Prediction model construction method and device for vehicle thermal management, control method and device and vehicle

Abstract

The invention relates to the technical field of vehicle thermal management, in particular to a prediction model construction method and device for vehicle thermal management, a control method and device and a vehicle, and the construction method comprises the steps: obtaining a historical sample data set of a vehicle thermal management system; preprocessing the historical sample data set; an initial prediction model is established based on a Transform network, the input of the initial prediction model is state data, the output of the initial prediction model is control data, and a loss function comprises a mean square error term and a physical constraint term; and taking the minimum loss function value as an optimization target, and performing iterative training optimization on the initial prediction model through the preprocessed historical sample data to obtain a prediction model of vehicle thermal management. According to the method, on the basis of strong nonlinear modeling capability of the Transform network, model parameters are optimized by introducing physical constraint terms, so that a model prediction result is ensured to be consistent with an actual physical rule, and a prediction error caused by violation of the physical rule is avoided, thereby improving the prediction precision.

Description

Technical Field

[0001] The present invention relates to the technical field of vehicle thermal management, and in particular to a prediction model construction method, a control method, a device and a vehicle for vehicle thermal management. Background Art

[0002] With the booming global automotive industry, and particularly the widespread adoption of new energy vehicles, automotive thermal management systems have become a key element in improving vehicle performance and reliability. The system's core goal is to optimize the temperature distribution of various vehicle components, ensuring they consistently operate within the optimal temperature range, thereby achieving the dual goals of reducing energy consumption and extending component life. For example, in electric vehicles, their thermal management systems dynamically adjust coolant flow and radiator operating status by real-time monitoring of multiple key parameters, including battery temperature, ambient temperature, vehicle speed, and power demand, to achieve precise temperature control while minimizing energy loss.

[0003] In traditional technology, parameter prediction for thermal management systems relies primarily on empirical formulas or physical modeling methods. However, when faced with complex and changing actual operating conditions, these methods exhibit significant limitations: they struggle to accurately capture the dynamic coupling relationships between system parameters, particularly when dealing with multidimensional input features and nonlinear output characteristics, leading to a significant decrease in prediction accuracy. Furthermore, traditional design processes rely heavily on manual intervention and extensive experimental verification, which not only leads to high development costs but also severely restricts the system's real-time responsiveness and flexible adaptability.

[0004] In recent years, machine learning techniques, particularly deep learning methods, have become a research hotspot for thermal management system parameter prediction. Time series models, such as recurrent neural networks (RNNs) and long short-term memory networks (LSTMs), have been widely used for time series data prediction. However, these models still have inherent flaws when dealing with multidimensional input features and complex nonlinear relationships. When faced with high-dimensional data, the computational complexity of the models increases exponentially, significantly reducing training efficiency. Furthermore, existing models generally lack physical constraints, which can lead to prediction results that violate actual physical laws, seriously affecting their accuracy and reliability. Summary of the Invention

[0005] The purpose of the present invention is to provide a prediction model construction method, control method, device and vehicle for vehicle thermal management, which can optimize model parameters by introducing physical constraints based on the powerful nonlinear modeling capabilities of the Transformer network, ensure that the model prediction results are consistent with the actual physical laws, avoid prediction errors caused by violations of physical laws, and thus improve prediction accuracy.

[0006] In order to achieve the above object, the technical solution adopted by the present invention is as follows:

[0007] In a first aspect, the present invention discloses a method for constructing a prediction model for vehicle thermal management, comprising:

[0008] Acquire a historical sample data set of the vehicle thermal management system, the historical sample data set including state data of the vehicle thermal management system and control data corresponding to the state data;

[0009] Preprocess the historical sample data set;

[0010] An initial prediction model is established based on a Transformer network. The input of the initial prediction model is state data, and the output is control data. The loss function includes a mean square error term and a physical constraint term. The mean square error term is the mean square error between the actual value of the control data and the predicted value of the control data. The physical constraint term is the absolute value of the error between the theoretical flow rate and the predicted flow rate of the vehicle thermal management system.

[0011] Taking the minimum loss function value as the optimization goal, the initial prediction model is iteratively trained and optimized through preprocessed historical sample data to obtain a prediction model for vehicle thermal management.

[0012] Furthermore, the loss function of the initial prediction model is L=α·MSE+β·E, where MSE is the mean square error term, α is the weight coefficient of the mean square error term, E is the physical constraint term, and β is the weight coefficient of the physical constraint term.

[0013] Furthermore, preprocessing the historical sample data set specifically includes: performing principal component analysis on the data set to achieve feature selection and data dimensionality reduction.

[0014] Furthermore, principal component analysis of the data set specifically includes:

[0015] Constructing an m×n-dimensional feature vector matrix of the historical sample data set, and normalizing all sample feature vectors of the historical sample data set to obtain a normalized feature vector matrix X of the historical sample data set;

[0016] Calculating the covariance matrix of the eigenvectors of the standardized historical sample data set, and performing eigenvalue decomposition on the covariance matrix to obtain a plurality of eigenvalues ​​and eigenvectors corresponding to the eigenvalues;

[0017] Sort the eigenvalues ​​from large to small and construct the projection matrix Y=(y1,y2,...,y k ) T , where y1,y2,...,y kThe eigenvalues ​​are the eigenvectors corresponding to the largest k eigenvalues ​​among the eigenvalues, where k is determined by a preset principal component percentage. The projection matrix Y is left-multiplied by the eigenvector normalization matrix X of the historical sample dataset to obtain a k×n dimensionality reduction vector YX.

[0018] Furthermore, the initial prediction model based on the Transformer network includes:

[0019] The input layer is used to convert the input data into an input vector of a preset length;

[0020] The position encoding layer is used to add position encoding to each input sequence element of the input vector to preserve the timing information;

[0021] The encoder layer is used to capture the dependencies between different positions in the input vector based on a multi-head self-attention mechanism and extract the feature information of the input vector based on a feedforward neural network;

[0022] The decoder layer is used to capture the dependencies within the decoder based on a multi-head self-attention mechanism and output the feature information of the input vector based on a feedforward neural network;

[0023] The output layer is used to receive the feature information of the input vector and generate the predicted value of the control data through linear transformation and activation function.

[0024] Furthermore, after preprocessing the historical sample data set, the preprocessed historical sample data set is divided into a training set and a test set according to a preset ratio. The training set is used to iteratively train and optimize the initial prediction model, and the test set is used to evaluate the performance of the trained prediction model.

[0025] Furthermore, the status data includes: vehicle status information, battery status information, electric drive status information, air conditioning status information, environmental status information and cooling status information.

[0026] Furthermore, the vehicle status information includes vehicle speed, driving mode, gear position, braking status, accelerator pedal opening and vehicle load;

[0027] The battery status information includes battery charging status, battery internal resistance, battery voltage, and battery current;

[0028] The electric drive status information includes electric drive speed, electric drive torque and electric drive current;

[0029] The air conditioning status information includes the air conditioning working status, air conditioning temperature and air conditioning air volume;

[0030] The environmental status information includes ambient temperature, ambient humidity, wind speed and altitude;

[0031] The cooling status information includes coolant temperature, coolant flow, coolant pump speed, cooling fan speed, radiator temperature, radiator pressure difference and water pump flow.

[0032] Furthermore, the control data includes control status data of each thermal management subsystem and temperature data of each thermal management component.

[0033] In a second aspect, the present invention discloses a method for controlling vehicle thermal management, comprising:

[0034] Obtain current status data of the vehicle thermal management system;

[0035] Inputting the current state data into a prediction model for vehicle thermal management for prediction to obtain prediction control data; the prediction model for vehicle thermal management is obtained using the above-mentioned method for constructing a prediction model for vehicle thermal management;

[0036] A vehicle thermal management system is controlled according to the predictive control data, the vehicle thermal management system including at least one of a passenger compartment thermal management system, an electric drive thermal management system, and a battery thermal management system.

[0037] In a third aspect, the present invention discloses a vehicle thermal management control device, comprising:

[0038] An acquisition module, used to obtain current status data of the vehicle thermal management system;

[0039] A prediction module, configured to input the current state data into a prediction model for vehicle thermal management to perform prediction and obtain prediction control data; the prediction model for vehicle thermal management is obtained using the above-mentioned method for constructing a prediction model for vehicle thermal management;

[0040] An adjustment module is used to control a vehicle thermal management system according to the predicted control data, wherein the vehicle thermal management system includes at least one of a passenger compartment thermal management system, an electric drive thermal management system, and a battery thermal management system.

[0041] In a fourth aspect, the present invention discloses a vehicle comprising the above-mentioned vehicle thermal management control device.

[0042] The present invention has the following unexpected beneficial effects:

[0043] 1. The vehicle thermal management prediction model construction method described in this invention incorporates a physical constraint term (the absolute value of the error between theoretical and predicted flow rates) into the loss function, forcing the model's predictions to conform to physical laws such as fluid dynamics. This avoids the deviation from reality caused by traditional purely data-driven models (such as LSTM and DNN) due to their disregard for physical rules. The physical constraint term and the mean squared error term together form the loss function, ensuring that the model adheres to physical laws while fitting the data, thereby enhancing the credibility of the prediction results.

[0044] 2. The vehicle thermal management prediction model construction method described in this invention is based on an initial prediction model established using a Transformer network. This network uses a self-attention mechanism to process multidimensional input features (such as vehicle speed, battery temperature, and coolant flow) in parallel. Compared to traditional recurrent neural networks or long-short-term memory networks, this network can more efficiently capture long-range dependencies and improve dynamic adaptability under complex operating conditions. Furthermore, the Transformer network, through its multi-head attention mechanism and positional encoding, can effectively process high-dimensional state data, avoiding the performance degradation of traditional neural networks due to excessive dimensionality.

[0045] 3. This invention solves the core problems of "physical unreliability" and "lack of dynamic adaptability" in traditional thermal management system predictions through the innovative architecture of "data-driven modeling + physical constraint optimization", and provides a solution for new energy vehicle thermal management systems that combines prediction accuracy, real-time performance and engineering reliability. It has significant technological breakthroughs and industrial application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the implementation or prior art description. Obviously, the drawings described below are only some embodiments of the present invention.

[0047] Figure 1 A flow chart illustrating an implementation of a method for constructing a prediction model for vehicle thermal management according to an embodiment of the present invention is shown.

[0048] Figure 2 A flow chart illustrating another implementation of the method for constructing a prediction model for vehicle thermal management according to an embodiment of the present invention is shown.

[0049] Figure 3 A schematic diagram of the process of preprocessing a historical sample data set according to an embodiment of the present invention is shown.

[0050] Figure 4 A schematic diagram of the structure of the initial prediction model established based on the Transformer network according to an embodiment of the present invention is shown.

[0051] Figure 5 A flow chart of a vehicle thermal management control method according to an embodiment of the present invention is shown.

[0052] Figure 6 A schematic structural diagram of a vehicle thermal management control device according to an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0053] The following describes the embodiments of the present invention with reference to the accompanying drawings and preferred embodiments. Those skilled in the art will readily appreciate the other advantages and benefits of the present invention from the disclosure herein. The present invention may also be implemented or applied through various other specific embodiments, and the various details in this specification may be modified or altered based on different viewpoints and applications without departing from the spirit of the present invention. It should be understood that the preferred embodiments are intended only to illustrate the present invention and are not intended to limit the scope of protection of the present invention.

[0054] In one embodiment, see Figure 1 As shown, the present invention provides a method for constructing a prediction model for vehicle thermal management, which includes the following steps:

[0055] S1. Acquire a historical sample dataset of the vehicle thermal management system. This dataset includes both state data and control data corresponding to the state data. By simultaneously collecting state data (e.g., temperature, pressure) and control data (e.g., pump speed, valve opening), the prediction model can learn the mapping relationship between system state and control strategy, providing a foundation for subsequent prediction of control parameters and avoiding model bias caused by missing features.

[0056] S2, preprocess the historical sample data set to remove noise, outliers and missing values ​​in the historical sample data set to prevent these defects from affecting the stability and accuracy of model training.

[0057] S3, establish an initial prediction model based on the Transformer network, the input of the initial prediction model is state data, the output is control data, and the loss function includes a mean square error term and a physical constraint term. The mean square error term is the mean square error between the true value of the control data and the predicted value of the control data, and the physical constraint term is the absolute value of the error between the theoretical flow and the predicted flow of the vehicle thermal management system.

[0058] The Transformer network's self-attention mechanism effectively captures long-range temporal dependencies in thermal management systems (such as the impact of coolant temperature changes on downstream component delays), outperforming traditional recurrent neural networks. Furthermore, compared to traditional recurrent neural networks, which must process sequences sequentially, the Transformer network can process inputs in parallel, significantly accelerating the training process and making it suitable for large datasets. The physical constraint term (flow error) in the loss function ensures that model predictions conform to the laws of thermodynamics, avoiding physically infeasible control recommendations and enhancing model reliability.

[0059] S4, with the minimum loss function value as the optimization goal, iteratively trains and optimizes the initial prediction model using preprocessed historical sample data to obtain a prediction model for vehicle thermal management. By minimizing the weighted sum of the mean square error and the physical constraint terms, the model maintains physical consistency while fitting historical data, achieving a balance between prediction accuracy and physical rationality. During the iterative training process, the model can automatically adjust parameters to adapt to the thermal management requirements under different operating conditions (such as the difference in cooling strategies between high-temperature and low-temperature environments). The optimized prediction model for vehicle thermal management can quickly output predicted values ​​of control parameters, meet the real-time response requirements of the vehicle thermal management system, and provide a decision-making basis for online control.

[0060] The vehicle thermal management prediction model construction method described in this paper incorporates a physical constraint term (the absolute value of the error between the theoretical and predicted flow rates) into the loss function, forcing the model's predictions to conform to physical laws such as fluid dynamics. This avoids the deviations from reality often seen in traditional data-driven models (such as LSTM and DNN) due to their neglect of physical rules. By combining the physical constraint term with the mean squared error term in the loss function, the model ensures that it adheres to physical laws while fitting the data, enhancing the credibility of the predictions.

[0061] The vehicle thermal management prediction model construction method described in this invention is based on an initial prediction model established by a Transformer network. This network uses a self-attention mechanism to process multidimensional input features (such as vehicle speed, battery temperature, and coolant flow) in parallel. Compared to traditional recurrent neural networks or long-short-term memory networks, this network can more efficiently capture long-range dependencies and improve dynamic adaptability under complex operating conditions. Furthermore, through a multi-head attention mechanism and positional encoding, the Transformer network can effectively process high-dimensional state data, avoiding the performance degradation of traditional neural networks due to excessive dimensionality.

[0062] Through the innovative architecture of "data-driven modeling + physical constraint optimization", this invention solves the core problems of "physical unreliability" and "lack of dynamic adaptability" in the prediction of traditional thermal management systems, and provides a solution for the thermal management system of new energy vehicles that combines prediction accuracy, real-time performance and engineering reliability. It has significant technological breakthroughs and industrial application value.

[0063] As a preferred embodiment of the present invention, the loss function of the initial prediction model is L=α·MSE+β·E, where MSE is the mean square error term, α is the weight coefficient of the mean square error term, E is the physical constraint term, and β is the weight coefficient of the physical constraint term.

[0064] The setting of the mean square error term (MSE) can directly optimize the model's fitting accuracy for historical control data, ensuring that the deviation between the predicted value and the actual value is minimized, and ensuring the prediction accuracy of the vehicle thermal management prediction model within the existing data range.

[0065] The physical constraint (E) ensures that model predictions conform to the principles of thermodynamics and fluid mechanics by limiting the error between theoretical and predicted flow rates. This means that the control parameters output by the model are within the system's safety range, reducing the risk of system failure. This prevents the prediction of cooling strategies that violate energy conservation, such as reducing temperature without energy input, and prevents unreasonable flow distribution, such as coolant flow exceeding the maximum capacity of the pump.

[0066] Furthermore, when historical data coverage is insufficient (e.g., there are few samples of extreme operating conditions), physical constraints can prevent the model from overfitting and improve its generalization capabilities in unseen scenarios. For example, in a high-temperature environment, even if the training data lacks corresponding operating conditions, physical constraints can still prevent the model from generating unreasonable cooling strategies. At the same time, physical constraints can filter out noise or incorrect annotations in historical data, allowing the model to learn more essential physical laws. For example, if a control parameter is recorded incorrectly in the historical data, physical constraints can prevent the model from learning the wrong mapping relationship.

[0067] The weight coefficients α and β adjust the relative importance of the mean square error term and the physical constraint term, allowing the model to find the optimal balance between "fitting historical data" and "adhering to physical laws." For example, when α < β, the prediction model is more likely to generate physically reasonable results. When α > β, the prediction model is closer to historical data.

[0068] Specifically, the mean square error term Where a is the sample size; M i is the true control data value of the i-th sample (such as actual coolant flow, fan speed, etc.), that is, the true value of the control data; is the predicted control data value of the prediction model for the i-th sample, that is, the control data predicted value.

[0069] Physical constraints b is the number of vehicle thermal management subsystems, E j is the absolute value of the error between the theoretical flow and the predicted flow of the jth vehicle thermal management subsystem, δ j is the physical constraint weight coefficient of the j-th vehicle thermal management subsystem.

[0070] in,

[0071] is the theoretical flow of the vehicle thermal management subsystem of the jth sample, and its calculation formula is:

[0072] Where, is the pipe geometry correction coefficient of the jth vehicle thermal management subsystem, is the temperature-viscosity coupling factor of the jth vehicle thermal management subsystem, is the turbulence effect compensation factor of the jth vehicle thermal management subsystem, η j is the fluid viscosity of the jth vehicle thermal management subsystem, L j is the pipe length of the jth vehicle thermal management subsystem, is the predicted pressure difference between the fluid inlet and outlet of the j-th vehicle thermal management subsystem predicted by the prediction model, ΔP t j is the actual pressure difference between the fluid inlet and outlet of the jth vehicle thermal management subsystem, r j is the pipe radius of the jth vehicle thermal management subsystem.

[0073] is the predicted flow of the vehicle thermal management subsystem for the jth sample, which is obtained through the prediction model output.

[0074] As a preferred embodiment of the present invention, preprocessing the historical sample data set specifically includes: performing principal component analysis (PCA) on the data set to achieve feature selection and data dimensionality reduction.

[0075] The raw data from a vehicle thermal management system may contain numerous features (e.g., temperature, pressure, flow, and other multi-sensor data). PCA can project these features into a low-dimensional space, preserving the principal components with the highest variance. For example, compressing 20-dimensional raw features into 5- to 10-dimensional principal components significantly reduces the computational complexity of model training.

[0076] Furthermore, PCA automatically filters out noise dimensions with low variance, improving the model's ability to capture valid signals. For example, low-variance components such as sensor measurement errors or environmental interference are naturally suppressed.

[0077] Furthermore, some features in a thermal management system may be highly correlated (such as coolant inlet and outlet temperatures). PCA generates uncorrelated principal components through orthogonal transformation to avoid model overfitting. For example, the original features "radiator inlet temperature" and "engine outlet temperature" may be strongly correlated. PCA transforms these into independent dimensions, eliminating feature redundancy.

[0078] Principal component analysis reduces the data dimension, making gradient calculations more efficient and significantly improving model training speed. Removing redundant features reduces model complexity and enhances generalization capabilities.

[0079] PCA preprocessing significantly improves the model's training efficiency and generalization capabilities through data dimensionality reduction, noise filtering, and feature de-redundancy, while retaining the key physical properties of the thermal management system and providing high-quality input for the subsequent efficient training of the Transformer network.

[0080] Further, see Figure 3 As shown in Figure 2, principal component analysis of a data set includes the following steps:

[0081] S21 , constructing an m×n dimensional feature vector matrix of the historical sample dataset, and normalizing all sample feature vectors of the historical sample dataset to obtain a historical sample dataset feature vector normalization matrix X. The mean of the historical sample dataset is set to 0, and the variance is set to 1.

[0082] Normalization can eliminate dimensionality effects, prevent variance overestimation due to differences in feature units (such as between temperature and pressure), and avoid ill-conditioned covariance matrices caused by differences in feature scales. Furthermore, the standardized data distribution is more conducive to gradient descent optimization and accelerates convergence.

[0083] S22, calculating the covariance matrix of the eigenvectors of the standardized historical sample data set, and performing eigenvalue decomposition on the covariance matrix to obtain a plurality of eigenvalues ​​and eigenvectors corresponding to the eigenvalues.

[0084] S23, sort the eigenvalues ​​from large to small, and construct the projection matrix Y=(y1,y2,...,y k ) T , where y1,y2,...,y k is the eigenvector corresponding to the largest k eigenvalues ​​among the eigenvalues, where k is determined by a preset principal component percentage.

[0085] S24 , multiplying the projection matrix Y by the normalized matrix X of the characteristic vectors of the historical sample dataset on the left to obtain a k×n dimensionality reduction vector YX.

[0086] As a preferred embodiment of the present invention, see Figure 4 As shown in Figure 2, the initial prediction model based on the Transformer network includes:

[0087] S31, the input layer, is used to convert input data into an input vector of a preset length. That is, input data of any length (such as sensor time series data) is converted into a fixed-length vector sequence through a sliding window or padding / truncation.

[0088] This setup unifies the input format, adapts to the Transformer network's requirement for fixed-length sequences, and supports batch training. By adjusting the length of the vector sequence, it is possible to capture both short-term (such as second-level temperature changes) and long-term (such as trip-level heat accumulation) dependencies.

[0089] S32, a position encoding layer, is used to add position encoding to each input sequence element of the input vector to preserve the timing information.

[0090] Specifically, for each input vector x at position i i Add position encoding PE i , usually using sine / cosine functions:

[0091] Where d is the vector dimension and pos is the position index.

[0092] Positional encoding compensates for the Transformer's self-attention mechanism's insensitivity to sequence order, preserving the temporal relationships between parameters such as temperature and flow. Compared to fixed-position embedding, sinusoidal encoding allows the model to handle sequence lengths not seen during training.

[0093] S33, the encoder layer, is used to capture the dependencies between different positions in the input vector based on the multi-head self-attention mechanism and extract the feature information of the input vector based on the feedforward neural network.

[0094] The multi-head self-attention mechanism can simultaneously focus on elements at any position in the sequence, realizing multi-scale feature fusion, such as the correlation between coolant temperature and engine load 5 minutes ago, and the impact of compressor start and stop on condenser pressure 10 seconds later.

[0095] S34, the decoder layer, is used to capture the dependencies within the decoder based on a multi-head self-attention mechanism and output the feature information of the input vector based on a feedforward neural network.

[0096] S35, the output layer, is used to receive the feature information of the input vector and generate the predicted value of the control data through linear transformation and activation function.

[0097] This architecture leverages the powerful expressive power of the Transformer network to provide a highly accurate and interpretable prediction model for vehicle thermal management systems, making it particularly suitable for handling complex timing dependencies and multivariate interaction problems.

[0098] As a preferred embodiment of the present invention, after preprocessing the historical sample data set, the preprocessed historical sample data set is divided into a training set and a test set according to a preset ratio, the training set is used to iteratively train and optimize the initial prediction model, and the test set is used to evaluate the performance of the trained prediction model.

[0099] The test set data is not used in model training and can truly reflect the model's predictive capabilities in unknown scenarios. For example, the training set contains urban road conditions data, while the test set contains highway conditions data to verify the model's adaptability to different driving scenarios.

[0100] Setting the test set can also prevent overfitting. For example, if the model performs well on the training set but has significant errors on the test set, it indicates overfitting and needs to be optimized by adjusting regularization parameters, reducing model complexity, etc.

[0101] The model is trained on the training set and the validation set (which can be split from the training set) is used to tune hyperparameters (such as the number of Transformer layers, the number of attention heads, and the learning rate) to avoid data leakage in the test set that affects the objectivity of the evaluation.

[0102] For example, the model training process sets the epoch to 150, the initial learning rate lr = 0.001, and the Adam optimizer is used to calculate the gradient, with its parameter betas = (0.9, 0.999).

[0103] As a preferred embodiment of the present invention, see Figure 2 As shown, the status data includes: vehicle status information, battery status information, electric drive status information, air conditioning status information, environmental status information and cooling status information.

[0104] This preferred implementation constructs a digital twin foundation for the vehicle thermal management system through systematic state data collection. The fusion of multi-dimensional data not only enhances the model’s ability to model complex thermal interactions, but also enables it to dynamically respond to environmental changes, driving behavior, and system status, providing comprehensive information support for achieving efficient, safe, and energy-saving thermal management control.

[0105] Furthermore, the vehicle status information includes vehicle speed, driving mode, gear position, braking status, accelerator pedal opening and vehicle load;

[0106] The battery status information includes battery charging status, battery internal resistance, battery voltage, and battery current;

[0107] The electric drive status information includes electric drive speed, electric drive torque and electric drive current;

[0108] The air conditioning status information includes the air conditioning working status, air conditioning temperature and air conditioning air volume;

[0109] The environmental status information includes ambient temperature, ambient humidity, wind speed and altitude;

[0110] The cooling status information includes coolant temperature, coolant flow, coolant pump speed, cooling fan speed, radiator temperature, radiator pressure difference and water pump flow.

[0111] This preferred implementation transforms the complexity of the physical world into computable data space by providing diversified state information. This allows the prediction model to capture the microscopic mechanisms within a single system (e.g., a battery) (e.g., heat generation characteristics at different SOCs) and also achieves horizontal collaboration, namely modeling the coupling relationships between multiple systems (e.g., electric drive, air conditioning, cooling) (e.g., energy flow from waste heat recovery). Through the continuous accumulation of multi-dimensional data, the model is intelligently upgraded from "rules-based" to "data-driven," ultimately improving prediction accuracy, reducing system energy consumption, and extending component life.

[0112] As a preferred embodiment of the present invention, see Figure 2 As shown, the control data includes control status data of each thermal management subsystem and temperature data of each thermal management component.

[0113] The control status data of the thermal management subsystem (such as valve opening, water pump speed, and fan power) is the direct output of the prediction model. It corresponds to the execution layer instructions of the thermal management system and is used to actively adjust the operating status of each subsystem. For example, increasing the coolant pump speed to improve heat dissipation capacity.

[0114] The temperature data of each thermal management component (such as battery temperature, electric drive temperature, and air conditioning outlet temperature) is an indirect output, reflecting the actual effect of the control command and used to evaluate the effectiveness of the control strategy. For example, after the water pump speed is increased, it is determined whether the battery temperature drops as expected.

[0115] This hierarchical design conforms to the "dual-loop control" concept in modern control theory. The inner loop is actuator control, and the outer loop is state feedback, ensuring that the model output not only meets real-time control requirements but also optimizes long-term control strategies through temperature feedback.

[0116] Control state data must satisfy fluid dynamics constraints (such as the linear relationship between pump speed and flow rate), while component temperature data must conform to thermodynamic laws (such as the dynamic balance between heat generation and heat dissipation). By simultaneously predicting both types of data, cross-dimensional verification of physical constraints is possible: if a control instruction (such as increasing fan speed) does not result in the expected temperature drop, the model can identify potential anomalies (such as a clogged radiator) and force a correction of the prediction deviation through the physical constraint term in the loss function.

[0117] In the Transformer network architecture, the control state data of the thermal management subsystem and the temperature data of the thermal management components can be modeled through independent output heads.

[0118] The control head uses continuous value regression, for example, the pump speed is a floating point value of 0-5000 rpm. The activation function uses a linear function to preserve the actual meaning of the physical quantity.

[0119] The temperature header uses interval constraint regression, for example: the battery temperature is limited to -30℃~60℃.

[0120] This decoupled modeling approach can avoid gradient update conflicts caused by dimensional differences between the two types of data (such as rotational speed vs. temperature), thereby improving training stability.

[0121] By comparing the correlation between control commands and temperature responses—for example, the magnitude of the battery temperature drop when the water pump speed increases by 10%—the control efficiency of each actuator can be analyzed, providing a basis for parameter adjustment. For example, if the radiator efficiency of a certain vehicle model is insufficient, the weighting of fan speed can be increased. Furthermore, when the deviation between the predicted and actual temperature exceeds a threshold, the corresponding control command can be retrospectively verified to determine if insufficient heat dissipation is caused by a valve opening that does not meet expectations. This helps locate faults in software algorithms or hardware systems.

[0122] In one embodiment, the present invention discloses a method for controlling vehicle thermal management, see Figure 5 As shown, the control method includes the following steps:

[0123] S10, obtaining current status data of the vehicle thermal management system.

[0124] S20: Input the current state data into a prediction model for vehicle thermal management to perform prediction and obtain prediction control data. The prediction model for vehicle thermal management is obtained using the method for constructing a prediction model for vehicle thermal management described in any of the above embodiments.

[0125] S30, controlling a vehicle thermal management system according to the predictive control data, the vehicle thermal management system including at least one of a passenger compartment thermal management system, an electric drive thermal management system, and a battery thermal management system.

[0126] In one embodiment, the present invention discloses a vehicle thermal management control device, see Figure 6 As shown, the control device 100 includes an acquisition module 200, a prediction module 300, and an adjustment module 400. The acquisition module 200 is used to obtain the current state data of the vehicle thermal management system. The prediction module 300 is used to input the current state data into the prediction model of the vehicle thermal management for prediction to obtain prediction control data; the prediction model of the vehicle thermal management is obtained using the vehicle thermal management prediction model construction method described in any of the above embodiments. The adjustment module 400 is used to control the vehicle thermal management system according to the prediction control data, and the vehicle thermal management system includes at least one of the passenger compartment thermal management system, the electric drive thermal management system, and the battery thermal management system.

[0127] In one embodiment, the present invention discloses a vehicle, which includes the above-mentioned vehicle thermal management control device.

[0128] The above embodiments are only preferred embodiments for fully illustrating the present invention, and the protection scope of the present invention is not limited thereto. Any equivalent substitution or modification made by those skilled in the art based on the present invention is within the protection scope of the present invention.

Claims

1. A method for constructing a prediction model for vehicle thermal management, characterized in that: include: Acquire a historical sample data set of the vehicle thermal management system, the historical sample data set including state data of the vehicle thermal management system and control data corresponding to the state data; Preprocess the historical sample data set; An initial prediction model is established based on a Transformer network. The input of the initial prediction model is state data, and the output is control data. The loss function includes a mean square error term and a physical constraint term. The mean square error term is the mean square error between the actual value of the control data and the predicted value of the control data. The physical constraint term is the absolute value of the error between the theoretical flow rate and the predicted flow rate of the vehicle thermal management system. Taking the minimum loss function value as the optimization goal, the initial prediction model is iteratively trained and optimized through preprocessed historical sample data to obtain a prediction model for vehicle thermal management.

2. The method for constructing a prediction model for vehicle thermal management according to claim 1, characterized in that: The loss function of the initial prediction model is L=α·MSE+β·E, where MSE is the mean square error term, α is the weight coefficient of the mean square error term, E is the physical constraint term, and β is the weight coefficient of the physical constraint term.

3. The method for constructing a prediction model for vehicle thermal management according to claim 1, characterized in that: The preprocessing of historical sample data sets specifically includes: performing principal component analysis on the data sets to achieve feature selection and data dimensionality reduction.

4. The method for constructing a prediction model for vehicle thermal management according to claim 3, characterized in that: Principal component analysis of a data set specifically includes: Constructing an m×n-dimensional feature vector matrix of the historical sample data set, and normalizing all sample feature vectors of the historical sample data set to obtain a normalized feature vector matrix X of the historical sample data set; Calculating the covariance matrix of the eigenvectors of the standardized historical sample data set, and performing eigenvalue decomposition on the covariance matrix to obtain a plurality of eigenvalues ​​and eigenvectors corresponding to the eigenvalues; Sort the eigenvalues ​​from large to small and construct the projection matrix Y=(y1,y2,...,y k ) T , where y1,y2,...,y k The eigenvalues ​​are the eigenvectors corresponding to the largest k eigenvalues ​​among the eigenvalues, where k is determined by a preset principal component percentage. The projection matrix Y is left-multiplied by the eigenvector normalization matrix X of the historical sample dataset to obtain a k×n dimensionality reduction vector YX.

5. The method for constructing a prediction model for vehicle thermal management according to claim 1, characterized in that: The initial prediction model based on the Transformer network includes: The input layer is used to convert the input data into an input vector of a preset length; The position encoding layer is used to add position encoding to each input sequence element of the input vector to preserve the timing information; The encoder layer is used to capture the dependencies between different positions in the input vector based on a multi-head self-attention mechanism and extract the feature information of the input vector based on a feedforward neural network; The decoder layer is used to capture the dependencies within the decoder based on a multi-head self-attention mechanism and output the feature information of the input vector based on a feedforward neural network; The output layer is used to receive the feature information of the input vector and generate the predicted value of the control data through linear transformation and activation function.

6. The method for constructing a prediction model for vehicle thermal management according to claim 1, characterized in that: After preprocessing the historical sample data set, the preprocessed historical sample data set is divided into a training set and a test set according to a preset ratio. The training set is used to iteratively train and optimize the initial prediction model, and the test set is used to evaluate the performance of the trained prediction model.

7. The method for constructing a prediction model for vehicle thermal management according to claim 1, characterized in that: The status data includes: vehicle status information, battery status information, electric drive status information, air conditioning status information, environmental status information and cooling status information.

8. The method for constructing a prediction model for vehicle thermal management according to claim 7, characterized in that: The vehicle status information includes vehicle speed, driving mode, gear position, braking status, accelerator pedal opening and vehicle load; The battery status information includes battery charging status, battery internal resistance, battery voltage, and battery current; The electric drive status information includes electric drive speed, electric drive torque and electric drive current; The air conditioning status information includes the air conditioning working status, air conditioning temperature and air conditioning air volume; The environmental status information includes ambient temperature, ambient humidity, wind speed and altitude; The cooling status information includes coolant temperature, coolant flow, coolant pump speed, cooling fan speed, radiator temperature, radiator pressure difference and water pump flow.

9. The method for constructing a prediction model for vehicle thermal management according to claim 1, characterized in that: The control data includes control status data of each thermal management subsystem and temperature data of each thermal management component.

10. A method for controlling vehicle thermal management, characterized in that: include: Obtain current status data of the vehicle thermal management system; Inputting the current state data into a prediction model for vehicle thermal management for prediction to obtain prediction control data; the prediction model for vehicle thermal management is obtained using the method for constructing a prediction model for vehicle thermal management according to any one of claims 1 to 9; A vehicle thermal management system is controlled according to the predictive control data, the vehicle thermal management system including at least one of a passenger compartment thermal management system, an electric drive thermal management system, and a battery thermal management system.

11. A vehicle thermal management control device, characterized in that: include: An acquisition module, used to obtain current status data of the vehicle thermal management system; a prediction module, configured to input the current state data into a prediction model for vehicle thermal management to perform prediction and obtain prediction control data; the prediction model for vehicle thermal management is obtained using the method for constructing a prediction model for vehicle thermal management according to any one of claims 1 to 9; An adjustment module is used to control a vehicle thermal management system according to the predicted control data, wherein the vehicle thermal management system includes at least one of a passenger compartment thermal management system, an electric drive thermal management system, and a battery thermal management system.

12. A vehicle, characterized in that: Comprising the vehicle thermal management control device as claimed in claim 11.

Images