Power battery real-time heat flow field regulation and control method based on edge calculation

Abstract

The invention discloses a power battery real-time heat flow field regulation and control method based on edge calculation, and the method comprises the following steps: S1, collecting and preprocessing temperature data, and constructing a time sequence data set; s2; organizing a Euclidean heat data structure according to space + time; s3, performing diffusion mapping to obtain non-Euclidean manifold features; s4, constructing a thermal field graph structure model; s5, inputting the graph structure to the graph adaptive convolutional neural network model of the edge node; s6, training the model by using historical data; s7, optimizing model training by using a sharpness perception minimization algorithm; s8, predicting future heat flow distribution in real time; s9, generating a thermal management strategy instruction set; and S10, controlling an execution module to carry out real-time regulation and control. According to the invention, the accuracy of thermal anomaly prediction and the adaptive ability of the cooling strategy are significantly improved, and the method can be widely applied to an intelligent thermal management scene.

Description

Technical Field

[0001] The present invention relates to the technical field of edge computing, and in particular, to a method for real-time heat flow field regulation of a power battery based on edge computing. Background Art

[0002] With the rapid development of new energy vehicles, as the core energy unit, the thermal management problem during the operation of power batteries has become a key factor affecting the safety, endurance performance and system stability of the whole vehicle. During the charging and discharging, high-rate operation and complex ambient temperature of power batteries, problems such as heat accumulation and uneven temperature difference are extremely likely to occur. If not effectively regulated, it may lead to thermal runaway, capacity attenuation, and even serious safety accidents such as fires. Therefore, establishing an efficient and reliable power battery thermal management system to achieve real-time monitoring and regulation of the heat flow field is an important research direction for ensuring the safe operation of electric vehicles.

[0003] Currently, most of the mainstream battery thermal management strategies rely on static threshold rules or control strategies based on a centralized processing architecture. The technical path mainly includes: arranging temperature sensors in the battery pack, collecting sensor data through the battery management system and then transmitting it to the central controller, and the central processor judges whether to start the cooling or heating equipment based on a fixed strategy. Such solutions are relatively mature in engineering deployment, but there are still significant technical limitations. First, since the regulation basis mainly relies on fixed thresholds or empirical rules, it cannot fully adapt to the unevenness and complex time-variation of the internal heat distribution of the battery pack, resulting in low regulation accuracy. Second, the thermal management response often has a delay, and it is unable to prospectively predict local rapid temperature rise or potential thermal runaway, easily forming a thermal response lag and reducing the thermal safety of the system. Third, the centralized processing structure has bandwidth and computing power bottlenecks in the processing of large-scale sensor data. Especially under the condition of limited computing power at the vehicle end, it is difficult to support high-frequency and high-dimensional heat flow field data analysis.

[0004] In recent years, with the rapid development of edge computing technology, sinking computing tasks from the cloud to local nodes close to the data source (such as vehicle-mounted controllers or embedded edge chips) has become an effective means to improve the system response speed and processing efficiency. Some studies have tried to introduce edge computing architectures for power battery temperature perception and regulation. For example, a simple thermal model is embedded in the vehicle-mounted controller to perform local processing on sensor data. However, such solutions generally still use temperature grid modeling in the traditional Euclidean space and fail to fully exploit the topological structure and heat diffusion behavior of the temperature distribution in space. At the same time, most of these methods rely on static graphs or fixed sensor layout modeling and are difficult to dynamically reflect the time-varying nature of the actual heat conduction path. In addition, in the model training and parameter optimization stages, existing methods often ignore the requirements of the power battery heat flow prediction task for "model smoothness" and "high robustness" and do not adopt a perturbation sensitivity optimization strategy that matches the physical process, resulting in unstable model prediction performance in complex edge environments.

[0005] To address the above technical problems, the present invention provides a method for regulating the heat flow of a power battery that integrates diffusion mapping modeling, graph adaptive convolution prediction, and spatial attention optimization. This method first maps the power battery thermal sensor data to a non-Euclidean manifold space to fully exploit the intrinsic relationship of heat diffusion between nodes; then constructs a dynamic graph structure model with diffusion-embedded features and inputs it into a graph neural network deployed on edge nodes for multi-step heat flow prediction; and introduces a SAM optimization strategy enhanced by local spatial attention to improve the prediction sensitivity of the hot spot area through weighted perturbation. This solution not only achieves high-precision modeling and prediction of the battery heat flow dynamics but also has extremely high response real-time performance and deployment adaptability, significantly improving the forward-looking recognition and dynamic regulation capabilities of the battery thermal management system for abnormal states, and is suitable for the thermal safety guarantee requirements in various intelligent electric vehicles, energy storage systems, and high-power battery scenarios.

[0006] Compared with the prior art, the present invention has the following remarkable advantages: First, through the diffusion mapping algorithm, the feature transformation from the Euclidean temperature grid to the non-Euclidean manifold is completed, overcoming the defect of ignoring the spatial structure of temperature propagation in traditional methods; second, a heat flow field modeling framework that dynamically evolves over time is constructed through a graph adaptive convolution network, which can continuously perceive local changes and achieve multi-time-step prediction; third, a local spatial heat perturbation attention mechanism is innovatively introduced, combined with the sharpness-aware minimization optimization objective, making the model more robust to the perturbation environment and having stronger generalization ability; fourth, the overall system runs on edge nodes, enabling closed-loop prediction and control locally in the vehicle, significantly reducing the computational delay and improving the heat response speed. In summary, the present invention breaks through the bottleneck problems of the traditional power battery thermal management system, such as poor real-time performance, weak prediction ability, and rough regulation method, and proposes a new technical path for future intelligent battery management, with broad application prospects and great engineering value.

[0007] Therefore, how to provide a real-time heat flow field regulation method for power batteries based on edge computing is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0008] An object of the present invention is to propose a real-time heat flow field regulation method for power batteries based on edge computing. The present invention comprehensively utilizes the diffusion mapping algorithm, graph adaptive convolutional network and sharpness-aware minimization optimization strategy enhanced by local spatial attention, and details the whole process of realizing dynamic modeling, prediction and regulation of the battery heat flow field under the condition of edge node deployment. By mapping traditional Euclidean temperature data into non-Euclidean manifold space expression, combining graph structure modeling and dynamic graph reasoning mechanism, the perception accuracy of complex heat diffusion behavior is improved. At the same time, spatial sensitive perturbation optimization means are adopted to enhance the model stability and prediction ability. This method has the advantages of high prediction accuracy, fast response speed, strong adaptability and strong thermal safety guarantee ability, and is suitable for the battery thermal management requirements in intelligent electric vehicles and high-performance energy storage systems.

[0009] A real-time heat flow field regulation method for power batteries based on edge computing according to an embodiment of the present invention includes the following steps:

[0010] S1. Arrange temperature sensors in the power battery system, collect temperature data, perform preprocessing, and construct a temperature time series data set;

[0011] S2. Organize the temperature time series data set into a high-dimensional heat data structure in Euclidean space according to the spatial position encoding and time window partitioning method;

[0012] S3. Use the diffusion mapping algorithm to perform feature transformation on the Euclidean data, construct a non-Euclidean manifold space expression, and extract the hidden heat diffusion structure information in the data;

[0013] S4. Establish a heat field graph structure model according to the non-Euclidean representation, where the graph nodes represent the measurement points of the battery module, and the graph edges are calculated and generated according to the diffusion distance and heat correlation degree;

[0014] S5. Load the graph adaptive convolutional network model in the edge computing node deployed on the vehicle side, and use the heat field graph structure as the model input;

[0015] S6. Use the historical heat flow evolution data to train the adaptive convolutional network model so that it can learn the dynamic change characteristics of the heat flow field on the graph structure;

[0016] S7. Use the sharpness-aware minimization algorithm to optimize the training process of the adaptive convolutional network model, and improve its generalization performance and robustness;

[0017] S8. During vehicle operation, continuously update the thermal field map structure using real-time sensor data and input it into the trained adaptive convolutional network model to obtain the predicted results of the thermal flow field for multiple future time steps;

[0018] S9. Generate a thermal management instruction set containing control parameters such as cooling intensity, fan speed regulation, and liquid cooling flow rate based on the thermal flow prediction results and the battery thermal safety boundary;

[0019] S10. Send the thermal management instruction set to the execution module of the battery thermal management system to drive the cooling or heating hardware for real-time regulation, so as to achieve the adaptive balance of the battery thermal flow field and the intervention of abnormal thermal states.

[0020] Optionally, the specific steps of S2 are as follows:

[0021] S21. Represent the original temperature data collected by each thermal sensor in the power battery system in the form of a triple:

[0022] {T i ,x i ,t i};

[0023] Among them, T i represents the temperature value collected by the i-th sensor at time t i , x i represents the spatial coordinate of the sensor in the battery pack, and t i represents the sampling timestamp;

[0024] S22. Set a time window Δt and evenly divide the total observation time length T total into N time segments, such that:

[0025] T total = N×Δt;

[0026] S23. The temperature measurement values of the sensor numbered i in the j-th time segment form a local time series which is expressed as:

[0027]

[0028] where M represents the total number of thermal sensors;

[0029] S24. For each time segment j, construct a thermal data matrix X j , where the i-th row is:

[0030]

[0031] Among them, Feature(·) represents a statistical feature extraction function for time series, extracting F eigenvalue features including the mean, variance, maximum value, minimum value, and first-order difference, etc.;

[0032] S25. Combine the statistical features of each sensor in this time segment with the position x i to form the thermal data tensor of time segment j in the Euclidean space:

[0033]

[0034] Optionally, the specific steps of S3 are as follows:

[0035] S31. For each item in the thermal data tensor at each time segment j ∈ [1, N] , construct a feature vector through concatenation operation to obtain a feature set;

[0036] S32. Construct a diffusion kernel function between any two feature vectors in the feature set to generate a symmetric kernel matrix K j , where the (i, k) element is defined as:

[0037]

[0038] where, ‖·‖ represents the Euclidean norm, ∈ is the diffusion kernel bandwidth parameter, and exp is the exponential function;

[0039] S33. Perform row normalization on the kernel matrix K j to generate a transition matrix P j , which is defined as:

[0040] P j = D -1 K j ;

[0041] where, D is a diagonal matrix, satisfying:

[0042]

[0043] S34. Perform eigen decomposition on the transition matrix P j to obtain the first d largest modulus eigenvalues λ1,…,λ d and their corresponding eigenvectors ψ1,…,ψ d ;

[0044] S35. According to the eigenvalues and eigenvectors, define the embedding expression of each sample in the non-Euclidean space as:

[0045]

[0046] Among them, represents the diffusion mapping embedding feature of the i-th thermal sensor in the j-th time segment;

[0047] S36. Take the embedding set under all time segments j ∈ [1, N] as the original Euclidean space data H j which is the conversion result in the diffusion manifold space for subsequent non-Euclidean modeling processing.

[0048] Optionally, the specific steps of S4 are as follows:

[0049] S41. For the embedding feature set obtained by diffusion mapping under each time segment j ∈ [1, N] construct an undirected graph where the set of graph nodes is:

[0050]

[0051] Each node corresponds to the embedding vector

[0052] S42. Define the edge weight between each pair of nodes and The edge weight value is jointly calculated based on the diffusion feature distance and the thermal feature similarity, and is expressed as follows:

[0053]

[0054] where σ y and σ T respectively represent the diffusion space scale parameter and the temperature feature space scale parameter, and exp represents the exponential function;

[0055] S43. Construct the edge weight matrix and accordingly define the adjacency matrix A of the graph j , satisfying:

[0056]

[0057] where τ is the edge weight screening threshold, used to limit the graph edge connection density;

[0058] S44. Take the graph structure constructed under each time segment j as the graph expression model of the heat flow field in the non-Euclidean space for subsequent input to the graph neural network model.

[0059] Optionally, the specific steps of S5 are as follows:

[0060] S51. Pre-deploy the graph adaptive convolutional network model in the vehicle side edge computing node. The adaptive convolutional network model consists of L graph convolutional modules, and each layer contains an adaptive adjacency adjustment mechanism and an activation function module;

[0061] S52. Take the heat map structure constructed under each time segment j ∈ [1, N] as the input graph structure. The graph node feature matrix is:

[0062]

[0063] where, is the embedded feature of the i-th node at time segment j obtained by diffusion mapping, and d is the diffusion embedding dimension;

[0064] S53. Input the adjacency matrix A of the graph j and the node feature matrix into the adaptive convolutional network model together. The first-layer graph convolution is calculated as follows:

[0065]

[0066] where, W (0) represents the first-layer weight matrix, h1 is the output dimension of this layer, σ(·) represents the activation function, represents the normalized adjacency matrix;

[0067] S54. In the l-th layer (l ∈ [2, L]), update the adjacency matrix through the adaptive adjacency mechanism as:

[0068]

[0069] where, φ(·, ·) is the similarity function between nodes, used to dynamically adjust the graph structure;

[0070] S55. Update the output features of each layer in the following way:

[0071]

[0072] where, W (l-1) is the (l - 1)-th layer weight matrix;

[0073] S56. Finally, output the heat flux prediction result:

[0074] O j = H (L) ;

[0075] where, q represents the prediction dimension of the output time step, and O j is the set of heat flux prediction values of all sensors in time segment j.

[0076] Optionally, S6 specifically includes:

[0077] S61. Collect the heat flux evolution data sequence in the vehicle operation history and construct a training dataset Among them, represents the heat map structure at time segment j, represents the input node feature matrix, represents the prediction target, that is, the heat flux characteristics of each sensor within the next q time steps;

[0078] S62. In each round of training, input the sample into the adaptive convolutional network model to generate the heat flux prediction output O j ;

[0079] S63. Calculate the loss function between the model output O j and the true label The loss function is defined in the form of mean square error as:

[0080]

[0081] Among them, and are respectively the heat flux values of the i-th node predicted by the model and the actual observation at the t-th time step;

[0082] S64. Based on the loss function perform gradient backpropagation on the model parameters Θ and use an optimizer (such as Adam) to iteratively update until the training stop condition is reached;

[0083] S65. Save the trained model parameters Θ * in the edge node device for subsequent real-time prediction task calls.

[0084] Optionally, S7 specifically includes:

[0085] S71. During the model training process, introduce a local spatial attention enhanced sharpness-aware minimization optimization mechanism. Based on the heat map structure at the current time segment j, first calculate the node temperature change rate:

[0086]

[0087] Among them, represents the set of adjacent nodes of node , is the temperature value of the i-th node;

[0088] S72. According to​ Constructive Space Attention Weight Coefficient Indicates the perturbation sensitivity of the node, defined as:

[0089]

[0090] Where, M represents the total number of thermal sensors;

[0091] S73. On the basis of the perturbation construction of the standard SAM, introduce attention weight adjustment, and weight the perturbation of each node channel in the parameter gradient of each layer to form: For each node channel in,

[0092]

[0093] Where, ∈ i Is the perturbation vector corresponding to the i-th node's corresponding channel, ρ is the perturbation intensity hyperparameter, g i Is the gradient component of the i-th node;

[0094] S74. Construct the perturbed model parameters:

[0095] Θ adv = Θ * + ∈ = Θ * + [∈1, ∈2, …, ∈ M ;

[0096] And calculate the loss of the output of the perturbed model:

[0097]

[0098] S75. Minimize Perform backpropagation and update on the model parameter Θ to obtain a thermal flow prediction model with stronger local spatial robustness.

[0099] Optionally, the specific content of S8 includes: during vehicle operation, the system continuously collects the temperature data of each measuring point of the power battery, and updates the thermal data input in real time. The latest temperature information is embedded through diffusion mapping, and combined with the current constructed thermal field map structure, the real-time input data is sent into the trained graph adaptive convolutional network model for online inference to obtain the thermal flow field prediction results of each measuring point at multiple future time steps, so as to realize continuous dynamic thermal state perception and provide input basis for subsequent thermal management strategies.

[0100] Optionally, the S9 specifically includes: The generation process of the thermal management instruction set includes: comparing the predicted values of the heat flow at each node in the future multiple time steps output by the GACNet model with a preset battery thermal safety threshold to identify areas with potential overheating risks; based on the predicted local temperature rise rate, the heat diffusion trend between nodes, and the overall battery temperature gradient distribution, combined with the current operating conditions of the vehicle, calculating thermal management parameters such as the adaptive cooling fan speed, liquid cooling circulation flow rate, and battery cooling plate temperature control value, and forming a thermal management strategy that can be directly recognized by the execution module in the form of structured instructions for output.

[0101] Optionally, the S10 specifically includes: After the thermal management instruction set is generated, it is sent to the execution module in the battery thermal management system in real time. The execution module includes a cooling fan control unit, a liquid cooling pump control unit, and a heating device control unit. The module automatically adjusts the operating states of each control component according to the received regulation instructions, and continuously collects the thermal field change information after the regulation response through a feedback loop to achieve dynamic balance adjustment of the thermal states of each part of the battery module, and triggers emergency cooling measures when detecting abnormal local temperature trends to ensure the thermal safety and stable operation of the battery system under complex operating conditions.

[0102] The beneficial effects of the present invention are: (Expand the above)

[0103] Certainly. The following are the beneficial effects achieved by the present invention summarized and refined in combination with the background technology you provided and the complete claims 1-10, expressed in a point-by-point and paragraph-by-paragraph form, which meets the requirements of patent specification writing:

[0104] Compared with the prior art, the present invention has the following beneficial effects:

[0105] (1) By mapping the temperature data of the power battery sensor from the traditional Euclidean space to the non-Euclidean manifold space, the present invention effectively captures the implicit heat diffusion structure information in the temperature distribution; then, based on the node embedding generated by diffusion mapping, a graph structure model is constructed, and the graph adaptive convolutional network is input for multi-time-step heat flow prediction, realizing accurate modeling and forward-looking perception of the internal dynamic thermal behavior of the battery. Compared with the traditional fixed threshold or static thermal model method, this scheme significantly improves both the prediction accuracy and the local thermal anomaly response speed.

[0106] (2) The present invention introduces a sharpness-aware minimization optimization algorithm enhanced by local spatial attention. During the model training phase, by guiding the perturbation direction to focus on high-heat gradient regions, the robustness and generalization ability of the model in scenarios of abnormal heat flux changes are improved. This optimization strategy can effectively suppress the prediction deviation problem caused by small-sample perturbations, enabling the model to still have stable prediction performance when facing changes in different battery module structures and operating conditions, meeting the high requirements for the reliability and safety of the model in the actual industrial environment.

[0107] (3) The overall architecture of the present invention is designed based on edge computing. Data acquisition, feature extraction, graph construction, model inference, and control instruction generation are all completed at the vehicle-side edge node, greatly reducing the dependence on communication bandwidth and the central server, and effectively solving the computational bottleneck problem of traditional centralized processing solutions in scenarios of high-frequency temperature data. The thermal management instructions generated through edge local inference can directly drive actuators such as cooling fans and liquid-cooling pumps to achieve dynamic thermal equilibrium regulation and emergency thermal intervention response of the battery system, improving the safety and intelligent level of system operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0108] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, and do not constitute a limitation to the present invention. In the drawings:

[0109] Figure 1 is a flowchart of a method for real-time thermal flow field regulation of a power battery based on edge computing proposed by the present invention;

[0110] Figure 2 is a schematic diagram of the feature conversion process of mapping Euclidean data of the thermal flow field to a non-Euclidean manifold proposed by the present invention;

[0111] Figure 3 is a schematic diagram of the perturbation weighting mechanism of the sharpness-aware minimization optimization algorithm during the training process of the graph adaptive convolutional neural network model proposed by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0112] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0113] Refer to Figures 1-3 , a method for real-time thermal flow field regulation of a power battery based on edge computing, includes the following steps:

[0114] S1. Arrange temperature sensors in the power battery system, collect temperature data, perform preprocessing, and construct a temperature time series dataset;

[0115] S2. Organize the temperature time series dataset into a high-dimensional thermal data structure in Euclidean space according to the spatial position encoding and time window partitioning method;

[0116] S3. Use the diffusion mapping algorithm to perform feature transformation on the Euclidean data, construct a non-Euclidean manifold space representation to extract the implicit thermal diffusion structure information in the data;

[0117] S4. Establish a thermal field map structure model based on the non-Euclidean representation, where the graph nodes represent the measurement points of the battery module, and the graph edges are calculated and generated according to the diffusion distance and thermal correlation degree;

[0118] S5. Load the graph adaptive convolutional network model in the edge computing nodes deployed on the vehicle side, and use the thermal field map structure as the model input;

[0119] S6. Use the historical thermal flow evolution data to train the adaptive convolutional network model so that it can learn the dynamic change characteristics of the thermal flow field on the graph structure;

[0120] S7. Use the sharpness-aware minimization algorithm to optimize the training process of the adaptive convolutional network model to improve its generalization performance and robustness;

[0121] S8. During vehicle operation, continuously update the thermal field map structure using real-time sensor data and input it into the trained adaptive convolutional network model to obtain the thermal flow field prediction results for multiple future time steps;

[0122] S9. Generate a thermal management instruction set including control parameters such as cooling intensity, fan speed regulation, and liquid cooling flow rate according to the thermal flow prediction results and the battery thermal safety boundary;

[0123] S10. Send the thermal management instruction set to the execution module of the battery thermal management system to drive the cooling or heating hardware for real-time regulation, and achieve adaptive equilibrium and abnormal thermal state intervention of the battery thermal flow field.

[0124] The present invention realizes real-time modeling, prediction, and active control of the thermal flow field of the power battery by constructing a complete edge-side thermal flow regulation process, improves the response efficiency, intelligent level, and thermal safety guarantee ability of the thermal management system, and is applicable to new energy vehicles and energy storage systems under complex operating conditions.

[0125] In this embodiment, the specific content of S2 includes:

[0126] S21. Represent the original temperature data collected by each thermal sensor in the power battery system in the form of a triple:

[0127] {T i , x i , t i};

[0128] Among them, T i represents the temperature value collected by the i-th sensor at time t i moment, x i represents the spatial coordinates of the sensor in the battery pack, and t i represents the sampling timestamp;

[0129] S22. Set the time window Δt, and divide the total observation time length T total evenly into N time segments, such that:

[0130] T total = N × Δt;

[0131] S23. The temperature measurement values of the sensor numbered i in the j-th time segment form a local time series which is expressed as:

[0132]

[0133] Among them, M represents the total number of thermal sensors;

[0134] S24. For each time segment j, construct a thermal data matrix X j , where the i-th row is:

[0135]

[0136] Among them, Feature(·) represents a statistical feature extraction function for the time series, extracting F eigenvalue features including the mean value, variance, maximum value, minimum value, and first-order difference, etc.;

[0137] S25. Combine the statistical features of each sensor in this time segment with the position x i to form a thermal data tensor for the time segment j in the Euclidean space:

[0138]

[0139] The present invention effectively enhances the structured expression ability of thermal data by introducing a spatial coding and time window partitioning mechanism to convert the original temperature time series data into an Euclidean structure that can be used for high-dimensional modeling, providing a stable and efficient input basis for subsequent non-linear dimensionality reduction and mapping.

[0140] In this embodiment, the specific steps of S3 are as follows:

[0141] S31. For each thermal data tensor in each time segment j ∈ [1, N], each item is constructed into a feature vector Obtain the feature set;

[0142] S32. Construct a diffusion kernel function between any two feature vectors in the feature set to generate a symmetric kernel matrix K j , where the i,k element is defined as:

[0143]

[0144] where, ‖·‖ represents the Euclidean norm, ∈ is the diffusion kernel bandwidth parameter, and exp is the exponential function;

[0145] S33. Row-normalize the kernel matrix K j to generate a transition matrix P j , which is defined as:

[0146] P j = D -1 K j ;

[0147] where, D is a diagonal matrix that satisfies:

[0148]

[0149] S34. Perform eigen-decomposition on the transition matrix P j to obtain the first d largest-magnitude eigenvalues λ1,…,λ d and their corresponding eigenvectors ψ1,…,ψ d ;

[0150] S35. According to the eigenvalues and eigenvectors, define the embedding expression of each sample in the non-Euclidean space as:

[0151]

[0152] where, represents the diffusion mapping embedding feature of the i-th thermal sensor in the j-th time segment;

[0153] S36. Take the embedding set at all time segments j∈[1,N] as the conversion result of the original Euclidean space data H j in the diffusion manifold space for subsequent non-Euclidean modeling processing.

[0154] The present invention maps Euclidean thermal data into non-Euclidean manifold embeddings through the diffusion mapping algorithm, effectively capturing the internal physical structure and local non-linear characteristics of heat diffusion, improving the representation accuracy of heat flow modeling, and providing a high-quality feature basis for graph structure construction and deep learning modeling.

[0155] In this embodiment, S4 specifically includes:

[0156] S41. For each time segment j ∈ [1, N], construct an undirected graph for the set of embedded features obtained by diffusion mapping where the set of graph nodes is: where each node

[0157]

[0158] corresponds to an embedding vector

[0159] S42. Define the edge weight between each pair of nodes The edge weight value is jointly calculated based on the diffusion feature distance and the thermal feature similarity, and is expressed as follows:

[0160]

[0161] where σ y and σ T respectively represent the diffusion space scale parameter and the temperature feature space scale parameter, and exp represents the exponential function;

[0162] S43. Construct an edge weight matrix and define the adjacency matrix A of the graph accordingly j , satisfying:

[0163]

[0164]

[0165] where τ is the edge weight screening threshold, used to limit the graph edge connection density; S44. Take the graph structure constructed in each time segment j

[0166] as the graph expression model of the heat flow field in the non-Euclidean space for the input of the subsequent graph neural network model.

[0167] The present invention constructs a heat field graph structure model based on the embedded features in the diffusion space, dynamically captures the heat diffusion relationship between sensors, and combines graph topology for modeling, making the heat flow structure expression more flexible and controllable, and enhancing the response ability of the model to local heat anomalies and dynamic propagation paths.

[0168] In this embodiment, S5 specifically includes:

[0169] S51. Pre-deploy a graph adaptive convolutional network model in the vehicle side edge calculation nodes. The adaptive convolutional network model consists of L graph convolutional modules, and each layer contains an adaptive adjacency adjustment mechanism and an activation function module;

[0169] S52. Take the heatmap structure constructed at each time segment \(j\in[1, N]\) as the input graph structure. The graph node feature matrix is:

[0170]

[0171] where is the embedding feature of the \(i\)-th node at time segment \(j\) obtained by diffusion mapping, and \(d\) is the diffusion embedding dimension;

[0172] S53. Input the adjacency matrix \(A\) of the graph j and the node feature matrix into the adaptive convolutional network model together. The first-layer graph convolution calculation is as follows:

[0173]

[0174] where \(W\) (0) represents the first-layer weight matrix, \(h1\) is the output dimension of this layer, \(\sigma(\cdot)\) represents the activation function, represents the normalized adjacency matrix;

[0175] S54. In the \(l\)-th layer (\(l\in[2, L]\)), update the adjacency matrix through the adaptive adjacency mechanism as:

[0176]

[0177] where \(\varphi(\cdot,\cdot)\) is the similarity function between nodes, which is used to dynamically adjust the graph structure;

[0178] S55. Update the output features of each layer in the following way:

[0179]

[0180] where \(W\) (l-1) is the \((l - 1)\)-th layer weight matrix;

[0181] S56. Finally, output the heat flow prediction result:

[0182] O j =H (L) ;

[0183] where \(q\) represents the prediction dimension of the output time step, and \(O\) j is the set of heat flow prediction values of all sensors in time segment \(j\).

[0184] After the graph adaptive convolutional neural network model is deployed on the edge node, it can adaptively adjust the graph convolution strategy based on the current heatmap structure, achieve efficient spatio-temporal joint inference, provide model support with high accuracy and low resource overhead for multi-step heat flow prediction under edge computing conditions, and be adapted to in-vehicle real-time applications.

[0185] In this embodiment, the specific steps of S6 are as follows:

[0186] S61. Collect the heat flow evolution data sequence in the vehicle operation history and construct a training data set Among them, represents the heatmap structure at time segment j, represents the input node feature matrix, represents the prediction target, that is, the heat flow characteristics of each sensor within the next q time steps;

[0187] S62. In each round of training, input the sample into the adaptive convolutional network model to generate the heat flow prediction output O j ;

[0188] S63. Calculate the loss function j between the model output O and the true label which is defined in the form of mean squared error as:

[0189]

[0190] Among them, and are respectively the heat flow values of the i-th node predicted by the model and the actual observation at the t-th time step;

[0191] S64. Based on the loss function perform gradient backpropagation on the model parameters Θ, and use an optimizer (such as Adam) to iteratively update until the training stop condition is reached;

[0192] S65. Save the trained model parameters Θ * in the edge node device for subsequent real-time prediction task calls.

[0193] By constructing a historical heat flow evolution data set and training the graph adaptive convolutional neural network model, it enables the model to have the learning ability for complex heat flow change trends, helps to discover potential thermal runaway patterns, and enhances the system's forward prediction ability and strategy planning accuracy for future thermal field states.

[0194] In this embodiment, the specific steps of S7 are as follows:

[0195] S71. During the model training process, introduce a local spatial attention enhanced sharpness-aware minimization optimization mechanism, based on the heatmap structure of the current time segment j First, calculate the node temperature change rate:

[0196]

[0197] where represents the set of adjacent nodes of node , and is the temperature value of the i-th node;

[0198] S72. According to , construct the spatial attention weight coefficient , which represents the perturbation sensitivity of this node and is defined as:

[0199]

[0200] where M represents the total number of thermal sensors;

[0201] S73. On the basis of the perturbation construction of the standard SAM, introduce attention weight adjustment, and weight the perturbation of each node channel in the parameter gradient of each layer to form:

[0202]

[0203] where ∈ i is the perturbation vector corresponding to the i-th node channel, ρ is the perturbation intensity hyperparameter, and g i is the gradient component of the i-th node;

[0204] S74. Construct the perturbed model parameters:

[0205] Θ adv = Θ * + ∈ = Θ * + [∈1, ∈2, …, ∈ M ;

[0206] And calculate the loss of the output of the perturbed model:

[0207]

[0208] S75. Minimize Perform backpropagation and update on the model parameter Θ to obtain a heat flow prediction model with stronger local spatial robustness.

[0209] The present invention adopts an optimization strategy of sharpness-aware minimization enhanced by local spatial attention, effectively improving the sensitivity and generalization ability of the model to high thermal gradient regions, enhancing the adaptability of the training process to the real perturbation distribution, and improving the stability of the model in a dynamic thermal environment.

[0210] In this embodiment, S8 specifically includes: during vehicle operation, the system continuously collects temperature data of each measurement point of the power battery and updates the thermal data input in real time. The latest temperature information is embedded through diffusion mapping, and combined with the currently constructed thermal field map structure, the real-time input data is sent into the trained graph adaptive convolutional network model for online inference to obtain the predicted results of the heat flux field at each measurement point in multiple future time steps, thereby realizing continuous dynamic thermal state perception and providing an input basis for subsequent thermal management strategies.

[0211] The present invention realizes the continuity and dynamics of the heat flux prediction process by updating the thermal map structure in real time and calling the optimized GACNet model for online inference, significantly improving the real-time performance of the battery thermal state response and the closed-loop ability of predictive control.

[0212] In this embodiment, S9 specifically includes: the generation process of the thermal management instruction set includes: comparing the predicted heat flux values at each node in multiple future time steps output by the graph adaptive convolutional neural network model with the preset battery thermal safety threshold to identify regions with potential overheating risks; based on the predicted local temperature rise rate, heat diffusion trend between nodes, and the overall battery pack temperature gradient distribution, combined with the current vehicle operating conditions, calculating thermal management parameters such as the adaptive cooling fan speed, liquid cooling circulation flow rate, and battery cooling plate temperature control value, and forming a thermal management strategy output that can be directly recognized by the execution module in the form of structured instructions.

[0213] The thermal management strategy generation process of the present invention combines the prediction results and thermal safety boundaries, has the ability to regulate multiple parameters, and can dynamically adjust the operating states of actuators such as fans and liquid cooling according to different heat diffusion situations, improving the system energy efficiency and the pertinence of thermal management strategies.

[0214] In this embodiment, S10 specifically includes: after the thermal management instruction set is generated, it is sent to the execution module in the battery thermal management system in real time. The execution module includes a cooling fan control unit, a liquid cooling pump control unit, and a heating device control unit. The module automatically adjusts the operating states of each control component according to the received regulation instructions, and continuously collects information on the thermal field changes after the regulation response through a feedback loop to realize dynamic balance adjustment of the thermal states of various parts of the battery module, and trigger emergency cooling measures when detecting local temperature abnormal trends to ensure the thermal safety and stable operation of the battery system under complex operating conditions.

[0215] By sending the thermal management instruction set to the execution module in real time, the present invention realizes a fast closed-loop control from the regulation instruction to the hardware response, enhances the system's intervention ability for sudden temperature rise events, and effectively improves the thermal stability and safety redundancy of the whole vehicle or energy storage device.

[0216] Example 1:

[0217] In order to verify the application effect of the present invention in the actual operation management of power batteries, in this example, a pure electric city bus equipped with a ternary lithium battery pack is taken as the test object, and a deployment and comparative experiment on its thermal management system is carried out under high load and high temperature environments. The electric bus is equipped with 96 battery modules, with a total assembly energy of 84.5 kWh. The driving environment of the whole vehicle is the urban main line during the high temperature period in July in a certain city, with a daily operation duration of more than 12 hours, a peak ambient temperature of 38.4 °C, an average one-way operation distance of 19 kilometers, and obvious high heat load operation conditions such as urban congestion and frequent acceleration and deceleration.

[0218] In this scenario, the bus originally used a conventional thermal management system. This system periodically reads the data of temperature sensors through a centralized controller and turns on the cooling strategy based on a fixed temperature threshold. For example, when the temperature of any module in the battery pack exceeds 45 °C, the liquid cooling cycle is started. However, it is found in actual operation that this strategy has two serious problems: First, the cooling response is delayed, and the cooling starts when the temperature approaches 50 °C, and the heat diffusion has caused local overheating to spread; Second, the cooling control is not precise enough, and the fan runs at full speed frequently, resulting in high energy consumption and uneconomical control.

[0219] To solve the above problems, the method of the present invention is embedded in the edge control unit of the bus and integrated with the original temperature control system. During the deployment process, a total of 64 thermal sensors are arranged on each battery module to collect real-time temperature data. These data are mapped to a non-Euclidean space through the diffusion mapping algorithm, so as to more truly restore the heat diffusion trend and the heat conduction relationship between modules. The constructed dynamic graph structure is sent to the graph adaptive convolutional network running on the edge node for prediction and inference. GACNet consists of 3 layers of graph convolution, including an adaptive edge weight mechanism and a position embedding module, and each sensor point can predict the heat flow trend within the next 15 minutes.

[0220] To improve the robustness of the model under complex operating conditions, the sharpness-aware minimization algorithm with enhanced local spatial attention proposed by the present invention is used during system training. This algorithm automatically adjusts the perturbation amplitude according to the temperature difference between nodes, making the model more sensitive to high heat gradient regions. During the real-time prediction stage, the system automatically updates the heat map structure every minute and generates a dynamic thermal management strategy for local hot spots, including adjusting the liquid cooling pump rate, fan speed, etc.

[0221] After 60 hours of on-site operation testing over 5 consecutive days, the system successfully identified early signals of local temperature rise in multiple locations and executed local cooling regulation in advance. In the traditional solution, the vehicle triggered the full-speed operation of the fan once per hour due to overheating on average, while the system of the present invention can predict the trend of local temperature rise 8 - 12 minutes in advance, and control the temperature below 42°C only through 20% segmented adjustment of the wind speed, significantly reducing energy consumption.

[0222] The test data shows that after the deployment of the present invention, the average cooling response time is advanced by 9.3 minutes, the energy consumption of the fan is reduced by 34.6%, the energy consumption of the liquid cooling pump is reduced by 21.2%, the total daily energy consumption of the cooling system is reduced from the original 5.3 kWh to 3.4 kWh, the maximum temperature difference of the battery module drops from 11.7°C to 6.1°C, and the thermal distribution balance is improved by nearly 48%. At the same time, the prediction accuracy rate reaches 93.7%, and the success rate of the system's abnormal response is increased to 96.5%, demonstrating powerful prediction ability and control accuracy.

[0223] Table 1: Comparative test data table of the present invention and the traditional thermal management system

[0224]

[0225]

[0226] As can be seen from Table 1, the method of the present invention has achieved significant improvements compared with the traditional thermal management strategy in multiple key indicators. In terms of energy consumption control, the system of the present invention effectively reduces the ineffective operation of the cooling system by introducing a predictive regulation mechanism based on graph neural networks. The average daily energy consumption of the fan is reduced from 3.8 kWh by the traditional method to 2.5 kWh, a decrease of 34.6%; the energy consumption of the liquid cooling pump is also reduced from 1.5 kWh to 1.2 kWh, a reduction of 21.2%. The overall energy consumption of the cooling system has decreased by nearly 36%, significantly improving the energy efficiency performance without sacrificing thermal safety.

[0227] In terms of cooling response ability, the graph adaptive heat flow prediction model adopted by the present invention can identify the temperature rise trend in advance, with an average cooling response time advanced by 9.3 minutes, breaking the limitation of the passive triggering mechanism of the traditional system and significantly reducing the risk of temperature control hysteresis of the system. The introduction of the prediction mechanism also brings more accurate regulation results. The maximum temperature difference of the battery module drops from 11.7°C to 6.1°C, and the thermal distribution balance is improved by nearly 48%, effectively alleviating the risk of thermal unevenness caused by local hot spots.

[0228] In addition, in terms of thermal anomaly recognition and trend prediction, the local spatial attention enhanced training mechanism of the present invention significantly improves the adaptability of the model to complex heat diffusion paths. Test data shows that the success rate of local overheat recognition has increased from 72.4% to 96.5%, and the accuracy of heat flow trend prediction has risen from 68.9% to 93.7%. The model maintains high robustness and prediction stability in a dynamic operating environment. At the same time, the number of times the fan runs at full speed has been reduced from 11 times per day to 3 times, greatly reducing noise and mechanical losses, and improving the comfort and equipment life of vehicle operation.

[0229] Based on the above analysis, it can be seen that the present invention not only improves the energy efficiency and safety of the power battery thermal management system, but also enhances its intelligence and autonomy levels. It is applicable to new energy vehicle-mounted scenarios and distributed energy storage devices with high requirements for real-time thermal response and operation stability, and has significant engineering promotion value.

[0230] This embodiment effectively proves that the present invention has significant practical value and engineering feasibility in an actual complex thermal environment, especially having prominent technical advantages in aspects such as improving real-time thermal response, energy-saving control accuracy, and ensuring thermal safety and stability.

[0231] The above is only a preferred specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, making equivalent replacements or changes, should be covered within the protection scope of the present invention.

Claims

1. A real-time heat flow field regulation method for power batteries based on edge computing, characterized in that It includes the following steps: S1. Deploy temperature sensors, collect temperature data, perform preprocessing, and construct a temperature time series dataset; S2. Organize the temperature time series dataset into a high-dimensional thermal data structure in Euclidean space; S3. Use the diffusion mapping algorithm to perform feature transformation on the Euclidean data and construct a non-Euclidean manifold space expression; S4. Establish a thermal field map structure model; S5. Load the graph adaptive convolutional network model in the edge computing node and use the thermal field map structure as the model input; S6. Use historical heat flow evolution data to train the graph adaptive convolutional network model; S7. Use the sharpness-aware minimization algorithm to optimize the training process of the graph adaptive convolutional network model; S8. Continuously update the thermal field map structure during vehicle operation, input it into the trained graph adaptive convolutional network model, and obtain the heat flow field prediction results for multiple future time steps; S9. Generate a thermal management instruction set according to the heat flow prediction results and the battery thermal safety boundary; S10. Send the thermal management instruction set to the execution module of the battery thermal management system to drive the hardware for real-time regulation.

2. The real-time heat flux field regulation method of a power battery based on edge computing according to claim 1, characterized in that The specific content of S2 includes: S21. Represent the original temperature data collected by each thermal sensor in the power battery system in the form of a triple {T i , x i , t i}, where T i represents the temperature value collected by the i-th sensor at time t i , x i represents the spatial coordinate of the sensor in the battery pack, and t i represents the sampling timestamp; S22. Set a time window Δt and divide the total observation time length T total evenly into N time segments such that T total = N × Δt; S23. The temperature measurement values of the sensor numbered i within the j-th time segment form a local time series which is expressed as: where M represents the total number of thermal sensors; S24. For each time segment j, construct the hot data matrix X j , where the i-th row is where Feature(·) represents the statistical feature extraction function performed on the time series; S25. Combine the statistical features of each sensor under this time segment with the position x i to form the thermal data tensor of time segment j in Euclidean space 3. The real-time heat flow field regulation method of a power battery based on edge computing according to claim 1, characterized in that The specific content of S3 includes: S31. For each time segment \(j\in[1, N]\), each item in the hot data tensor \(H\) j is constructed into a feature vector through a splicing operation to obtain a feature set;​​ S32. Construct a diffusion kernel function between any two feature vectors in the feature set to generate a symmetric kernel matrix K j ; S33. Normalize the rows of the kernel matrix K j to generate the transition matrix P j ; S34. Perform eigen - decomposition on the transition matrix P j to obtain the first d largest - modulus eigenvalues λ1, …, λ d and their corresponding eigen - vectors ψ1, …, ψ d ; S35. Define each sample eigenvector according to the eigenvalue and eigenvector The embedding expression in the non-Euclidean space is where represents the diffusion mapping embedding feature of the i-th thermal sensor in the j-th time segment; S36. Take the set of embeddings at all time segments \(j\in[1,N]\) as the original Euclidean space data \(H\) j which is the conversion result in the diffusion manifold space.

4. A method for real-time heat flow field regulation of a power battery based on edge computing according to claim 1, characterized in that, The specific content of S4 includes: S41. Construct an undirected graph where the graph node set is each node corresponds to an embedding vector S42. Define each pair of nodes and the edge weight between them The edge weight value is jointly calculated based on the diffusion feature distance and the thermal feature similarity; S43. Construct an edge weight matrix Define the adjacency matrix A of the graph accordingly j ; S44. Take the graph structure constructed under each time segment j as the graph expression model of the heat flow field in the non-Euclidean space.

5. A real-time heat flow field regulation method for power batteries based on edge computing according to claim 1, characterized in that The specific content of S5 includes: S51. Pre-deploy the graph adaptive convolutional network model in the vehicle-side edge computing node. The adaptive convolutional network model consists of L graph convolutional modules, and each layer contains an adaptive adjacency adjustment mechanism and an activation function module; S52. Take the heat map structure as the input graph structure, and the graph node feature matrix is S53. Input the adjacency matrix A of the graph j and the node feature matrix into the adaptive convolutional network model together to perform the first-layer graph convolution calculation H (1) ; S54. In the layer, update the adjacency matrix through the adaptive adjacency mechanism as follows: where φ(·,·) is the similarity function between nodes; S55. The output features of each layer are updated in the following manner: Among them, is the layer weight matrix, and σ is the activation function. S56. Finally, output the heat flux prediction result O according to the output features j .

6. The method 1 of real-time heat flow field regulation for power batteries based on edge computing according to claim 1, characterized in that, The specific content of S6 includes: S61. Collect the sequence of heat flux evolution data in the vehicle operation history and construct a training dataset Among them, represents the heat map structure at time segment j, represents the input node feature matrix, represents the prediction target; S62. In each round of training, input the heat map structure sample into the adaptive convolutional network model to generate a heat flow prediction output; S63. Calculate the loss function between the output of the calculation model and the true label between It is defined in the form of mean squared error as follows: Among them, and are the heat flux values of the i-th node of the model prediction and the actual observation at the t-th time step, respectively, and M represents the total number of heat sensors; S64. Based on the loss function Perform gradient backpropagation on the model parameters Θ, and use the optimizer to iteratively update until the training stop condition is reached; S65. Save the model parameters Θ * after training in the edge node device.

7. A real-time heat flow field regulation method for power batteries based on edge computing according to claim 1, characterized in that The specific content of S7 includes: S71. During the model training process, introduce a local spatial attention enhanced sharpness-aware minimization optimization mechanism, and based on the heatmap structure of the current time segment j First, calculate the node temperature change rate S72. According to Constructed space attention weight coefficient Indicates the perturbation sensitivity of the node, defined as: S73. On the basis of the perturbation construction of the standard sharpness-aware minimization algorithm, introduce attention weight adjustment, and weight and perturb each node channel in the parameter gradient of each layer respectively to form: where, ∈ i is the perturbation vector of the channel corresponding to the i-th node, ρ is the perturbation intensity hyperparameter, and g i is the gradient component of the i-th node; S74. Construct the perturbed model parameters: Θ adv = Θ * + ∈ = Θ * + [∈1, ∈2, …, ∈ M ; Calculate the loss of the output of the perturbed model: S75, Minimization Backpropagate and update the model parameters Θ to obtain a heat flow prediction model with stronger local spatial robustness.

8. A real-time heat flux field regulation method for power batteries based on edge computing according to claim 1, characterized in that The specific content of S8 includes: During vehicle operation, the system continuously collects the temperature data of each measuring point of the power battery, updates the thermal data input in real time, embeds the latest temperature information through the diffusion mapping method, combines the currently constructed thermal field map structure, and sends the real-time input data into the trained graph adaptive convolutional network model for online inference to obtain the heat flow field prediction results of each measuring point for multiple future time steps.

9. A real-time heat flow field regulation method for power batteries based on edge computing according to claim 1, characterized in that The specific content of S9 includes: Generate a thermal management instruction set, compare the heat flow prediction values of each node for multiple future time steps output by the graph adaptive convolutional neural network model with the preset battery thermal safety threshold, identify the areas with potential overheating risks, and calculate the adaptive cooling fan speed, liquid cooling circulation flow rate, battery cooling plate temperature control values and other thermal management parameters based on the predicted local temperature rise rate, heat diffusion trend between nodes, and the overall battery temperature gradient distribution, and form a thermal management strategy output that can be directly recognized by the execution module in the form of structured instructions.

10. A real-time heat flux field control method for power batteries based on edge computing according to claim 1, characterized in that, The specific steps of S10 are as follows: After the thermal management instruction set is generated, it is sent to the execution module in the battery thermal management system in real time. The execution module includes a cooling fan control unit, a liquid cooling pump control unit, and a heating device control unit. The module automatically adjusts the operating states of each control component according to the received regulation instructions, continuously collects the thermal field change information after the regulation response through a feedback loop, realizes the dynamic balance adjustment of the thermal states of each part of the battery module, and triggers an emergency cooling measure when detecting a local temperature anomaly trend.

Images