A temperature field control method and system for lithium battery dynamic working conditions

Abstract

The present application relates to the technical field of lithium battery, and provides a temperature field control method and system for dynamic working condition of lithium battery, the method comprising: constructing a dynamic temperature field prediction model comprising a multi-modal feature encoder, a Fourier neural operator, a spatial high-frequency compensation branch, a weight network and a spatial gating fusion module; wherein the multi-modal feature encoder is used to construct a global multi-physical feature matrix, the Fourier neural operator is used to obtain low-frequency macro features for representing macro temperature rise trend, the spatial high-frequency compensation branch is used to obtain high-frequency detail features, the weight network is used to obtain dynamic weight coefficients, and the spatial gating fusion module is used to superimpose the high-frequency detail features to the low-frequency macro features according to the dynamic weight coefficients to obtain a predicted three-dimensional temperature field; and performing temperature field control based on the trained dynamic temperature field prediction model. The present application avoids the problem of response lagging of existing temperature field prediction algorithms and can realize millisecond-level real-time deduction.

Description

Technical Field

[0001] This invention relates to the field of lithium battery technology, and more specifically, to a method and system for temperature field control under dynamic operating conditions of lithium batteries. Background Technology

[0002] With the widespread application of electric vehicles, energy storage systems, and various high-energy-density devices, lithium-ion batteries play a crucial role in these fields. However, lithium batteries generate a significant amount of heat during charging and discharging, especially under dynamic operating conditions. Due to frequent load fluctuations, battery heat generation is drastic and variable. If this heat is not dissipated effectively and promptly, it will lead to an abnormal rise in battery temperature, accelerating battery capacity degradation and potentially causing serious safety accidents such as thermal runaway. Faced with the high-frequency fluctuations in dynamic charging and discharging conditions during actual operation, the lag of traditional passive or static liquid cooling strategies is becoming increasingly apparent.

[0003] Currently, the coupled calculation of flow and temperature fields mainly relies on traditional computational fluid dynamics numerical simulation, which can provide highly accurate multiphysics coupling results. However, its calculation process involves highly intensive nonlinear solutions, which are extremely time-consuming and cannot meet the timeliness requirements of systems under dynamic conditions for millisecond-level online prediction and active feedback control. Summary of the Invention

[0004] To address the problem that existing methods for coupled flow and temperature field calculations are extremely time-consuming and cannot meet the timeliness requirements of millisecond-level online prediction and active feedback control under dynamic operating conditions, this invention provides a temperature field control method and system for lithium batteries under dynamic operating conditions. The specific technical solution is as follows:

[0005] A method for temperature field control under dynamic operating conditions of lithium batteries includes:

[0006] A dynamic temperature field prediction model is constructed, comprising a multimodal feature encoder, a Fourier neural operator, a spatial high-frequency compensation branch, a weight network, and a spatial gated fusion module. The multimodal feature encoder extracts independent high-dimensional features from the multidimensional boundary conditions of the battery pack, obtaining latent feature vectors of the same dimension. These latent feature vectors are then concatenated along the feature channel dimension to construct a global multi-physical feature matrix. The Fourier neural operator processes this global multi-physical feature matrix to obtain low-frequency macroscopic features characterizing the macroscopic temperature rise trend. The spatial high-frequency compensation branch extracts local features and performs nonlinear mapping on the low-frequency macroscopic features, obtaining high-frequency detail features reflecting local thermodynamic abrupt changes. The weight network extracts the overall severity of the current operating condition based on the low-frequency macroscopic features, obtaining dynamic weight coefficients to determine the fusion ratio for detail adjustment. The spatial gated fusion module superimposes high-frequency detail features onto low-frequency macroscopic features based on the dynamic weight coefficients to obtain the predicted three-dimensional temperature field.

[0007] The dynamic temperature field prediction model is trained to obtain the trained dynamic temperature field prediction model, and temperature field control is performed based on the trained dynamic temperature field prediction model.

[0008] The temperature field control method for lithium batteries under dynamic operating conditions constructs a dynamic temperature field prediction model that includes a multimodal feature encoder, a Fourier neural operator, a spatial high-frequency compensation branch, a weight network, and a spatial gating fusion module. It does not rely on external temperature sensors for hysteresis feedback. By utilizing the real-time multidimensional boundary conditions of the battery pack, it transforms the slow macroscopic physical heat transfer process into a fast state mapping within a microsecond-level algorithm space. This reduces the risk of control overshoot and thermal runaway caused by thermal response delay, avoiding the problems of high computation time and response hysteresis in existing temperature field prediction algorithms. Furthermore, the model, once solidified, has a low online computation dimension, enabling millisecond-level real-time extrapolation. This solves the problem that existing coupled calculations of flow field and temperature field are extremely time-consuming and cannot meet the timeliness requirements of millisecond-level online prediction and active feedback control under dynamic operating conditions.

[0009] Preferably, the specific methods for training the dynamic temperature field prediction model include:

[0010] The Latin hypercube sampling method is used to perform hierarchical orthogonal sampling within a set physical threshold range to generate N sets of multidimensional boundary condition combinations. Among them, the N sets of multidimensional boundary condition combinations include the battery multidimensional dynamic electrical state sequence and the coolant targeted flow velocity allocation vector of the multi-inlet under the corresponding working conditions.

[0011] The N sets of multidimensional boundary conditions are input into the computational fluid dynamics solver to perform CFD numerical simulation and obtain the true three-dimensional temperature field on the spatial grid nodes.

[0012] The battery multidimensional dynamic electrical state sequence and the coolant targeted flow velocity allocation vector are input into the dynamic temperature field prediction model to obtain the predicted three-dimensional temperature field.

[0013] Error comparison and parameter optimization are performed based on the true three-dimensional temperature field and the predicted three-dimensional temperature field.

[0014] Preferably, the specific methods for error comparison and parameter optimization include:

[0015] Construct a physical embedded joint loss function that includes data fidelity loss, physical equation residual loss, and high-frequency regularization loss;

[0016] Based on the physical embedded joint loss function, the joint loss for each iteration cycle is calculated in real time, and the state is determined based on the convergence threshold and the joint loss.

[0017] If the joint loss is not less than the convergence threshold, the Adam optimizer is started, and the gradient descent algorithm is used to backpropagate the joint error, dynamically adjust and optimize the weight parameters inside the dynamic temperature field prediction model.

[0018] If the joint loss is less than the convergence threshold, the dynamic temperature field prediction model is solidified.

[0019] Preferably, the specific methods for temperature field control include:

[0020] Obtain the current multidimensional boundary conditions of the lithium battery and the multi-inlet flow rate allocation scheme currently being executed by the cooling mechanism;

[0021] Input the current multidimensional boundary conditions and multi-inlet velocity allocation scheme into the trained dynamic temperature field prediction model to obtain the predicted three-dimensional temperature field under the current multi-inlet velocity allocation scheme.

[0022] The deviation between the predicted three-dimensional temperature field and the target temperature field is obtained, and the temperature field is controlled based on the deviation.

[0023] Preferably, the specific method for controlling the temperature field based on the degree of deviation includes:

[0024] If the deviation is less than the over-limit threshold, maintain the original instruction;

[0025] If the deviation is not less than the threshold, the improvement mechanism is activated, and the particle swarm optimization (PSO) algorithm is used to generate a series of trial flow rate schemes based on the current real flow rate allocation scheme.

[0026] The trial flow velocity scheme is fed back into the dynamic temperature field prediction model for a new round of rapid prediction. Through continuous optimization and iteration, the optimal flow velocity scheme for multiple inlets is obtained.

[0027] A temperature field control system for lithium batteries under dynamic operating conditions, used to implement the aforementioned temperature field control method, includes a dynamic temperature field prediction model, which comprises:

[0028] The multimodal feature encoder is used to extract high-dimensional features independently from the multidimensional boundary conditions of the battery pack, obtain latent feature vectors of the same dimension, and concatenate the independent latent feature vectors in the feature channel dimension to construct a global multi-physical feature matrix.

[0029] Fourier neural operators are used to process the global multi-physics feature matrix to obtain low-frequency macroscopic features that characterize the macroscopic temperature rise trend.

[0030] The spatial high-frequency compensation branch is used to extract local features and perform nonlinear mapping on low-frequency macroscopic features to obtain high-frequency detailed features that reflect local thermodynamic abrupt changes.

[0031] A weighted network is used to extract the overall severity of the current working condition based on low-frequency macroscopic features, and to obtain dynamic weight coefficients for determining the fusion ratio of detailed adjustments.

[0032] The spatial gating fusion module is used to superimpose high-frequency detail features onto low-frequency macroscopic features based on dynamic weighting coefficients to obtain the predicted three-dimensional temperature field.

[0033] Preferably, the temperature field control system further includes:

[0034] The multidimensional boundary condition combination acquisition module uses the Latin hypercube sampling method to perform hierarchical orthogonal sampling within a set physical threshold range to generate N sets of multidimensional boundary condition combinations. Among them, the N sets of multidimensional boundary condition combinations include the battery multidimensional dynamic electrical state sequence and the coolant targeted flow velocity allocation vector of the multi-inlet under the corresponding working conditions.

[0035] Computational fluid dynamics solver, used to perform CFD numerical simulation based on N sets of multidimensional boundary conditions, and obtain the true three-dimensional temperature field on the spatial grid nodes;

[0036] The physical embedded joint optimization module is used to compare errors and optimize parameters based on the true three-dimensional temperature field and the predicted three-dimensional temperature field.

[0037] Preferably, the physical embedded joint optimization module includes:

[0038] The loss function construction unit is used to construct a physical embedded joint loss function that includes data fidelity loss, physical equation residual loss, and high-frequency regularization loss.

[0039] The state determination unit is used to determine the state based on the convergence threshold and the joint loss. If the joint loss is not less than the convergence threshold, the Adam optimizer is started, and the gradient descent algorithm is used to backpropagate the joint error to dynamically adjust and optimize the weight parameters inside the dynamic temperature field prediction model. Otherwise, the dynamic temperature field prediction model is fixed to obtain a trained dynamic temperature field prediction model.

[0040] Preferably, the temperature field control system further includes:

[0041] The battery pack current parameter acquisition module is used to acquire the current multidimensional boundary conditions of the lithium battery and the multi-inlet flow velocity allocation scheme executed by the cooling mechanism at the current moment, and input the current multidimensional boundary conditions and multi-inlet flow velocity allocation scheme into the trained dynamic temperature field prediction model to obtain the predicted three-dimensional temperature field under the current multi-inlet flow velocity allocation scheme.

[0042] The optimization module is used to obtain the degree of deviation between the predicted three-dimensional temperature field and the target temperature field, and to obtain the optimal flow velocity scheme for multiple inlets based on the degree of deviation.

[0043] Preferably, the optimization module includes:

[0044] The trial flow rate scheme generation unit is used to activate the improvement mechanism when the deviation is not less than the over-limit threshold. It uses the particle swarm optimization algorithm (PSO) to generate a series of trial flow rate schemes based on the current real flow rate allocation scheme.

[0045] The optimal flow rate scheme acquisition unit is used to feed back the trial flow rate scheme to the dynamic temperature field prediction model for a new round of rapid prediction. Through continuous optimization and iteration, the optimal flow rate scheme for multiple inlets is obtained. Attached Figure Description

[0046] The invention will be further understood from the following description taken in conjunction with the accompanying drawings. The components in the drawings are not necessarily drawn to scale, but rather the emphasis is on illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0047] Figure 1 This is a schematic diagram of the overall process of a temperature field control method for lithium battery dynamic operating conditions in one embodiment of the present invention.

[0048] Figure 2 This is a flowchart illustrating a specific method for training a dynamic temperature field prediction model in one embodiment of the present invention.

[0049] Figure 3 This is a flowchart illustrating a specific method for error comparison and parameter optimization in one embodiment of the present invention.

[0050] Figure 4This is a schematic flowchart of a specific method for temperature field control in one embodiment of the present invention;

[0051] Figure 5 This is a flowchart illustrating a specific method for controlling the temperature field based on the degree of deviation in one embodiment of the present invention.

[0052] Figure 6 This is a schematic diagram of the overall structure of a dynamic temperature field prediction model in one embodiment of the present invention;

[0053] Figure 7 This is a schematic diagram of the training process of a dynamic temperature field prediction model in one embodiment of the present invention;

[0054] Figure 8 This is a schematic diagram of the dynamic temperature field prediction model working in a cooling system according to an embodiment of the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to its embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of the invention.

[0056] Before describing the specific embodiments of the present invention, a brief introduction to the prior art will be given first.

[0057] With the widespread application of electric vehicles, energy storage systems, and various high-energy-density devices, lithium-ion batteries play a crucial role in these fields. However, lithium batteries generate a significant amount of heat during charging and discharging, especially under dynamic operating conditions where frequent load fluctuations lead to dramatic and variable heat generation. If this heat is not dissipated effectively and promptly, it can cause abnormally high battery temperatures, accelerating battery capacity degradation and potentially triggering serious safety incidents such as thermal runaway.

[0058] Faced with the high-frequency fluctuations of dynamic charging and discharging conditions in actual operation, the lag of traditional passive or static liquid cooling strategies is becoming increasingly apparent. To achieve high-precision temperature control and suppress the risk of thermal runaway, an ideal active thermal management system must possess feedforward regulation capabilities based on the transient heat generation characteristics of the battery. This requires the system to dynamically adjust the coolant flow rate distribution in the multi-inlet heat dissipation structure to reduce the amplitude of temperature changes inside the battery pack. The core challenge in achieving this goal lies not only in ensuring the real-time high-fidelity prediction of the three-dimensional temperature field, but also in incorporating the active response delay of the liquid cooling actuator and the time difference in temperature field evolution caused by system thermal inertia into the design of the control loop.

[0059] Currently, the coupled calculation of flow and temperature fields mainly relies on traditional computational fluid dynamics numerical simulation, which can provide highly accurate multiphysics coupling results. However, its calculation process involves highly intensive nonlinear solutions, which are extremely time-consuming and completely fail to meet the timeliness requirements of millisecond-level online prediction and active feedback control under dynamic operating conditions. Meanwhile, existing liquid-cooled feedback control systems mostly rely on temperature sensors placed on the battery pack casing or the surface of the water-cooling plate. However, there is a significant physical thermal inertia from heat generation inside the battery cell to heat conduction through multiple layers of materials to the surface sensor. By the time the external sensor detects a temperature anomaly and triggers a system response, a large amount of heat has often accumulated inside the battery, or even localized overheating has occurred.

[0060] In summary, existing technologies have the following drawbacks:

[0061] 1. Traditional CFD calculations are time-consuming and cannot predict in real time.

[0062] Traditional computational fluid dynamics (CFD) simulations can accurately calculate multiphase flow and heat transfer processes, but due to their extremely high computational resource requirements and slow solution process, they cannot achieve real-time temperature field prediction at the second or millisecond level under dynamic conditions. They can only be used in the heat dissipation system design stage and cannot output results to the heat dissipation system in real time under dynamic conditions.

[0063] 2. Conventional machine learning methods have weak generalization ability and exhibit prediction lag.

[0064] Existing data-driven temperature field prediction models (such as traditional neural networks) rely excessively on static training samples, resulting in poor generalization ability and difficulty in handling boundary drift under dynamic operating conditions. Furthermore, these algorithms lack integration with physical mechanisms, making it difficult to resolve the strong nonlinear coupling between multi-inlet flow fields and three-dimensional temperature fields, thus failing to provide accurate feedforward guidance for thermal management systems. In addition, existing liquid cooling strategies largely rely on discrete data from sparse sensors for passive feedback control. This prediction method requires waiting for heat generated inside the battery to be conducted to the temperature sensor location before responding. Due to the inherent physical delay caused by the system's thermal inertia, such delayed responses are difficult to compensate for in advance, easily leading to temperature control overshoot and irreversible local heat accumulation, thereby exacerbating the risk of battery thermal runaway.

[0065] 3. Active liquid cooling systems have a response time lag.

[0066] Existing battery temperature field prediction and thermal management control systems are mostly based on instantaneous state feedback logic, generally ignoring the inherent time lag effect between actuator actions and heat transfer physical processes. When the control system adjusts the flow rate or volume of the multi-inlet coolant, the fluid transport process in the pipeline and the thermal inertia of the materials of various components inside the battery pack result in a time lag between the change in flow field distribution being transmitted to the entire three-dimensional temperature field and the achievement of the expected cooling or temperature equalization effect. This severe thermal lag is particularly prominent under high-frequency fluctuating dynamic charge and discharge conditions. Without a feedforward prediction mechanism to compensate for this time lag, the thermal management system will always be in a passive response state. This not only leads to excessive intervention by the control system during regulation but may also result in irreversible local heat accumulation inside the battery before the coolant reaches the designated area and performs effective heat exchange. This situation will drastically increase the potential risk of battery thermal runaway, seriously threatening the safe operation of the system and accelerating battery aging.

[0067] One of the objectives of this invention is to address the problem that existing coupled calculations of flow field and temperature field are extremely time-consuming and cannot meet the timeliness requirements of systems under dynamic operating conditions for millisecond-level online prediction and active feedback control. This invention designs a temperature field control method and system that does not rely on external temperature sensors for hysteresis feedback. By utilizing the real-time multidimensional boundary conditions of the battery pack, the slow macroscopic physical heat transfer process is transformed into a fast state mapping in the microsecond-level algorithm space, thereby reducing the risk of control overshoot and thermal runaway caused by thermal response delay.

[0068] Therefore, such as Figure 1 As shown, an embodiment of the present invention provides a temperature field control method for lithium batteries under dynamic operating conditions, comprising the following steps:

[0069] S1. A dynamic temperature field prediction model is constructed, comprising a multimodal feature encoder, a Fourier neural operator, a spatial high-frequency compensation branch, a weight network, and a spatial gated fusion module. The multimodal feature encoder is used to independently extract high-dimensional features from the multidimensional boundary conditions of the battery pack, obtaining latent feature vectors of the same dimension. These latent feature vectors are then concatenated along the feature channel dimension to construct a global multi-physical feature matrix. The Fourier neural operator processes this global multi-physical feature matrix to obtain low-frequency macroscopic features characterizing the macroscopic temperature rise trend. The spatial high-frequency compensation branch performs local feature extraction and nonlinear mapping on the low-frequency macroscopic features to obtain high-frequency detail features reflecting local thermodynamic abrupt changes. The weight network extracts the overall severity of the current operating condition based on the low-frequency macroscopic features and obtains dynamic weight coefficients to determine the fusion ratio for detail adjustment. The spatial gated fusion module superimposes high-frequency detail features onto low-frequency macroscopic features based on the dynamic weight coefficients to obtain the predicted three-dimensional temperature field.

[0070] Specifically, the dynamic temperature field prediction model first maps the multidimensional boundary conditions of the battery pack into high-dimensional latent feature vectors. Then, the Fourier Neural Operator (FNO) solves for the low-frequency macroscopic features representing the macroscopic temperature rise trend in the frequency domain. Based on this, the spatial high-frequency compensation branch is responsible for generating high-frequency detail features such as local hot spots, while the weight network dynamically determines the fusion ratio of detail features according to the severity of the current operating conditions. Finally, the spatial gated fusion module superimposes the high-frequency detail features onto the low-frequency macroscopic features, outputting the predicted three-dimensional temperature field.

[0071] As a preferred technical solution, step S1 specifically includes:

[0072] S11 processes the raw data.

[0073] The original input contains time-series and scalar data of different physical dimensions (such as real-time current, voltage, inlet velocity vector, ambient temperature, etc.), and uses these time-series and scalar data of different physical dimensions as multi-dimensional boundary conditions for the battery pack. The multimodal feature encoder is responsible for projecting and aligning these battery pack physical boundary conditions into a unified high-dimensional latent feature space, thus completing the coupling of multi-physics fields. First, the boundary conditions are divided into three categories.

[0074] Electrical state Includes real-time current. ,Voltage State of charge (SOC).

[0075] Fluid boundary conditions Coolant flow velocity vector containing multiple inlets .

[0076] Environmental and static scalar Includes ambient temperature Initial temperature .

[0077] For the three types of data, the multimodal feature encoder uses a multilayer perceptron (MLP) to extract high-dimensional features independently, and outputs high-dimensional latent feature vectors of the same dimension:

[0078] Electrical characteristic projection:

[0079] Fluid characteristic projection:

[0080] Environmental feature projection:

[0081] in, Latent feature vectors representing different types of data. and Represents the weights and biases of different types of data. This represents the activation function of the Gaussian Error Linear Unit (GELU).

[0082] After obtaining the latent feature vectors of different types of data in the same dimension, the independent physical features are first concatenated along the feature channel dimension to construct a global multi-physical feature matrix. : .

[0083] S12, Fourier neural operator computation.

[0084] The high-dimensional latent feature vectors, after independent encoding and alignment, are concatenated and stitched along the channel dimension to obtain a global multi-physical feature matrix. Subsequently, this global multi-physical feature matrix is ​​broadcast along the 3D spatial grid of the battery module and coupled to the 3D spatial coordinate system. Together, they are spliced ​​to construct the initial input field.

[0085] Let the three-dimensional mesh be For any point on the grid, we have Enter the initial input field of the Fourier neural operator FNO. for:

[0086] A shallow linear mapping is performed on the input field in the first layer of Fourier operators. The number of channels is increased to a wider hidden layer dimension, ensuring sufficient information is contained in the subsequent frequency domain. The initial hidden field after dimensionality increase. for: .

[0087] The initial hidden field after dimensional ascension and implicit fusion This allows for iteration using the core component of the Fourier neural operator, the Fourier layer. In the... In each Fourier layer, the data is divided into two paths: one path enters the frequency domain for global convolution, and the other path performs residual linear mapping in the spatial domain. Finally, the two paths are added together and activated. .

[0088] in: and These represent the Fast Fourier Transform and its inverse transform, respectively. A learnable complex weight matrix representing the truncated low-frequency mode in the frequency domain; Residual connectivity weights representing the spatial domain; This represents the activation function of the Gaussian Error Linear Unit (GELU).

[0089] After four layers of Fourier iteration, the low-frequency macroscopic characteristic map of the temperature field can be obtained.

[0090] S13, high-frequency compensation branch.

[0091] The low-frequency macroscopic feature map is directly used as the input clue for the spatial high-frequency compensation branch. This spatial high-frequency compensation branch abandons frequency domain transformation and operates entirely in the original three-dimensional physical space domain. It uses a deep three-dimensional convolutional neural network, U-Net, to perform local feature extraction and nonlinear mapping, specifically for reconstructing lost microscopic extreme features.

[0092] High-frequency compensation feature map derived from branch path for: .

[0093] in, This represents the low-frequency macroscopic feature map output after 4 layers of Fourier iteration in step S12; The deep spatial mapping network function representing the high-frequency compensation branch; Internally, it incorporates multi-level three-dimensional local receptive field convolution (Conv3D) and nonlinear activation operations. Through the above mapping, the network can autonomously learn the nonlinear spatial geometric relationship between macroscopic temperature gradients and local thermal accumulation, accurately generating high-frequency detail features that reflect local thermodynamic abrupt changes.

[0094] S14, the weighted network evaluates and outputs the final predicted three-dimensional temperature field.

[0095] Weighted networks As input, the overall severity of the current transient condition is extracted through global average pooling, and then evaluated and output as a scalar dynamic weight coefficient through nonlinear mapping. : .

[0096] in, This represents the multilayer perceptron mapping function within the weighted network; The activation function will output the weight coefficients. Constrained within the (0,1) interval, this coefficient physically reflects the strength of the current operating condition's dependence on local high-frequency detail compensation.

[0097] Finally, the system utilizes a spatially gated fusion module to superimpose high-frequency features onto the low-frequency macroscopic feature map using dynamic weighting coefficients, and then restores it to physical temperature values ​​through a final decoding projection layer. The predicted final three-dimensional temperature field... : .

[0098] in, This represents the element-wise multiplication operation of the characteristic matrix; and This represents the learnable weights and bias matrix of the final linear decoding layer.

[0099] Through the aforementioned adaptive gating mechanism, when the battery is in a stable low-current operating condition, the dynamic weighting coefficient... The temperature field automatically approaches zero, and is dominated by Fourier neural operators to maintain macroscopic smoothness and extremely fast inference; when under extreme conditions such as high-rate heat generation, the dynamic weighting coefficients... By adaptively increasing the frequency of the feature superposition, the system dynamically activates the high-frequency feature superposition, thereby achieving a high-fidelity temperature field reconstruction that balances global timeliness and local extremum accuracy within the algorithm space.

[0100] S2, train the dynamic temperature field prediction model, obtain the trained dynamic temperature field prediction model, and perform temperature field control based on the trained dynamic temperature field prediction model.

[0101] In summary, the proposed temperature field control method for lithium batteries under dynamic operating conditions constructs a dynamic temperature field prediction model that includes a multimodal feature encoder, a Fourier neural operator, a spatial high-frequency compensation branch, a weighted network, and a spatial gating fusion module. This method does not rely on external temperature sensors for delayed feedback. By utilizing the real-time multidimensional boundary conditions of the battery pack, it transforms the slow macroscopic physical heat transfer process into a rapid state mapping within a microsecond-level algorithm space. This reduces the risk of control overshoot and thermal runaway caused by thermal response delays, avoiding the problems of high computation time and response delays in existing temperature field prediction algorithms. Furthermore, the model, once solidified, has a low online computation dimension, enabling millisecond-level real-time extrapolation. This solves the problem that existing coupled calculations of flow field and temperature field are extremely time-consuming and cannot meet the timeliness requirements of millisecond-level online prediction and active feedback control under dynamic operating conditions.

[0102] In one embodiment, such as Figure 2 As shown, the specific methods for training the dynamic temperature field prediction model include:

[0103] S21. The Latin hypercube sampling method is used to perform hierarchical orthogonal sampling within the set physical threshold range to generate N sets of multidimensional boundary condition combinations. Among them, the N sets of multidimensional boundary condition combinations include the battery multidimensional dynamic electrical state sequence and the coolant targeted flow velocity allocation vector of the corresponding operating conditions.

[0104] This step mainly involves constructing a training dataset for multiple operating conditions based on the Latin hypercube sampling method.

[0105] like Figure 7As shown, to enable the dynamic temperature field prediction model to achieve high-fidelity generalization under complex operating conditions across the entire domain, a training dataset covering multiple operating states of the battery must first be constructed. Considering the extremely high dimensionality of the parameter space formed by the battery's heat generation power and the flow rate of the multi-inlet coolant, this system employs Latin Hypercube Sampling (LHS) to perform hierarchical orthogonal sampling within a set physical threshold range to generate... Combination of multidimensional boundary conditions : .

[0106] in, Representing the Multidimensional dynamic electrical state sequence of batteries in group sampling; This represents the corresponding working condition including Coolant targeted flow rate allocation vector for each inlet.

[0107] S22, input N sets of multidimensional boundary conditions into the computational fluid dynamics solver to perform CFD numerical simulation and obtain the true three-dimensional temperature field on the spatial grid nodes.

[0108] The generated A combination of multidimensional boundary conditions is batch-input into the computational fluid dynamics (CFD) solver. During this process, the solver, through the attached equivalent circuit model (ECM) and user-defined functions (UDFs), processes the input multidimensional dynamic electrical state sequence of the battery. The process involves real-time decoupling and transformation into a transient volumetric heat generation power source within the battery cell. The computational domain is defined by a pre-established three-dimensional geometric mesh of a square battery module and liquid cooling channels. The transient solution of the Navier-Stokes equations and energy conservation equations for incompressible fluids, involving momentum conservation, is performed at the microscopic physical scale. Through high-precision numerical iteration, the true three-dimensional temperature field distribution on the spatial grid nodes is extracted and defined as the true value label. : .

[0109] in, This represents the numerical calculation process of fluid-thermal coupling based on the finite volume method; Represents the coordinates of a grid node in three-dimensional physical space.

[0110] S23, input the battery's multidimensional dynamic electrical state sequence and the coolant targeted flow velocity allocation vector into the dynamic temperature field prediction model to obtain the predicted three-dimensional temperature field.

[0111] The product generated in step S21 Combination of multidimensional boundary conditions Multidimensional dynamic electrical state sequence of battery and coolant targeted flow rate distribution vector The data is input into the dynamic temperature field prediction model for calculation, and the result is: as well as .

[0112] in, Represents the initial three-dimensional input field; This represents the currently learnable weight parameters. Predictive network model; This represents the prediction result of the current model, i.e., the predicted three-dimensional temperature field, which is compared with the true three-dimensional temperature field in step S21. They are jointly transmitted to the subsequent physical embedded joint optimization module for error comparison and parameter optimization.

[0113] S24, compare the error and optimize the parameters based on the true three-dimensional temperature field and the predicted three-dimensional temperature field.

[0114] As a preferred technical solution, such as Figure 3 As shown, the specific methods for error comparison and parameter optimization include:

[0115] S241, construct a physical embedded joint loss function that includes data fidelity loss, physical equation residual loss and high-frequency regularization loss.

[0116] S242 calculates the joint loss for each iteration cycle in real time based on the physical embedded joint loss function, and makes a state determination based on the convergence threshold and the joint loss.

[0117] S243. If the joint loss is not less than the convergence threshold, the Adam optimizer is started, and the gradient descent algorithm is used to backpropagate the joint error, dynamically adjust and optimize the weight parameters inside the dynamic temperature field prediction model.

[0118] S244, If the joint loss is less than the convergence threshold, then solidify the dynamic temperature field prediction model.

[0119] Specifically, to ensure that the dynamic temperature field prediction model not only numerically fits the data representation but also strictly adheres to the underlying thermodynamic laws in the algorithm space, this embodiment constructs a physically embedded joint loss function in the error comparison stage. : .

[0120] in, Represents the data fidelity loss, used to calculate the mean square error (MSE) between the predicted field and the true field on the global 3D grid nodes, guiding the model to approximate the actual working condition distribution; This represents the residual loss in the physical equations, calculated by substituting the predicted results into the partial differential equations for energy conservation and heat transfer in incompressible fluids. By penalizing the physical residuals, the update trajectory of the network weights is forced to adhere to the constraints of the first law of thermodynamics. Represents high-frequency regularization loss, used to suppress local numerical distortions that may be caused by spatial compensation branches that do not conform to physical norms, and to ensure the spatial smoothness of the overall temperature field. This represents the adaptive penalty weighting coefficient that balances the gradients of various errors.

[0121] To strengthen physical constraints and ensure that the predicted three-dimensional temperature field satisfies the first law of thermodynamics, a thermodynamic consistency loss can also be introduced into the physical embedded joint loss function. In other words, the physical embedded joint loss function is represented as a weighted sum of thermodynamic consistency loss, data fidelity loss, physical equation residual loss, and high-frequency regularization loss.

[0122] in, The Laplace operator, cell heat generation rate, and thermal conductivity are represented respectively, with N representing the total number of spatial grid nodes. Specifically, the Laplace operator represents the second spatial derivative of the temperature field, used to simulate the heat transfer process from the high-temperature region to the low-temperature region; This represents the predicted temperature value of the model at grid point i in the spatial grid. Let be the square of the L2 norm, representing the sum of squares of the errors; The main task is to calculate the physical residuals of the predicted three-dimensional temperature field.

[0123] This thermodynamic consistency loss strengthens physical consistency by calculating the difference between the predicted temperature field and the theoretical value of the heat conduction equation, ensuring that the predicted temperature distribution conforms to the first law of thermodynamics. As a regularization term during training, it can prevent the model from overfitting to data noise and improve generalization ability. Especially under dynamic conditions, it can help the model capture the physical nature of heat conduction and reduce prediction errors.

[0124] Furthermore, in order to consider the effect of fluid inertia on the hysteresis of the temperature field, a fluid continuity loss is obtained to characterize the coupling relationship between fluid mechanics and heat conduction. The weighted sum of fluid continuity loss, thermodynamic consistency loss, data fidelity loss, physical equation residual loss, and high-frequency regularization loss is used as the physical embedded joint loss function.

[0125] For example, fluid continuity loss .in, The vector fields are represented in sequence as the velocity vector field of the j-th coolant inlet, the gradient of the predicted three-dimensional temperature field, the thermal diffusivity, and the Laplace operator of the predicted temperature field. These represent the weighting coefficients of the divergence constraint and the divergence of the velocity field, respectively.

[0126] Specifically, the divergence of the velocity field is used for the fluid source / sink strength, and the weighting coefficient of the divergence constraint is a hyperparameter, typically between 0.1 and 0.5. The dot product of coolant flow rate and temperature gradient reflects the energy transport caused by fluid motion. It mainly represents the effect of thermal diffusion, which conforms to Fourier's law; The term can be understood as a simplified form of forcing the predicted temperature field to satisfy the energy conservation equation. It is mainly used to constrain the model prediction results to conform to physical laws and avoid temperature distributions that violate thermodynamics. The term can be understood as forcing the fluid field to satisfy the law of conservation of mass. The squared L2 norm is mainly used to penalize regions with non-zero divergence. A divergence of 0 in the velocity field indicates that the fluid is incompressible and has no source / sink (ideal conditions). This item is mainly used to prevent the model from generating flow velocity distributions that violate mass conservation, such as coolant disappearing or appearing out of thin air.

[0127] In summary, incorporating fluid continuity loss into the physical embedded joint loss function allows the predicted temperature field to simultaneously satisfy convection-diffusion balance and fluid continuity, avoiding the hysteresis effect of fluid inertia on the temperature field, further optimizing system control delay, and improving cooling response speed.

[0128] like Figure 8 As shown, in the parameter optimization stage, the joint loss for each iteration cycle is calculated in real time, and a convergence threshold is introduced. Implement a state determination mechanism. If... This indicates that the model has not yet fully grasped the nonlinear mapping law of electrothermal flow. The system then activates the Adam optimizer, using the gradient descent algorithm to backpropagate the joint error, dynamically adjusting and optimizing the weight parameters within the prediction network. When the convergence limit is reached, the dynamic temperature field prediction model is determined to have reached its convergence limit and possesses high-precision cross-modal physical evolution inference capability. At this point, the system immediately cuts off the error backpropagation path and globally solidifies all operator weights and spatial convolution kernel parameters within the prediction network to obtain the trained dynamic temperature field prediction model.

[0129] In one embodiment, such as Figure 4 As shown, the specific methods for temperature field control include:

[0130] S25, obtain the current multidimensional boundary conditions of the lithium battery and the multi-inlet flow rate allocation scheme currently executed by the cooling mechanism.

[0131] S26. Input the current multidimensional boundary conditions and multi-inlet velocity distribution scheme into the trained dynamic temperature field prediction model to obtain the predicted three-dimensional temperature field under the current multi-inlet velocity distribution scheme.

[0132] S27, obtain the degree of deviation between the predicted three-dimensional temperature field and the target temperature field, and control the temperature field according to the degree of deviation.

[0133] Specifically, the current multidimensional boundary condition electrical state is first collected in real time by sensors. And the multi-inlet flow rate distribution scheme currently being implemented by the cooling mechanism. Subsequently, the collected actual electrical and fluid physical states are used as baseline inputs and transmitted to the fixed dynamic temperature field prediction model. Due to the complete freezing of the model's internal parameters, the system can rapidly output the predicted three-dimensional temperature field under the current multi-inlet velocity distribution scheme within milliseconds. : ;in, A dynamic temperature field prediction model representing solidified parameters.

[0134] After the prediction is completed, the results are transmitted to the optimization module. The optimization module has a preset target temperature field characterizing the battery's optimal safety and lifespan. Maximum temperature limit Minimum temperature limit Identical boundary conditions. The optimization module constructs a cost evaluation function. Calculate the degree of deviation between the current predicted baseline field and the target safe field: .

[0135] By evaluating the function This assessment evaluates whether the current cooling strategy is reasonable and efficient, providing a basis for decision-making regarding whether further optimization and improvement are needed.

[0136] As a preferred technical solution, such as Figure 5 As shown, specific methods for temperature field control based on the degree of deviation include:

[0137] S271, if the deviation is less than the over-limit threshold, maintain the original instruction.

[0138] S272. If the deviation is not less than the over-limit threshold, the improvement mechanism is activated. Using the particle swarm optimization (PSO) algorithm, a series of trial flow rate schemes are generated based on the current actual flow rate allocation scheme.

[0139] S273 feeds back the trial flow velocity scheme to the dynamic temperature field prediction model for a new round of rapid prediction. Through continuous optimization and iteration, the optimal flow velocity scheme for multiple inlets is obtained.

[0140] Specifically, obtain the baseline cost evaluation function value. Then, the system triggers the control branch decision logic.

[0141] If the cost function value meets the preset safety and energy efficiency convergence criteria (i.e., meets expectations), it indicates that the current actual flow rate scheme is sufficient to cope with the current cell heating, and the system maintains the original instructions without additional intervention. If the cost function value is not less than the over-limit threshold, the improvement mechanism of the optimization module is immediately activated. At this time, the optimization module uses the particle swarm optimization algorithm (PSO) to allocate the scheme according to the current actual flow rate. Based on this, a series of trial flow rate schemes are generated. .

[0142] These trial variables are frequently fed back into the dynamic temperature field prediction model for a new round of rapid predictions: .

[0143] The optimization algorithm is based on the temperature field cost function value corresponding to each trial flow velocity. Continuous iterative evolution. Thanks to the extremely high computational efficiency of the solidified surrogate model, the system can complete hundreds of optimization cycles within the brief control time window before a substantial temperature rise occurs due to physical thermodynamics. The optimization iteration terminates when a set of trial schemes minimizes the cost function or meets the stringent expected objective. The optimization module outputs the improved multi-inlet optimal flow rate scheme. Ultimately, this optimal solution is used as a feedforward improvement instruction and sent to the underlying multi-inlet flow rate control mechanism. The actuator dynamically adjusts the physical opening and pumping power of each fluid inlet to achieve targeted and precise cooling of the actual battery module.

[0144] In this embodiment, the present invention achieves real-time adaptive matching of heat dissipation strategies by combining a temperature field prediction algorithm with a multi-objective optimization control module. Based on transient heat generation characteristics, this method dynamically deduces and executes a globally optimal coolant flow field distribution scheme, thereby improving the system's convective heat transfer efficiency. This not only reduces drastic temperature fluctuations in the battery pack under complex dynamic conditions but also improves the spatial uniformity of the three-dimensional temperature field, thus controlling the temperature field within the optimal operating range of the battery, delaying battery thermal aging, and extending the battery pack's lifespan.

[0145] An embodiment of the present invention also provides a temperature field control system for lithium batteries under dynamic operating conditions, used to implement the aforementioned temperature field control method, which includes a dynamic temperature field prediction model, such as... Figure 6 As shown, the dynamic temperature field prediction model includes a multimodal feature encoder, a Fourier neural operator, a spatial high-frequency compensation branch, a weight network, and a spatial gating fusion module.

[0146] The multimodal feature encoder is used to extract high-dimensional features independently from the multidimensional boundary conditions of the battery pack, obtain latent feature vectors of the same dimension, and concatenate the independent latent feature vectors in the feature channel dimension to construct a global multi-physical feature matrix.

[0147] Fourier neural operators are used to process the global multi-physical feature matrix to obtain low-frequency macroscopic features that characterize the macroscopic temperature rise trend; spatial high-frequency compensation branches are used to extract local features and perform nonlinear mapping on low-frequency macroscopic features to obtain high-frequency detail features that reflect local thermodynamic abrupt changes; weighted networks are used to extract the overall severity of the current working condition based on low-frequency macroscopic features to obtain dynamic weight coefficients that determine the fusion ratio of detail adjustments.

[0148] The spatial gating fusion module is used to superimpose high-frequency detail features onto low-frequency macroscopic features based on dynamic weighting coefficients to obtain the predicted three-dimensional temperature field.

[0149] As a preferred technical solution, the temperature field control system also includes a multi-dimensional boundary condition combination acquisition module, a computational fluid dynamics solver, and a physics-embedded joint optimization module.

[0150] The multidimensional boundary condition combination acquisition module uses the Latin hypercube sampling method to perform hierarchical orthogonal sampling within a set physical threshold range to generate N sets of multidimensional boundary condition combinations. Among them, the N sets of multidimensional boundary condition combinations include the battery multidimensional dynamic electrical state sequence and the coolant targeted flow velocity allocation vector of the corresponding operating conditions.

[0151] The computational fluid dynamics solver is used to perform CFD numerical simulations based on N sets of multidimensional boundary conditions to obtain the true three-dimensional temperature field on the spatial grid nodes; the physics-embedded joint optimization module is used to compare errors and optimize parameters based on the true three-dimensional temperature field and the predicted three-dimensional temperature field.

[0152] The physical embedded joint optimization module includes a loss function construction unit and a state determination unit.

[0153] The loss function construction unit is used to construct a physical embedded joint loss function that includes data fidelity loss, physical equation residual loss, and high-frequency regularization loss. The state determination unit is used to determine the state based on the convergence threshold and the joint loss. If the joint loss is not less than the convergence threshold, the Adam optimizer is started, and the gradient descent algorithm is used to backpropagate the joint error to dynamically adjust and optimize the weight parameters inside the dynamic temperature field prediction model. Otherwise, the dynamic temperature field prediction model is fixed to obtain a trained dynamic temperature field prediction model.

[0154] The temperature field control system also includes a battery pack current parameter acquisition module and an optimization module.

[0155] The battery pack current parameter acquisition module is used to acquire the current multidimensional boundary conditions of the lithium battery and the multi-inlet flow velocity allocation scheme currently executed by the cooling mechanism. It then inputs the current multidimensional boundary conditions and multi-inlet flow velocity allocation scheme into the trained dynamic temperature field prediction model to obtain the predicted three-dimensional temperature field under the current multi-inlet flow velocity allocation scheme. The optimization module is used to acquire the degree of deviation between the predicted three-dimensional temperature field and the target temperature field, and obtain the optimal multi-inlet flow velocity scheme based on the degree of deviation.

[0156] The optimization module includes a trial flow rate scheme generation unit and an optimal flow rate scheme acquisition unit.

[0157] The trial flow rate scheme generation unit is used to activate the improvement mechanism when the deviation is not less than the over-limit threshold. Using the particle swarm optimization algorithm (PSO), a series of trial flow rate schemes are generated based on the current real flow rate allocation scheme. The optimal flow rate scheme acquisition unit is used to feed back the trial flow rate schemes to the dynamic temperature field prediction model for a new round of rapid prediction. Through continuous optimization and iteration, the optimal flow rate scheme for multiple inlets is obtained.

[0158] In summary, the temperature field control system for lithium batteries under dynamic operating conditions described in this invention employs a physically embedded surrogate model, avoiding the high computational time of existing temperature field prediction algorithms. Furthermore, the model, once solidified, has a low online computational dimensionality, enabling millisecond-level real-time simulation. In addition, the condition-triggered feedforward optimization avoids the thermal response delay of traditional passive feedback, allowing for timely adjustment of the optimal cooling flow rate based on real-time prediction results, achieving zero-hysteresis active heat dissipation and effectively improving system thermal safety.

[0159] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A method for temperature field control under dynamic operating conditions of lithium batteries, characterized in that, include: A dynamic temperature field prediction model is constructed, comprising a multimodal feature encoder, a Fourier neural operator, a spatial high-frequency compensation branch, a weight network, and a spatial gated fusion module. The multimodal feature encoder extracts independent high-dimensional features from the multidimensional boundary conditions of the battery pack, obtaining latent feature vectors of the same dimension. These latent feature vectors are then concatenated along the feature channel dimension to construct a global multi-physical feature matrix. The Fourier neural operator processes this global multi-physical feature matrix to obtain low-frequency macroscopic features characterizing the macroscopic temperature rise trend. The spatial high-frequency compensation branch extracts local features and performs nonlinear mapping on the low-frequency macroscopic features, obtaining high-frequency detail features reflecting local thermodynamic abrupt changes. The weight network extracts the overall severity of the current operating condition based on the low-frequency macroscopic features, obtaining dynamic weight coefficients to determine the fusion ratio for detail adjustment. The spatial gated fusion module superimposes high-frequency detail features onto low-frequency macroscopic features based on the dynamic weight coefficients to obtain the predicted three-dimensional temperature field. The dynamic temperature field prediction model is trained to obtain the trained dynamic temperature field prediction model, and temperature field control is performed based on the trained dynamic temperature field prediction model.

2. The temperature field control method for lithium battery dynamic operating conditions as described in claim 1, characterized in that, Specific methods for training dynamic temperature field prediction models include: The Latin hypercube sampling method is used to perform hierarchical orthogonal sampling within a set physical threshold range to generate N sets of multidimensional boundary condition combinations. Among them, the N sets of multidimensional boundary condition combinations include the battery multidimensional dynamic electrical state sequence and the coolant targeted flow velocity allocation vector of the multi-inlet under the corresponding working conditions. The N sets of multidimensional boundary conditions are input into the computational fluid dynamics solver to perform CFD numerical simulation and obtain the true three-dimensional temperature field on the spatial grid nodes. The battery multidimensional dynamic electrical state sequence and the coolant targeted flow velocity allocation vector are input into the dynamic temperature field prediction model to obtain the predicted three-dimensional temperature field. Error comparison and parameter optimization are performed based on the true three-dimensional temperature field and the predicted three-dimensional temperature field.

3. The temperature field control method for lithium battery dynamic operating conditions as described in claim 2, characterized in that, Specific methods for error comparison and parameter optimization include: Construct a physical embedded joint loss function that includes data fidelity loss, physical equation residual loss, and high-frequency regularization loss; Based on the physical embedded joint loss function, the joint loss for each iteration cycle is calculated in real time, and the state is determined based on the convergence threshold and the joint loss. If the joint loss is not less than the convergence threshold, the Adam optimizer is started, and the gradient descent algorithm is used to backpropagate the joint error, dynamically adjust and optimize the weight parameters inside the dynamic temperature field prediction model. If the joint loss is less than the convergence threshold, the dynamic temperature field prediction model is solidified.

4. The temperature field control method for lithium battery dynamic operating conditions as described in claim 3, characterized in that, Specific methods for temperature field control include: Obtain the current multidimensional boundary conditions of the lithium battery and the multi-inlet flow rate allocation scheme currently being executed by the cooling mechanism; Input the current multidimensional boundary conditions and multi-inlet velocity allocation scheme into the trained dynamic temperature field prediction model to obtain the predicted three-dimensional temperature field under the current multi-inlet velocity allocation scheme. The deviation between the predicted three-dimensional temperature field and the target temperature field is obtained, and the temperature field is controlled based on the deviation.

5. The temperature field control method for lithium battery dynamic operating conditions as described in claim 4, characterized in that, Specific methods for temperature field control based on the degree of deviation include: If the deviation is less than the over-limit threshold, maintain the original instruction; If the deviation is not less than the threshold, the improvement mechanism is activated, and the particle swarm optimization (PSO) algorithm is used to generate a series of trial flow rate schemes based on the current real flow rate allocation scheme. The trial flow velocity scheme is fed back into the dynamic temperature field prediction model for a new round of rapid prediction. Through continuous optimization and iteration, the optimal flow velocity scheme for multiple inlets is obtained.

6. A temperature field control system for lithium batteries under dynamic operating conditions, used to implement the temperature field control method as described in any one of claims 1-5, characterized in that, This includes a dynamic temperature field prediction model, which includes: The multimodal feature encoder is used to extract high-dimensional features independently from the multidimensional boundary conditions of the battery pack, obtain latent feature vectors of the same dimension, and concatenate the independent latent feature vectors in the feature channel dimension to construct a global multi-physical feature matrix. Fourier neural operators are used to process the global multi-physics feature matrix to obtain low-frequency macroscopic features that characterize the macroscopic temperature rise trend. The spatial high-frequency compensation branch is used to extract local features and perform nonlinear mapping on low-frequency macroscopic features to obtain high-frequency detailed features that reflect local thermodynamic abrupt changes. A weighted network is used to extract the overall severity of the current working condition based on low-frequency macroscopic features, and to obtain dynamic weight coefficients for determining the fusion ratio of detailed adjustments. The spatial gating fusion module is used to superimpose high-frequency detail features onto low-frequency macroscopic features based on dynamic weighting coefficients to obtain the predicted three-dimensional temperature field.

7. A temperature field control system for dynamic operating conditions of lithium batteries as described in claim 6, characterized in that, The temperature field control system also includes: The multidimensional boundary condition combination acquisition module uses the Latin hypercube sampling method to perform hierarchical orthogonal sampling within a set physical threshold range to generate N sets of multidimensional boundary condition combinations. Among them, the N sets of multidimensional boundary condition combinations include the battery multidimensional dynamic electrical state sequence and the coolant targeted flow velocity allocation vector of the multi-inlet under the corresponding working conditions. Computational fluid dynamics solver, used to perform CFD numerical simulation based on N sets of multidimensional boundary conditions, and obtain the true three-dimensional temperature field on the spatial grid nodes; The physical embedded joint optimization module is used to compare errors and optimize parameters based on the true three-dimensional temperature field and the predicted three-dimensional temperature field.

8. A temperature field control system for dynamic operating conditions of lithium batteries as described in claim 7, characterized in that, The physical embedded joint optimization module includes: The loss function construction unit is used to construct a physical embedded joint loss function that includes data fidelity loss, physical equation residual loss, and high-frequency regularization loss. The state determination unit is used to determine the state based on the convergence threshold and the joint loss. If the joint loss is not less than the convergence threshold, the Adam optimizer is started, and the gradient descent algorithm is used to backpropagate the joint error to dynamically adjust and optimize the weight parameters inside the dynamic temperature field prediction model. Otherwise, the dynamic temperature field prediction model is fixed to obtain a trained dynamic temperature field prediction model.

9. A temperature field control system for lithium batteries under dynamic operating conditions as described in claim 8, characterized in that, The temperature field control system also includes: The battery pack current parameter acquisition module is used to acquire the current multidimensional boundary conditions of the lithium battery and the multi-inlet flow velocity allocation scheme executed by the cooling mechanism at the current moment, and input the current multidimensional boundary conditions and multi-inlet flow velocity allocation scheme into the trained dynamic temperature field prediction model to obtain the predicted three-dimensional temperature field under the current multi-inlet flow velocity allocation scheme. The optimization module is used to obtain the degree of deviation between the predicted three-dimensional temperature field and the target temperature field, and to obtain the optimal flow velocity scheme for multiple inlets based on the degree of deviation.

10. A temperature field control system for dynamic operating conditions of lithium batteries as described in claim 9, characterized in that, The optimization module includes: The trial flow rate scheme generation unit is used to activate the improvement mechanism when the deviation is not less than the over-limit threshold. It uses the particle swarm optimization algorithm (PSO) to generate a series of trial flow rate schemes based on the current real flow rate allocation scheme. The optimal flow rate scheme acquisition unit is used to feed back the trial flow rate scheme to the dynamic temperature field prediction model for a new round of rapid prediction. Through continuous optimization and iteration, the optimal flow rate scheme for multiple inlets is obtained.

Images