Power battery three-dimensional temperature field prediction method based on boundary condition extrapolation and deep learning

Abstract

The invention relates to the technical field of power battery thermal management and intelligent prediction, and discloses a power battery three-dimensional temperature field prediction method based on boundary condition extrapolation and deep learning, and the method comprises the steps: building a three-dimensional unsteady state heat conduction simulation model on an open source finite element platform OpenFOAM based on the geometric structure and thermophysical parameters of a battery pack; generating three-dimensional temperature field data under multiple working conditions through the simulation model, and taking the three-dimensional temperature field data as a supervised learning label to construct a training data set; performing boundary condition extrapolation by adopting a Gaussian process regression method, and constructing an approximate continuous boundary field; a deep learning network based on a Transform is established; introducing physical constraint optimization model training, and verifying model performance through a test set; and the model is pruned, quantified and deployed in the BMS, the three-dimensional temperature field is output in real time, closed-loop control is formed by the three-dimensional temperature field and the BMS, high-precision real-time reconstruction of the three-dimensional temperature field of the power battery is achieved, data support is provided for battery thermal runaway early warning and active thermal management strategy making, and the operation safety and the intelligent level are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of power battery thermal management and intelligent prediction technology, specifically to a method for predicting the three-dimensional temperature field of a power battery based on boundary condition extrapolation and deep learning. Background Technology

[0002] With the rapid development of the new energy vehicle industry, the operational safety and thermal management capabilities of power battery systems, as core energy storage components, have become key factors restricting the industry's development. During charging and discharging, power batteries generate a large amount of Joule heat. If this heat cannot be dissipated in time, it will lead to uneven temperature distribution within the battery pack, causing localized overheating. In severe cases, this can trigger thermal runaway, resulting in safety accidents such as fires and explosions. Therefore, accurately obtaining complete three-dimensional temperature field information inside the battery pack is a prerequisite for achieving refined thermal management and safety early warning.

[0003] Current technologies for obtaining the temperature distribution of power batteries mainly fall into two categories: one is through actual measurement using temperature sensors, and the other is through simulation based on physical models such as finite element method or CFD. However, both of these technologies have significant drawbacks in practical applications: In terms of sensor testing, due to the compact internal space of the battery pack, the high cost of sensors, and the complexity of wiring, only a limited number of sensors can be deployed in a few key locations in practical applications. This sparse deployment method cannot cover all critical areas inside the battery, making it difficult to reflect the complete three-dimensional temperature distribution characteristics. As a result, potential safety hazards such as localized overheating cannot be detected in a timely manner, seriously affecting the safety management and control of the battery system.

[0004] In terms of physical simulation, existing thermal simulation methods heavily rely on accurate information such as the thermal conductivity and specific heat capacity of battery materials, as well as structural characteristics and environmental boundary conditions. However, in actual operating conditions, these parameters change dynamically with battery aging and environmental changes, leading to significant deviations between the simulation model's output and the actual temperature distribution. Furthermore, physical simulation is computationally complex, making it difficult to meet the real-time requirements of BMS systems and thus unsuitable for direct application in online prediction scenarios.

[0005] Furthermore, existing technologies generally lack deep integration with battery management systems (BMS), and most can only achieve unidirectional temperature field prediction, failing to support advanced functions such as fault diagnosis, thermal anomaly warning, and active thermal management adjustment. Currently, the industry lacks an integrated technical solution that can operate in real time within a BMS, reconstruct a complete three-dimensional temperature field using limited sensor data, and participate in control decisions.

[0006] In summary, current power battery temperature field prediction technology has significant shortcomings in terms of sensor utilization, simulation accuracy, real-time performance, and system integration. Summary of the Invention

[0007] This invention aims to provide a method for predicting the three-dimensional temperature field of a power battery based on boundary condition extrapolation and deep learning. This method can integrate multi-source data, balance accuracy and real-time performance, and can be seamlessly integrated with a BMS to reconstruct the three-dimensional temperature field, thereby achieving the goal of accurately predicting the three-dimensional temperature field of a power battery and improving the safety and intelligence level of the power battery system.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: Basic Solution: A method for predicting the three-dimensional temperature field of a power battery based on boundary condition extrapolation and deep learning, including the following steps: S1: Collect local temperature data under different operating conditions through N temperature sensors preset in the power battery pack. The data includes sensor coordinates (xi,yi,zi), current I, voltage V, SOC and ambient temperature T, where N is an integer from 10 to 30. S2: Perform outlier removal, missing value imputation, and normalization on the collected data. Outlier removal adopts the 3σ rule: Where T is the average temperature. The standard deviation of temperature; Missing value imputation uses Kriging interpolation: ,in For the coordinates of the missing data points, For interpolation weights; Normalization uses the Min-Max algorithm: Where X is the original data, , These are the minimum and maximum values ​​of the data, respectively. S3: A three-dimensional unsteady-state heat conduction simulation model is established on the open-source finite element platform OpenFOAM based on the battery pack geometry and thermal properties. The model satisfies the heat conduction equation: , For density, For specific heat capacity, For thermal conductivity, This is the temperature difference value; Boundary conditions include the convective heat transfer boundary at the contact surface of the cooling plate: h is the surface convective heat transfer coefficient. This refers to the temperature of the cooling plate. Radiative heat transfer boundary of the battery pack casing: h conv Indicates the convective heat transfer coefficient of the outer shell surface. Indicates the temperature of the battery pack casing; S4: Generate three-dimensional temperature field data under multiple operating conditions using the simulation model, and use this data as supervised learning labels to construct a training dataset. The operating conditions coverage must meet the following requirements: , Where 0.2C ≤ expansion rate ≤ 3C, -20℃ ≤ T ≤ 60℃, and 10% ≤ SOC ≤ 90%; S5: The boundary conditions are extrapolated using the Gaussian process regression method to construct an approximately continuous boundary field. The regression model is as follows: , Kernel function selection: , Hyperparameters Optimize using maximum likelihood estimation; S6: Construct a high-dimensional input vector, which includes sensor temperature values ​​and spatial coordinate encoding, extrapolated boundary temperature values, and operating parameters. The spatial coordinate encoding uses sinusoidal encoding. , A one-dimensional mapping representing spatial location, This represents the hidden layer dimension of the Transformer model. Represents the position encoding matrix; S7: Construct a deep learning network based on Transformer, which includes a positional encoding layer, a multi-head attention layer, and a residual module. The multi-head attention weights are calculated to satisfy the following: , This represents the feature matrix obtained by projecting the input sequence. This represents the dimension of a vector. For the number of attention heads, This represents the normalized attention weight matrix; Residual regularization satisfies: , Indicates the pyrogen source item. Represents the Laplace of the predicted temperature field; S8: Introduce physical constraints to optimize model training. The loss function includes mean squared error loss, temperature gradient penalty loss, and boundary consistency loss. The total loss function is: , Indicates the first Individual element models predict temperature. Indicates the first Finite element simulation temperature of individual elements The weighting coefficients for the physical constraint loss term; The gradient penalty loss satisfies: , This represents the predicted temperature gradient vector. Indicates the gradient magnitude threshold; S9: Verify model performance using the test set, employing the 3D spatial mean temperature prediction error (MSE). , This represents the predicted temperature field value. Represents the true value of the temperature field; and the spatial gradient error: As an evaluation indicator This represents the three-dimensional space of the entire battery pack. These represent the predicted and true temperature gradient fields, respectively. S10: Prune and quantize the model, convert it to TnsorRT or ONNX format, deploy it in BMS, output the three-dimensional temperature field in real time, and form a closed-loop control with BMS.

[0009] Beneficial effects: Through multi-step collaborative design, it achieves accurate mapping from limited sensor data to a complete three-dimensional temperature field, balancing prediction accuracy and real-time performance. It solves the problems of insufficient sensor coverage and disconnect between simulation and actual working conditions in existing technologies, and can be directly deployed in BMS systems to support thermal management needs in all scenarios.

[0010] This invention addresses the problems of insufficient sensor coverage, limited accuracy of physical simulation, and poor integration of BMS in existing temperature field prediction. By fusing limited sensor data, boundary condition extrapolation, open-source finite element simulation, and deep learning algorithms, it achieves high-precision real-time reconstruction of the three-dimensional temperature field of power batteries. This provides data support for battery thermal runaway early warning and active thermal management strategy formulation, significantly improving the operational safety and intelligence level of power battery systems.

[0011] Preferably, the temperature sensor placement locations in step S1 include the central heating area of ​​the battery pack, the module boundary, the contact surface of the cooling plate, and the cell tab location. The number of sensors N satisfies: total number of battery pack cells × 5% ≤ N ≤ total number of battery pack cells × 15%.

[0012] Beneficial effects: By limiting the location and number of sensors, the representativeness of the collected data can be ensured while controlling costs and wiring complexity. This provides high-quality basic data for subsequent boundary extrapolation and model training, and avoids the decline in prediction accuracy caused by unreasonable sensor layout.

[0013] Preferably, the thermophysical property parameters mentioned in step S3 include thermal conductivity kx = 10⁻¹² W / (m²). K), ky = 2-4W / (m K), kz = 7-9W / (m K), specific heat capacity c = 700-900 J / (kg) K), density ρ = 2500-2800 kg / m³.

[0014] Beneficial effects: By clarifying the range of values ​​for thermophysical parameters, the simulation model can better reflect the actual physical characteristics of power batteries, reduce the deviation between simulation and actual measurement, improve the reliability of the training dataset, and provide a physical basis for high-precision prediction by the model.

[0015] Preferably, the number of iterations for optimizing the hyperparameters of the kernel function in the Gaussian process regression described in step S5 is no less than 100, and the convergence condition is that the hyperparameter update amount is less than 1×e. -5 .

[0016] Beneficial effects: By limiting the number of iterations and convergence conditions for hyperparameter optimization, the stability and accuracy of boundary condition extrapolation results are ensured, and boundary field distortion caused by insufficient hyperparameter optimization is avoided, providing high-quality continuous boundary information for neural network input.

[0017] Preferably, in step S7, the Transformer network has 4-8 multi-head attention heads, 256-512 hidden layer neurons, and uses the tanh function as the activation function.

[0018] Beneficial effects: By optimizing the key structural parameters of the Transformer network, the computational complexity and feature extraction capability of the model are balanced. This not only effectively captures the long-distance spatial correlation of the temperature field, but also controls the model size, meeting the deployment requirements of the BMS embedded platform.

[0019] Preferably, the weights of each loss term in the total loss function in step S8 satisfy the following: mean square error loss weight Wmse=1.0, temperature gradient penalty loss weight Wgrad=0.3-0.7, and boundary consistency loss weight Wbound=0.5-1.0.

[0020] Beneficial effects: By reasonably allocating the weights of each loss term, the model training can simultaneously take into account both numerical accuracy and physical consistency, avoid prediction results that do not conform to thermodynamic laws, and improve the robustness of the model under complex working conditions.

[0021] Preferably, the qualification criterion for model evaluation in step S9 is: the three-dimensional spatial average temperature prediction error. ≤0.8℃², spatial gradient error ≤0.15℃ / mm.

[0022] Beneficial effects: By quantifying the model evaluation criteria, the accuracy requirements of the model are clarified, ensuring that the trained model can meet the needs of actual engineering applications and providing reliable temperature data support for battery thermal management and safety early warning.

[0023] Preferably, after model quantization in step S10, the number of parameters is compressed to 30%-50% of the original model, and the inference latency is ≤50ms.

[0024] Beneficial effects: By designing a lightweight model, the computational complexity and storage requirements of the model are significantly reduced while ensuring prediction accuracy. This ensures that the model can run in real time on the BMS embedded chip, meeting the real-time requirements of the battery system for temperature field prediction.

[0025] Preferably, the method further includes a closed-loop control step with the BMS: when the temperature T ≥ 80℃ in any region of the predicted three-dimensional temperature field, a first-level warning is triggered and maximum intensity cooling is started; when 60℃ ≤ T < 80℃, a second-level warning is triggered and cooling power is adjusted; when T < 60℃, the conventional thermal management strategy is maintained.

[0026] Beneficial effects: By clarifying the correspondence between temperature thresholds and control strategies, the deep integration of temperature field prediction and BMS thermal management is achieved, forming a closed-loop control, responding promptly to abnormal temperature conditions, and effectively preventing the risk of thermal runaway.

[0027] Preferably, the method further includes a multimodal fusion extension step: integrating the SOC, SOH, internal resistance and thermal stress parameters of the battery pack into a high-dimensional input vector, and achieving collaborative prediction of the three-dimensional temperature field and battery health status by fine-tuning the multi-head attention weights of the Transformer network, with the mean square error of collaborative prediction ≤1.0℃² and the health status prediction error ≤5%.

[0028] Beneficial effects: By fusing multimodal parameters and fine-tuning the network, the accuracy of temperature field prediction is maintained while the collaborative prediction function of battery health status is expanded, realizing the integrated perception of "temperature field-health status", providing more comprehensive data support for the thermal health management of power batteries throughout their entire life cycle, and further enhancing the engineering practical value and scalability of the method.

[0029] Working principle and advantages: (I) Working Principle The core idea of ​​this invention is to solve the problem of accurate prediction of the three-dimensional temperature field of power batteries through multi-source data fusion and cross-domain technology collaboration. Its working principle is as follows: 1. Multi-source data acquisition and preprocessing: Local temperature data and operating parameters during actual vehicle operation are collected through sparsely distributed temperature sensors. After outlier processing, missing value imputation and normalization, high-quality input data is obtained. 2. Open source simulation modeling and label generation: Based on OpenFOAM, a three-dimensional unsteady heat conduction simulation model is built to simulate the temperature evolution process under multiple working conditions and generate high-resolution three-dimensional temperature field data as supervision labels for deep learning; 3. Boundary condition extrapolation: Gaussian process regression method is used to extrapolate the approximately continuous boundary field based on finite sensor boundary data, which makes up for the lack of boundary information and improves the integrity of input data; 4. Deep learning model training: Construct a Transformer network that integrates spatial coding and physical priors. With sensor data, extrapolated boundary values ​​and operating parameters as inputs and simulated temperature field as labels, train the model through a loss function that includes physical constraints to ensure that the prediction results meet both numerical accuracy and thermodynamic laws. 5. Lightweight Model Deployment and Closed-Loop Control: The trained model is pruned and quantized to achieve lightweight processing. It is then deployed on the BMS embedded platform to receive sensor data in real time and output a three-dimensional temperature field. This forms a closed-loop control with the BMS thermal management module, enabling temperature anomaly early warning and proactive thermal management.

[0030] (II) Advantages 1. Three-dimensional full-domain temperature sensing with more complete coverage: Compared with the traditional sparse sensor point measurement method, the present invention realizes three-dimensional temperature field reconstruction through deep learning, which can restore the temperature information of any location inside the battery pack, including blind areas not covered by sensors, effectively avoiding the problem of missed detection of local overheating. 2. Data-driven and physical constraints work together for higher prediction accuracy: By integrating open-source simulation data and real vehicle sensor data, the problem of scarce real vehicle data samples is solved. Furthermore, by introducing physical constraints such as heat conduction equations and temperature gradient penalties, the model avoids producing prediction results that do not conform to thermodynamic laws. The prediction error MSE ≤ 0.8℃², which is significantly better than existing pure data-driven or pure physical simulation methods. 3. Enhanced robustness through boundary condition extrapolation optimization: By using Gaussian process regression to extrapolate boundary conditions, the missing boundary information caused by limited sensors is compensated for, enabling the model to maintain stable prediction performance under different operating conditions and battery aging states, and significantly improving generalization ability. 4. Lightweight design and better BMS integration: Through model pruning and quantization, the number of parameters is compressed to 30%-50% of the original model, and the inference latency is ≤50ms. It can be directly deployed on ARM CorTx-M type embedded chips and seamlessly integrated with BMS to realize real-time temperature field prediction and closed-loop control. 5. Controllable cost and strong engineering applicability: Only 10-30 temperature sensors (accounting for 5%-15% of the total number of battery cells) need to be deployed, which greatly reduces the cost of sensors and wiring complexity. At the same time, the cost of real vehicle testing is reduced through the open source simulation platform, which has significant engineering promotion value. Attached Figure Description

[0031] Figure 1 This is a flowchart of the prediction method of the present invention. Detailed Implementation

[0032] The following detailed description illustrates the specific implementation method: like Figure 1 As shown, the three-dimensional temperature field prediction method for power batteries based on boundary condition extrapolation and deep learning of the present invention includes the following steps: Step 1: Multi-source heterogeneous data acquisition Within the power battery pack, several key temperature sensors are pre-installed, forming a sparsely distributed sensing network that covers typical heat-generating areas, boundaries, and locations prone to thermal gradients. During data acquisition, the sensors record local temperature data in real time under the battery pack's operating conditions, including timestamps (t), current (I), voltage (V), state of charge (SOC), ambient temperature (T), and other operating parameters, forming a structured raw operating dataset. Sensor coordinates are obtained through pre-setting or calibration during the design phase, ensuring the spatial integrity of the data.

[0033] Assuming the battery pack contains N temperature sensors, the spatiotemporal data matrix is ​​constructed as follows: . Step 2: Data Cleaning and Normalization The raw data is preprocessed to remove extreme outliers (such as transient mutations, communication interruptions, etc.). Use 3 The rules govern outlier handling: ,in, The current value, This is the mean.

[0034] Use Kriging interpolation to fill in missing data: ,in yes The weight value, In order to be in Temperature value at any given time.

[0035] Then, different physical quantities (temperature, current, voltage, etc.) are normalized separately to eliminate dimensional differences and improve model convergence speed and training stability. Min-Max normalization is selected: , For the minimum value of the corresponding physical quantity, This represents the maximum value of the corresponding physical quantity.

[0036] Step 3: Construct a three-dimensional simulation temperature field model Based on the actual geometry of the battery pack, its thermal properties (such as thermal conductivity, specific heat capacity, and density), and power distribution characteristics, a three-dimensional unsteady-state heat conduction equation is established on an open-source CFD simulation platform (such as OpenFOAM). , Density, unit: kg / m³ 3 ; Specific heat capacity, unit ; Thermal conductivity, unit ; This represents the temperature difference.

[0037] Among them: Joule heat source item: Multiple simulation experiments were conducted with different boundary conditions and operating conditions (charge / discharge rate, ambient temperature, cooling fan speed, etc.). Boundary condition settings: Cooling plate contact surface: , in The coefficient of thermal conductivity is represented by h, which represents the surface convective heat transfer coefficient. This indicates the temperature of the cooling plate.

[0038] Battery pack casing: . in h represents thermal conductivity. conv Indicates the convective heat transfer coefficient of the outer shell surface. This indicates the temperature of the battery pack casing.

[0039] Output the temperature value of each grid node in three-dimensional space as it evolves over time, forming a high-resolution complete thermal field dataset.

[0040] Step 4: Simulation data is used for supervised learning label construction. Using the three-dimensional thermal field obtained from the simulation in step S3 as a supervised learning label, and combining it with the actual number and location of measurement points, input-output paired samples are constructed. Multiple thermal evolution paths are simulated at different charge / discharge rates, SOC, and ambient temperatures, covering potential operating conditions, and generating a representative and generalizable training set.

[0041] Operating conditions covered: . Step 5: Approximate Continuity Boundary Estimation Considering the limited number of sensor points, a spatial interpolation Gaussian process regression method based on physical laws is used to reconstruct the boundary to compensate for the lack of boundary data. , in Represents the mean function, This represents the covariance function.

[0042] Kernel function selection: , in This represents the Euclidean distance between two points. Represents the signal variance. Indicates a length scale.

[0043] Hyperparameters Optimize using maximum likelihood estimation.

[0044] Based on existing boundary point data, the neighborhood temperature distribution is inferred, constructing a continuous boundary value estimation field. This process provides more realistic and consistent boundary information input for subsequent neural networks.

[0045] Step 6: Construct the neural network input vector This results in a high-dimensional input feature vector that includes the following elements: The temperature values ​​and spatial coordinate positions of each sensor are encoded; The boundary temperature field value after interpolation; External operating parameters (current, voltage, ambient temperature, time).

[0046] All input data are synchronized and aligned within a unified time step to form the input end of the training samples.

[0047] Step 7: Design the deep neural network structure A neural network architecture integrating spatial encoding and prior physical knowledge is constructed. Based on the Transformer, the modeling ability for long-range spatial dependencies is enhanced; at the same time, a physical constraint module is introduced to assist in feature extraction. Gradient penalty term: . in This represents the predicted temperature gradient vector. This represents the gradient magnitude threshold.

[0048] The network can include structures such as position encoding, multi-head attention, and residual modules to improve the model's ability to capture complex heat conduction processes.

[0049] Position encoding: Introducing sinusoidal encoding for spatial coordinates: , in A one-dimensional mapping representing spatial location, This represents the hidden layer dimension of the Transformer model. This represents the position encoding matrix.

[0050] Multi-head attention: Computing spatial correlation weights: , in This represents the feature matrix obtained by projecting the input sequence. This represents the dimension of a vector. For the number of attention heads, This represents the normalized attention weight matrix.

[0051] Residual regularization: .

[0052] in Indicates the pyrogen source item. This represents the Laplace expression for predicting the temperature field.

[0053] Step 8: Model Training and Physical Consistency Optimization Supervised learning is employed, using simulated temperature fields as labels for model training. The loss function comprises several components: basic mean squared error (MSE), gradient consistency loss (the difference between the temperature gradient and the physical derivative), boundary consistency loss, etc. , in Indicates the first Individual element models predict temperature. Indicates the first Finite element simulation temperature of individual elements This represents the weighting coefficient of the physical constraint loss term.

[0054] To ensure that the model output has both numerical accuracy and conforms to thermodynamic laws, an adaptive weight adjustment strategy can be used to balance the various loss terms.

[0055] Step 9: Model Evaluation and Test Set Validation Using the preserved test dataset, the model's predictive performance under unseen operating conditions is evaluated.

[0056] Indicator 1: Mean Temperature Prediction Error (MSE) in Three-Dimensional Space: .

[0057] in This represents the predicted temperature field value. This represents the true value of the temperature field.

[0058] Indicator 2: Spatial gradient error: , in This represents the three-dimensional space of the entire battery pack. These represent the predicted and true temperature gradient fields, respectively.

[0059] Assess whether key thermal field features such as drastic boundary changes and hotspot edges are correctly identified.

[0060] Step 10: Lightweight Model and Edge Deployment After model training, model pruning, knowledge distillation, and quantization are used to compress the model size and reduce computational complexity. The model is then converted to formats suitable for edge computing, such as TnsorRT and ONNX, and deployed on embedded platforms such as BMS control chips, NPUs, and FPGAs to achieve low-power, low-latency operation and meet practical operational requirements.

[0061] Step 11: Online Inference and Real-Time Temperature Field Reconstruction After the model is deployed, the system collects the current sensor temperature values ​​and operating parameters in real time during battery operation and inputs them into the model for inference. The model outputs a complete three-dimensional temperature field at the current moment, realizing the dynamic reconstruction of the internal heat distribution of the battery pack, and periodically refreshes the thermal field information (e.g., once per second), providing high-frequency thermal status feedback for the BMS system.

[0062] Step 12: Integrate with BMS to form a closed-loop control system The model output is deeply integrated with the BMS thermal management module, forming a feedback control closed loop. When the temperature or gradient in a certain area of ​​the predicted thermal field exceeds a set threshold, it can automatically trigger operations such as turning on the cooling fan, adjusting the charging power, and issuing fault warnings. The system supports expansion to multi-modal fusion prediction, such as combining SOC, SOH, internal resistance, and thermal stress, to further improve the thermal safety and health management level of the battery system.

[0063] Specifically, Example 1: A three-dimensional temperature field prediction method and system applicable to passenger vehicle power battery packs (a) Implementation of the method 1. Data Acquisition: A ternary lithium battery pack for a passenger vehicle (total of 192 cells) was selected, and 16 temperature sensors (accounting for 8.3%) were deployed. The sensor locations included 4 in the center area of ​​the battery pack, 6 at the module boundary, 3 at the contact surface of the cooling plate, and 3 at the electrode tabs. The collected parameters included sensor coordinates (xi,yi,zi), current I, voltage V, SOC, and ambient temperature T. The sampling frequency was 1Hz.

[0064] 2. Data preprocessing: Outliers were removed using the 3σ rule (mean temperature T=32℃, standard deviation σT=5℃, data with Ti>47℃ or Ti<17℃ were removed); missing data were filled in using Kriging interpolation; and all parameters were mapped to the [0,1] interval using Min-Max normalization.

[0065] 3. Simulation Modeling: Based on the actual geometric dimensions of the battery pack (500mm long × 200mm wide × 150mm high), the thermal property parameters are set as follows: kx = 11.2W / (m³). K), ky = 3.1 W / (m K), kz = 8.3W / (m K), specific heat capacity c = 800 J / (kg) K), density ρ = 2650 kg / m³; a three-dimensional unsteady-state heat transfer model was established in OpenFOAM, with boundary conditions set as follows: the heat transfer coefficient between the cooling plate and the surface is h = 50 W / (m²). K), the emissivity of the battery pack casing ε=0.8; the simulated operating conditions cover 0.2C, 0.5C, 1C, 2C, 3C charge and discharge rates, ambient temperatures of -20℃, -10℃, 0℃, 25℃, 40℃, 60℃, and SOC of 10%, 30%, 50%, 70%, 90%, for a total of 150 operating conditions.

[0066] 4. Boundary Extrapolation: Gaussian process regression is used for boundary extrapolation, with a squared exponential kernel function. Hyperparameter optimization is iterated 150 times, and the convergence condition is that the update amount < 1e. -5 This yields a continuous boundary temperature field.

[0067] 5. Model Training: A Transformer network was constructed with 6 multi-head attention heads and 384 hidden layer neurons. The activation function was tanh. The weights of the total loss function were set to Wmse=1.0, Wgrad=0.5, and Wbound=0.8. The Adam optimizer was used with a learning rate of 1e. -4 The training iterations were repeated 1000 times.

[0068] 6. Model lightweighting and deployment: Structured pruning was used to remove redundant channels, reducing the number of parameters from 8.2M to 3.1M (compression rate of 62.2%). Through INT8 quantization, the inference latency was reduced to 38ms. The model was converted to TnsorRT format and deployed on a BMS embedded chip (ARMCorTx-M7).

[0069] 7. Closed-loop control: Set temperature thresholds: T≥80℃ triggers a first-level warning and starts the cooling fan at its maximum speed (3000r / min); 60℃≤T<80℃ triggers a second-level warning and adjusts the cooling fan speed to 2000r / min; T<60℃ maintains the normal speed (1000r / min).

[0070] (II) System Implementation Systems that implement this method include: 1. Data acquisition module: 16 NTC temperature sensors (measuring range -40℃-125℃, accuracy ±0.5℃), current sensor, voltage sensor, and ambient temperature sensor; 2. Data Preprocessing Module: A real-time data processing unit based on MCU, which realizes outlier deletion, interpolation completion, and normalization; 3. Simulation Modeling Module: The OpenFOAM simulation platform deployed in the cloud is used to generate training labels; 4. Model Training Module: Cloud-based GPU training cluster (NVIDIA A100) for model training and optimization; 5. Model Deployment Module: BMS embedded chip (ARMCorTx-M7), running lightweight models; 6. Closed-loop control module: Communicates with the BMS thermal management unit and outputs temperature field data and control commands.

[0071] (III) Implementation Results 1. Model accuracy: Three-dimensional spatial average temperature prediction error =0.62℃², spatial gradient error =0.12℃ / mm, which meets the requirements of engineering applications; 2. Real-time performance: Inference latency of 38ms, temperature field data updated 26 times per second, meeting the real-time requirements of BMS; 3. Safety warning: Under 3C discharge conditions, it can detect that the temperature of the electrode area exceeds 80°C 2.3 seconds in advance, triggering a first-level warning and effectively preventing thermal runaway; 4. Thermal management optimization: Through precise temperature field data, the cooling system energy consumption is reduced by 15%, and the battery pack cycle life is improved by 8%.

[0072] (iv) Differences from and non-obviousness of existing technologies 1. Compared with the prior art CN115879370A: This prior art only uses a high-precision heat generation model combined with a neural network to predict the temperature field, without involving boundary condition extrapolation optimization, and does not specify a closed-loop integration scheme with BMS; This invention solves the problem of missing boundary information from sparse sensors by extrapolating boundary conditions through Gaussian process regression, while achieving lightweight model and deep integration with BMS, resulting in lower inference latency (38ms vs. the prior art not specified, speculated to be ≥100ms) and better integration; 2. Compared with the prior art CN118654772A: This prior art relies on the difference in thermocouple deployment between the sample battery pack and the battery pack under test to achieve prediction, and does not integrate physical simulation data and boundary extrapolation technology; This invention generates massive amounts of labeled data through open-source simulation, without relying on multiple sets of battery pack experiments, reducing experimental costs, while the boundary extrapolation technology improves prediction accuracy (MSE=0.62℃² vs. the prior art's predicted MSE≥1.2℃²). 3. Compared with the prior art CN120337774A: the prior art focuses on the prediction of the surface temperature field of the battery module, does not involve the reconstruction of the three-dimensional global temperature field, and does not consider the integration with the BMS; the present invention realizes the prediction of the three-dimensional global temperature field, which can cover the key areas inside the battery, and at the same time, it is adapted to the BMS embedded platform through lightweight design, making it more practical for engineering.

[0073] Example 2: A three-dimensional temperature field prediction method and system for commercial vehicle power battery packs (a) Implementation of the method 1. A lithium iron phosphate battery pack for a commercial vehicle (total of 384 cells) was selected, and 24 temperature sensors (accounting for 6.25%) were installed. The sensor locations include 6 in the center area of ​​the battery pack, 8 at the module boundary, 5 at the contact surface of the cooling plate, and 5 at the electrode tabs. 2. The steps of data preprocessing, simulation modeling, boundary extrapolation, and model training are the same as in Example 1, only the thermophysical parameter is adjusted: kx = 10.8 W / (m K), ky = 2.8 W / (m K), kz = 7.9 W / (m K), specific heat capacity c=850J / (kg) K), density ρ = 2700 kg / m³; 3. After model lightweighting, the number of parameters was reduced from 9.5M to 3.8M (compression rate of 59%), and the inference latency was 42ms; 4. Closed-loop control threshold: Level 1 warning is triggered when T ≥ 85℃, and Level 2 warning is triggered when 65℃ ≤ T < 85℃.

[0074] (II) Implementation Results 1. Model accuracy: Three-dimensional spatial average temperature prediction error =0.75℃², spatial gradient error =0.14℃ / mm; 2. Real-time performance: Inference latency of 42ms meets the real-time requirements of commercial vehicle BMS; 3. Application effect: Under 2C charging conditions, the temperature of the central area of ​​the battery pack exceeds 85°C 3.1 seconds in advance, triggering the cooling system to run at full load and preventing the temperature from continuing to rise.

[0075] The non-obviousness of this invention: In the prior art, commercial vehicle power battery packs are large in size and have a large number of cells, so more sensors (accounting for ≥20%) are usually required to ensure the temperature monitoring effect. This invention, through the fusion technology of boundary extrapolation and deep learning, can achieve high-precision prediction of three-dimensional temperature field with only 6.25% of the sensors, breaking through the traditional understanding that "the number of sensors determines the monitoring accuracy". The technical solution is non-obvious. The inventiveness of this invention: Compared with the prior art, this invention reduces sensor cost by 68% while improving prediction accuracy by more than 30%, and achieves seamless integration with BMS. It solves the technical problem of "difficulty in balancing cost and accuracy" in the thermal management of commercial vehicle power batteries, and has significant inventiveness.

[0076] Comparative Example: Unbounded Extrapolation Comparative Experiment (I) Experimental Design Using the passenger vehicle power battery pack of Example 1, only the "boundary extrapolation" step was removed, while the remaining steps remained the same, to construct a comparative model.

[0077] (II) Experimental Results Comparison of the three-dimensional spatial average temperature prediction error of the model =1.85℃², spatial gradient error =0.32℃ / mm, inference delay 36ms; compared with Example 1, MSE is improved by 198%, spatial gradient error is improved by 167%, and multiple temperature abrupt points that do not conform to thermodynamic laws appear in the boundary region.

[0078] (III) Conclusion The boundary extrapolation step can effectively compensate for the lack of boundary information in sparse sensors, significantly improve the model prediction accuracy and physical consistency, and is one of the key technologies for achieving high-precision temperature field prediction in this invention, further proving the inventiveness of the technical solution of this invention.

[0079] The above descriptions are merely embodiments of the present invention, and common knowledge such as specific technical solutions and / or characteristics are not described in detail here. It should be noted that those skilled in the art can make various modifications and improvements without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application should be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.

Claims

1. A method for predicting the three-dimensional temperature field of a power battery based on boundary condition extrapolation and deep learning, characterized in that, Includes the following steps: S1: Collect local temperature data under different operating conditions through N temperature sensors preset in the power battery pack. The data includes sensor coordinates (xi,yi,zi), current I, voltage V, SOC and ambient temperature T, where N is an integer from 10 to 30. S2: Perform outlier removal, missing value imputation, and normalization on the collected data. Outlier removal adopts the 3σ rule: Where T is the average temperature. This represents the standard deviation of temperature. Missing value imputation uses Kriging interpolation: ,in For the coordinates of the missing data points, For interpolation weights; Normalization uses the Min-Max algorithm: Where X is the original data, , These are the minimum and maximum values ​​of the data, respectively. S3: A three-dimensional unsteady-state heat conduction simulation model is established on the open-source finite element platform OpenFOAM based on the battery pack geometry and thermal properties. The model satisfies the heat conduction equation: , For density, For specific heat capacity, For thermal conductivity, This is the temperature difference value; Boundary conditions include the convective heat transfer boundary at the contact surface of the cooling plate: h is the surface convective heat transfer coefficient. This refers to the temperature of the cooling plate. Radiative heat transfer boundary of the battery pack casing: h conv Indicates the convective heat transfer coefficient of the outer shell surface. Indicates the temperature of the battery pack casing; S4: Generate three-dimensional temperature field data under multiple operating conditions using the simulation model, and use this data as supervised learning labels to construct a training dataset. The operating conditions coverage must meet the following requirements: , Where 0.2C ≤ expansion rate ≤ 3C, -20℃ ≤ T ≤ 60℃, and 10% ≤ SOC ≤ 90%; S5: The boundary conditions are extrapolated using the Gaussian process regression method to construct an approximately continuous boundary field. The regression model is as follows: ， Kernel function selection: , hyperparameters Optimize using maximum likelihood estimation; S6: Construct a high-dimensional input vector, which includes sensor temperature values ​​and spatial coordinate encoding, extrapolated boundary temperature values, and operating parameters. The spatial coordinate encoding uses sinusoidal encoding. , A one-dimensional mapping representing spatial location, This represents the hidden layer dimension of the Transformer model. Represents the position encoding matrix; S7: Construct a deep learning network based on Transformer, which includes a positional encoding layer, a multi-head attention layer, and a residual module. The multi-head attention weights are calculated to satisfy the following: , This represents the feature matrix obtained by projecting the input sequence. This represents the dimension of a vector. For the number of attention heads, This represents the normalized attention weight matrix; Residual regularization satisfies: , Indicates the pyrogen source item. Represents the Laplace of the predicted temperature field; S8: Introduce physical constraints to optimize model training. The loss function includes mean squared error loss, temperature gradient penalty loss, and boundary consistency loss. The total loss function is: , Indicates the first Individual element models predict temperature. Indicates the first Finite element simulation temperature of individual elements The weighting coefficients for the physical constraint loss term; The gradient penalty loss satisfies: , This represents the predicted temperature gradient vector. Indicates the gradient magnitude threshold; S9: Verify model performance using the test set, employing the 3D spatial mean temperature prediction error (MSE). , This represents the predicted temperature field value. Represents the true value of the temperature field; and the spatial gradient error: As an evaluation indicator This represents the three-dimensional space of the entire battery pack. These represent the predicted and true temperature gradient fields, respectively. S10: Prune and quantize the model, convert it to TnsorRT or ONNX format, deploy it in BMS, output the three-dimensional temperature field in real time, and form a closed-loop control with BMS.

2. The method according to claim 1, characterized in that, The temperature sensor placement locations in step S1 include the central heating area of ​​the battery pack, the module boundary, the contact surface of the cooling plate, and the cell tab location. The number of sensors N satisfies: total number of battery pack cells × 5% ≤ N ≤ total number of battery pack cells × 15%.

3. The method according to claim 1, characterized in that, The thermophysical parameters mentioned in step S3 include thermal conductivity kx = 10⁻¹² W / (m²). K), ky = 2-4W / (m K), kz = 7-9W / (m K), specific heat capacity c = 700-900 J / (kg) K), density ρ = 2500-2800 kg / m³.

4. The method according to claim 1, characterized in that, In step S5, the hyperparameter optimization iterations of the kernel function for Gaussian process regression are no less than 100 times, and the convergence condition is that the hyperparameter update amount is less than 1×e. -5 .

5. The method according to claim 1, characterized in that, In step S7, the Transformer network has 4-8 multi-head attention heads, 256-512 hidden layer neurons, and uses the tanh function as the activation function.

6. The method according to claim 1, characterized in that, The weights of each loss term in the total loss function described in step S8 satisfy the following conditions: mean square error loss weight Wmse = 1.0, temperature gradient penalty loss weight Wgrad = 0.3-0.7, and boundary consistency loss weight Wbound = 0.5-1.

0.

7. The method according to claim 1, characterized in that, The qualification criterion for model evaluation in step S9 is: the three-dimensional spatial average temperature prediction error. ≤0.8℃², spatial gradient error ≤0.15℃ / mm.

8. The method according to claim 1, characterized in that, After model quantization in step S10, the number of parameters is compressed to 30%-50% of the original model, and the inference latency is ≤50ms.

9. The method according to claim 1, characterized in that, The method also includes a closed-loop control step with the BMS: when the temperature T ≥ 80℃ in any region of the predicted three-dimensional temperature field, a first-level warning is triggered and maximum intensity cooling is started; when 60℃ ≤ T < 80℃, a second-level warning is triggered and cooling power is adjusted; when T < 60℃, the conventional thermal management strategy is maintained.

10. The method according to claim 1, characterized in that, The method further includes a multimodal fusion extension step: the SOC, SOH, internal resistance and thermal stress parameters of the battery pack are integrated into a high-dimensional input vector, and the multi-head attention weights of the Transformer network are fine-tuned to achieve collaborative prediction of the three-dimensional temperature field and the battery health status. The mean square error of the collaborative prediction is ≤1.0℃², and the health status prediction error is ≤5%.

Images