task characteristics, thereby enhancing the pertinence of feature extraction and the generalization ability of the model, a learnable weight coefficient is introduced. To enhance the expressive ability and flexibility of the attention mechanism, a bias matrix is introduced. provides additional learnable parameters for the attention mechanism; to enable different heads to learn and focus on different feature patterns, thereby enhancing the model's ability to capture diverse features, an independent query transformation matrix is introduced. and an output transformation matrix. .

[0089] Therefore, the expression of the multi-head attention mechanism of the Transformer encoder is:

[0090] ;

[0091] where Q, K, and V represent the query matrix, key matrix, and value matrix respectively, d is the dimension of the attention head, R is the number of attention heads, is the weight coefficient of the r-th attention head; is the query transformation matrix of the r-th attention head; is the bias matrix of the r-th attention head; is the output transformation matrix of the r-th attention head, and T is the matrix transpose operation.

[0092] The computational expression of the feed-forward neural network is:

[0093] ;

[0094] where W1 = [64, 256], W2 = [256, 64], b1 =

[256] , b2 =

[64] .

[0095] The Cross attention decoder includes a cross-attention layer, a feed-forward neural network, and an activation function layer. The Cross attention decoder receives the output features of the TCN encoder and the Transformer encoder;

[0096] The cross-attention layer, the feed-forward neural network, and the activation function layer are sequentially used for feature extraction to obtain the output features of the Cross attention decoder:

[0097] Among them, residual connections are set in both the cross-attention layer, the feed-forward neural network, and the activation function layer.

[0098] The computational expression of the cross-attention layer is:

[0099] ;

[0100] Among them, Q comes from the decoder, and K and V come from the outputs of the TCN encoder and the Transformer encoder. In a specific implementation, the calculation of the cross-attention layer satisfies the following formula:

[0101] Cross-attention layer calculation: ;

[0102] Feed-forward network calculation: ;

[0103] Output layer calculation: ;

[0104] Among them, W1 = [64, 256], W2 = [256, 64], b1 =

[256] , b2 =

[64] , X is the input feature, Multi Head represents the multi-head attention operation, and Layer Norm is the layer normalization.

[0105] FFN is the feed-forward neural network. The model adopts the dropout mechanism to prevent overfitting and uses a linear activation function in the final output layer to generate the prediction result. This architecture design can not only effectively capture local temporal features but also model long-term dependencies through the self-attention and cross-attention mechanisms, thereby improving the prediction accuracy of the control parameters of the thermal management system. In addition, to improve the generalization ability of the model, the Dropout mechanism is added between key layers to prevent overfitting, and residual connections are used to ensure the training stability of the deep network.

[0106] In step S12, the multi-dimensional data is input into the hybrid deep learning model to obtain corresponding control information, and the thermal management system of the vehicle is controlled according to the control information. The control information includes the opening degree of the electronic expansion valve of the battery pack cooling system, the opening degree of the electronic expansion valve of the occupant compartment air conditioning system, and the compressor speed.

[0107] Specifically, the hybrid deep learning model masters the internal logic of generating control information based on multi-dimensional data. By inputting the real-time data collected by various vehicle sensors, such as the average temperature of the battery pack cells, the output current of the battery pack, the output voltage of the battery pack, the light intensity, the opening degree of the external circulation air damper, the ambient temperature, and the average temperature of the occupant compartment, into the hybrid deep learning model, the opening degree of the electronic expansion valve of the battery pack cooling system, the opening degree of the electronic expansion valve of the occupant compartment air conditioning system, and the compressor speed can be obtained.

[0108] Among them, in an optional embodiment of the present invention, before the multi-dimensional data is input into the hybrid deep learning model, the data is preprocessed to improve the prediction accuracy. For example, data cleaning and enhancement processing are first performed to achieve data outlier detection and processing, missing value filling, and data smoothing. Then, multi-modal data normalization processing is performed on the processed data to eliminate the influence of different dimensions and magnitudes and enhance the comparability of the data. After obtaining the corresponding predicted control information, at this time the data is a normalized value, and the demand needs to be inversely converted to the real value to obtain the final opening of the electronic expansion valve of the battery pack cooling system, the opening of the electronic expansion valve of the passenger compartment air conditioning system, and the compressor speed.

[0109] Exemplarily, when combined with the specific application scenario in the automotive field, the trained model parameters and structure can be converted into a format suitable for in-vehicle controllers, such as the TensorRT format, and the model can be deployed to the electronic control unit of the new energy vehicle thermal management system through the CAN bus communication protocol. In the actual deployment process, combined with the fault diagnosis and anomaly detection mechanism, it is ensured that the model can make stable and reliable real-time control decisions under various working conditions, realizing the precise adjustment of the opening of the electronic expansion valve of the battery pack cooling system, the opening of the electronic expansion valve of the passenger compartment air conditioning system, the electronic fan speed, and the compressor speed, thereby optimizing the performance of the vehicle's thermal management system.

[0110] In summary, the control method of the vehicle thermal management system based on multi-dimensional data in the above embodiments of the present invention is designed with a hybrid architecture of a TCN encoder, a Transformer encoder, and a Cross attention decoder. This architecture realizes multi-scale time series feature extraction through multi-layer causal convolutions and different dilation coefficients of multi-layer causal convolutions, uses the multi-head self-attention mechanism of the Transformer encoder to capture the long-term dependence relationship of features, and combines the Cross attention decoder to calculate the attention weights between the features output by the TCN encoder and the Transformer encoder, realizing the adaptive fusion of multi-source information, effectively solving the limitations of traditional methods in processing long sequence data and the lack of in-depth mining of time series features and other defects, improving the prediction accuracy of the model, and further significantly improving the control accuracy of the new energy vehicle thermal management system. It solves the problems of insufficient data processing ability of the control model of the vehicle thermal management system in the prior art and inaccurate control caused by poor multi-source data fusion effect.

[0111] Example 2

[0112] This embodiment also proposes a control method for a vehicle thermal management system based on multi-dimensional data. The difference between the control method for a vehicle thermal management system based on multi-dimensional data in this embodiment and the control method for a vehicle thermal management system based on multi-dimensional data proposed in Embodiment 1 is as follows:

[0113] Obtain a training data set composed of historical multi-dimensional data and corresponding control information. First, perform data cleaning and augmentation, outlier detection and handling, missing value filling, and data smoothing on the collected training data set. Then, perform multi-modal data standardization processing, and divide the training data set into a training set, a test set, and a validation set according to a preset ratio;

[0114] Use the training set to train the neural network architecture model, and use the validation set to perform systematic hyperparameter tuning on the trained hybrid deep learning model;

[0115] Use the pre-divided test set to evaluate the optimized hybrid deep learning model, evaluate the deviation degree between the predicted value and the actual value, and verify the robustness and generalization ability of the model.

[0116] Among them, first perform data cleaning and augmentation on the collected original training data set to detect and handle outliers, fill in missing values, and smooth the data. Then perform multi-modal data standardization processing on the processed data to eliminate the influence of different dimensions and magnitudes, enhance the comparability of the data. The standardization formula uses the min-max processing method to ensure that the data is within a unified range, which is convenient for subsequent model training and prediction. The formula for standardization processing is:

[0117] ;

[0118] Among them, is the normalized value, is the original value, is the maximum value in the original value, is the minimum value in the original value.

[0119] Perform time window segmentation processing on the standardized data, using 1 second as the window length. This refined time granularity design helps the TCN encoder to better capture local temporal features, and at the same time enables the Transformer encoder to more effectively learn long-range dependencies.

[0120] In terms of dataset division, a standard three - segment division strategy is adopted: the overall data is sequentially divided into a training set, a validation set, and a test set according to the ratio of 50% - 25% - 25%. Specifically, the first 50% of the dataset is used as the training set for parameter learning and optimization of the model. The focus is on training the convolutional layer parameters of the TCN encoder, the multi - head self - attention mechanism parameters of the Transformer encoder, and the cross - attention weights of the Cross - attention decoder. The middle 25% of the data is used as the validation set for performance evaluation and hyperparameter tuning during the model training process, especially for optimizing key hyperparameters such as the dilation coefficient of the TCN encoder and the number of attention heads of the Transformer encoder. The last 25% of the data is used as the test set to evaluate the generalization ability and actual prediction effect of the model.

[0121] For the training implementation of the hybrid deep - learning model (the hybrid model of TCN encoder - Transformer encoder - Cross - attention decoder), this embodiment provides multiple technical path options: it can be implemented through self - programming in MATLAB or Python, or a mature deep - learning framework such as MindSpore can be selected.

[0122] In this embodiment, considering the rich ecosystem of deep - learning frameworks and the native support for the attention mechanism, the PyTorch framework based on Python is selected for model training and optimization. This implementation method can not only make full use of the dynamic computing characteristics and GPU acceleration ability of PyTorch, but also conveniently implement complex attention mechanisms and causal convolution operations, and at the same time facilitate flexible adjustment and optimization of the model structure.

[0123] The hyperparameters of the trained model are adjusted using the validation set data. In terms of training parameters, the TCN encoder, the Transformer encoder, and the Cross - attention decoder all adopt a Dropout rate of 0.2, the initial learning rate is set to 0.001, combined with a warmup strategy of 500 steps, and the batch size is set to 128. To prevent overfitting, an early - stopping mechanism is introduced, and training stops when the improvement of the validation set loss does not exceed 10 -4 during 10 consecutive epochs.

[0124] In addition, in terms of model training optimization, to ensure prediction accuracy, the mean squared error (MSE) is used as the loss function. Early stopping is adopted to avoid overfitting, and grid search combined with Bayesian optimization is used for parameter search. The MSE on the validation set is used as the evaluation metric. To prevent overfitting, an early stopping mechanism is introduced, and training stops when the validation set loss does not improve for multiple consecutive epochs. Through this systematic hyperparameter tuning process, it is ensured that the model can achieve optimal prediction accuracy while maintaining good generalization ability.

[0125] After model training, the change trend of the loss function is as Figure 6 shown. From the training process, the hybrid model of the TCN encoder, Transformer encoder, and Cross attention decoder shows good convergence performance on both the training set and the validation set. The loss function shows a steady downward trend, and there is no obvious overfitting phenomenon, which verifies the effectiveness of the regularization mechanism with a dropout rate of 0.2 in the model.

[0126] The verification results on the test set are as Figure 7 、 Figure 8 、 Figure 9 shown. The results show that the hybrid model of the TCN encoder, Transformer encoder, and Cross attention decoder exhibits excellent prediction performance. Among them, the TCN encoder effectively captures the local temporal features of parameter changes through a multi-layer causal convolution structure, while the multi-head self-attention mechanism of the Transformer encoder successfully models long-term dependencies. The two achieve effective feature fusion through the cross-attention mechanism of the Cross attention decoder. This deeply integrated architecture enables the model to accurately predict key control parameters such as the opening degree of the electronic expansion valve of the battery pack cooling system, the opening degree of the electronic expansion valve of the occupant compartment air conditioning system, and the compressor speed.

[0127] From Figure 10 the comparison of the occupant compartment temperature changes under different driving cycle conditions shown, it can be seen that the technical solution of the embodiment of the present invention has a better temperature control effect compared with the technical solution of the prior art. Under each condition, the temperature control curve of the technical solution of the embodiment of the present invention is smoother, especially in the stage of rapid temperature change, showing stronger response speed and control stability. From Figure 11 the comparison results of COP under different conditions shown, it can be seen that the control strategy of the embodiment of the present invention has significant advantages compared with the existing LSTM method.

[0128] In summary, the control method of the vehicle thermal management system based on multi-dimensional data in the above embodiments of the present invention is designed with a hybrid architecture of a TCN encoder, a Transformer encoder, and a Cross attention decoder. This architecture realizes multi-scale temporal feature extraction through multi-layer causal convolution and different dilation coefficients of multi-layer causal convolution, uses the multi-head self-attention mechanism of the Transformer encoder to capture the long-term dependence of features, and combines the Cross attention decoder to calculate the attention weights between the features output by the TCN encoder and the Transformer encoder, realizing the adaptive fusion of multi-source information. It effectively solves the limitations of traditional methods in processing long-sequence data and the lack of in-depth mining of temporal features, improves the accuracy of model prediction, and thus significantly improves the control accuracy of the new energy vehicle thermal management system. It solves the problem in the prior art that the data processing ability of the control model of the vehicle thermal management system is insufficient and the control is inaccurate due to the poor multi-source data fusion effect.

[0129] Example 3

[0130] Please refer to Figure 12 , which shows the control system of the vehicle thermal management system based on multi-dimensional data proposed in the third embodiment of the present invention. The system includes:

[0131] The acquisition module 100 is used to collect multi-dimensional data related to the control of the vehicle's thermal management system in real time during the vehicle's driving process;

[0132] The training module 200 is used to establish a neural network architecture composed of a TCN encoder, a Transformer encoder, and a Cross attention decoder, and obtain a hybrid deep learning model obtained by training the model with the neural network architecture;

[0133] The control module 300 is used to input the multi-dimensional data into the hybrid deep learning model to obtain corresponding control information, and control the vehicle's thermal management system according to the control information. The control information includes the opening degree of the electronic expansion valve of the battery pack cooling system, the opening degree of the electronic expansion valve of the occupant compartment air conditioning system, the rotation speed of the electric fan, and the rotation speed of the compressor;

[0134] Among them, the TCN encoder is used to realize multi-scale temporal feature extraction through multi-layer causal convolution and different dilation coefficients of multi-layer causal convolution. The Transformer encoder is used to capture the long-term dependence among features through the multi-head self-attention mechanism. The Cross attention decoder is used to calculate the attention weights between the features output by the TCN encoder and the Transformer encoder, realizing the adaptive fusion of multi-source information.

[0135] The functions or operation steps implemented when each of the above modules is executed are substantially the same as those in the above method embodiments, and will not be described in detail herein.

[0136] Example 4

[0137] On the other hand, the present invention also provides a readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the steps of the method described in any one of the above Embodiments 1 to 2 are implemented.

[0138] Example 5

[0139] On the other hand, the present invention also provides an electronic device, which includes a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the program, the steps of the method described in any one of the above Embodiments 1 to 2 are implemented.

[0140] The technical features of each of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered to be within the scope described in this specification.

[0141] Those skilled in the art can understand that the logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any storage medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can fetch and execute instructions from the instruction execution system, apparatus, or device), or used in combination with these instruction execution systems, apparatus, or devices. For the purposes of this specification, a "storage medium" can be any device that can contain, store, communicate, propagate, or transmit a program for use by or in combination with an instruction execution system, apparatus, or device.

[0142] More specific examples (nonexhaustive list) of the storage medium include the following: electrical connection parts (electronic devices) having one or more wirings, portable computer disk cartridges (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber devices, and portable compact disc read-only memory (CDROM). Additionally, the storage medium can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other media, then editing, interpreting, or processing it in other suitable ways when necessary, and then storing it in a computer memory.

[0143] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, any one or a combination of the following techniques well known in the art can be used: discrete logic circuits having logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits having suitable combinational logic gate circuits, programmable gate arrays (PGA), field programmable gate arrays (FPGA), etc.

[0144] In the description of this specification, the description with reference to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples", etc. means that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples.

[0145] The above-described embodiments merely represent several implementation manners of the present invention, and their descriptions are relatively specific and detailed, but should not be construed as limiting the scope of the patent of the present invention. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present invention, several modifications and improvements can still be made, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention should be subject to the appended claims.

Claims

1. A control method for a vehicle thermal management system based on multi-dimensional data, characterized in that The method includes: During the driving process of the vehicle, multi-dimensional data related to the control of the vehicle's thermal management system is collected in real time; A neural network architecture composed of a TCN encoder, a Transformer encoder, and a Cross attention decoder is established, and a hybrid deep learning model obtained by training the model with the neural network architecture is obtained; The multi-dimensional data is input into the hybrid deep learning model to obtain corresponding control information, and the vehicle's thermal management system is controlled according to the control information. The control information includes the opening degree of the electronic expansion valve of the battery pack cooling system, the opening degree of the electronic expansion valve of the occupant compartment air conditioning system, and the compressor speed; Among them, the TCN encoder is used to extract multi-scale time series features through multi-layer causal convolution and different dilation coefficients of multi-layer causal convolution. The Transformer encoder is used to capture long-term dependencies in the features through the multi-head self-attention mechanism. The Cross attention decoder is used to calculate the attention weights between the features output by the TCN encoder and the Transformer encoder to achieve adaptive fusion of multi-source information; The TCN encoder includes a first-layer causal convolution layer and a second-layer causal convolution layer. The first-layer causal convolution layer uses 16 convolutional kernels, the dilation coefficient is 1, and the convolutional kernel size is 3×1; the second-layer causal convolution layer uses 32 convolutional kernels, the dilation coefficient is 2, and the convolutional kernel size is 3×1; The TCN encoder receives multi-dimensional data, performs causal convolution processing using the first-layer causal convolution layer, and performs residual connection with the multi-dimensional data to obtain preliminary input features; The preliminary input features are sequentially processed through Layer Normalization and the ReLU activation function, and the obtained target input features are processed through causal convolution using the second-layer causal convolution layer and residual connection with the target input features to obtain the final input features, The final input features are sequentially processed through Layer Normalization and the ReLU activation function to obtain the output features of the TCN encoder, so as to achieve multi-scale time series feature extraction; The processing formula for causal convolution processing is: ； Among them, is the output function, is the input data at time point t, is the time window size, is the weight system, used to adjust the importance of data at different time points, is the step size, used to control the sampling interval, means pushing forward time units, represents historical data points, is the parameterized transformation function, used to perform non-linear transformation on the input data, is the sigmoid function, is the regularization coefficient, used to balance the model complexity, is the set of regularization functions, represents the number of regularization terms, represents the l-th regularization function; The Transformer encoder includes positional encoding, multi-head attention mechanism, and a first feed-forward neural network; The Transformer encoder receives multi-dimensional data, and after positional encoding, the features generated by the multi-head attention mechanism are residually connected with the features after positional encoding to obtain preliminary input features; The preliminary input features are processed through Layer Normalization and the activation function, and the features generated by the first feed-forward neural network are residually connected with the features after being processed through Layer Normalization and the activation function to obtain the target input features; The target input features are processed through Layer Normalization and the activation function to obtain the output features of the Transformer encoder; The expression of the multi-head attention mechanism of the Transformer encoder is: ； Among them, Q, K, and V represent the query matrix, key matrix, and value matrix respectively, is the dimension of the attention head, R is the number of attention heads, is the weight coefficient of the r-th attention head; is the query transformation matrix of the r-th attention head; is the bias matrix of the r-th attention head; is the output transformation matrix of the r-th attention head, and T is the matrix transpose operation. The calculation expression of the first feedforward neural network is as follows: ； where W1 = [64, 256], W2 = [256, 64], b1 = [256], b2 = [64]; The Cross attention decoder includes a cross-attention layer, a second feedforward neural network, and an activation function layer; The Cross attention decoder receives the output features of the TCN encoder and the Transformer encoder, and sequentially uses the cross-attention layer, the second feedforward neural network, and the activation function layer to extract features to obtain the output features of the Cross attention decoder; Among them, residual connections are set in the cross-attention layer, the second feedforward neural network, and the activation function layer.

2. The control method of the vehicle thermal management system based on multi-dimensional data according to claim 1, characterized in that, The calculation expression of the cross-attention layer is as follows: ； ； The calculation expression of the second feedforward neural network is as follows: ； The calculation expression of the activation function layer is as follows: ； Among them, W1 = [64, 256], W2 = [256, 64], b1 = [256], b2 = [64], from the Cross attention decoder, and the outputs from the TCN encoder and the Transformer encoder, represent the dimension of the attention head, X is the output features of the TCN encoder and the Transformer encoder received by the Cross attention decoder, MultiHead represents the multi-head attention operation, is layer normalization.

3. A control system for a vehicle thermal management system based on multi-dimensional data, characterized in that, A control method for a vehicle thermal management system based on multi-dimensional data according to claim 1, the system comprising: An acquisition module for real-time acquiring multi-dimensional data related to the control of the thermal management system of the vehicle during vehicle driving; A training module for establishing a neural network architecture composed of a TCN encoder, a Transformer encoder, and a Cross attention decoder, and obtaining a hybrid deep learning model obtained by model training of the neural network architecture; A control module for inputting the multi-dimensional data into the hybrid deep learning model to obtain corresponding control information, and controlling the thermal management system of the vehicle according to the control information, where the control information includes the opening degree of the electronic expansion valve of the battery pack cooling system, the opening degree of the electronic expansion valve of the occupant compartment air conditioning system, the rotation speed of the electric fan, and the rotation speed of the compressor; Among them, the TCN encoder is used to achieve multi-scale time series feature extraction through multi-layer causal convolution and different dilation coefficients of multi-layer causal convolution, the Transformer encoder is used to capture long-term dependencies in the features through the multi-head self-attention mechanism, and the Cross attention decoder is used to calculate the attention weights between the features output by the TCN encoder and the Transformer encoder to achieve adaptive fusion of multi-source information.

4. A readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, the steps of the method according to any one of claims 1 to 2 are implemented.

5. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored on the memory and running on the processor. When the processor executes the program, the steps of the method according to any one of claims 1 to 2 are implemented.

Images