Battery temperature prediction method and system based on Shenchang differential equation

Abstract

The invention provides a battery temperature prediction method and system based on a Shenchang differential equation. The method comprises the steps of obtaining multi-source operation data in the operation process of a battery system; performing time interpolation processing on the data, and uniformly mapping the data of different sampling frequencies to a continuous time dimension to obtain a continuous time input feature; respectively constructing global features and discrimination features based on the input features, wherein the discrimination features are used for representing key thermal influence factors such as battery calorific value and cooling medium flow; fusing the global features, the discriminant features and the current temperature state of the battery to construct comprehensive features; and inputting the comprehensive characteristics into a thermodynamic model constructed based on a Shenchang differential equation, and solving a continuous evolution process of the battery temperature through numerical integration to realize temperature prediction one by one. According to the method, the thermodynamic continuous evolution law can be followed, and the space coupling characteristic and the multi-working-condition robustness can be considered.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery temperature prediction, and particularly relates to a battery temperature prediction method and system based on neural ordinary differential equation. BACKGROUND

[0002] The battery system is accompanied by complex thermal physical evolution in the charging and discharging process, and there is strong spatial thermal coupling characteristics between single cells. Accurate temperature prediction is the cornerstone for the battery management system to realize thermal runaway warning and life management. At present, battery thermal modeling mainly has two paths: One is an analytical or numerical model based on physical equations, such as finite element, lumped parameter model, etc. Such a model has strong physical interpretability, but is highly dependent on the accurate calibration of key parameters such as thermal capacity and thermal resistance. Under the complex and variable working conditions in actual operation, the nonlinear drift of physical parameters makes it difficult for the model to maintain high accuracy for a long time, and the large-scale consumption of computing resources limits its online application.

[0003] The second is a statistical or machine learning model based on data-driven, such as LSTM, CNN, etc. Such a model is good at capturing nonlinear characteristics, but it is essentially a black box mapping of time steps, which lacks the description of the continuous evolution process of thermodynamics. When facing input signal mutation or non-uniform sampling, the robustness and time continuity of the model are poor.

[0004] In recent years, the emergence of neural ordinary differential equation (Neural ODE) provides a new idea for continuous-time dynamic modeling. Neural ODE fits the vector field through a neural network, and converts the hierarchical structure of deep learning into a continuous time evolution process, which is more consistent with the physical nature of battery thermal dynamics in theory. However, when directly applying Neural ODE to the modeling of dynamic battery systems, the following problems still exist: High-dimensional state representation difficulty: the high dimension of the thermal state space of large-scale battery arrays results in heavy computational burden of ordinary differential equation system solving; Lack of spatiotemporal coupling mechanism: existing models are difficult to balance the continuous evolution of time dimension and the complex thermal conduction coupling of space dimension; Insufficient generalization and stability: under dynamic current, variable environment and other wide working conditions, the model is prone to time interpolation instability and prediction drift phenomenon.

[0005] Therefore, there is an urgent need for a battery temperature prediction method and system based on neural ordinary differential equation, which can not only follow the continuous evolution law of thermodynamics, but also consider the spatial coupling characteristics and multi-condition robustness. SUMMARY

[0006] In view of the above prior art defects, the purpose of the present application is to provide a battery temperature prediction method and system based on neural ordinary differential equation, aiming to solve the technical problems of complex high-dimensional state solving, inaccurate spatiotemporal coupling description and poor generalization in multiple working conditions when Neural ODE is applied to battery thermal modeling.

[0007] To achieve the above purpose, in a first aspect, the present application provides a battery temperature prediction method based on neural ordinary differential equation, comprising the following steps: S1, obtaining multi-source operation data in the operation process of the battery system; S2, performing time interpolation processing on the multi-source operation data, and uniformly mapping data with different sampling frequencies and unequal time intervals to a continuous time dimension to obtain continuous time input features; S3, based on the continuous time input features, respectively constructing global features and discriminative features, wherein the discriminative features are used to at least represent thermal influencing factors corresponding to battery heat generation and cooling medium flow; S4, fusing the global features, discriminative features and current temperature state of the battery to construct comprehensive features for describing the thermal dynamics state of the battery; S5, inputting the comprehensive features into a thermal dynamics model constructed based on neural ordinary differential equation, solving the continuous evolution process of the battery temperature with time by numerical integration method to obtain the single-cell temperature prediction result of the battery system at the target time.

[0008] As a further improvement of the above scheme, the multi-source operation data at least includes sampling time, ambient temperature, battery single-cell heat generation, battery working current, cooling medium flow and cooling medium inlet temperature.

[0009] As a further improvement of the above scheme, the thermal dynamics model parameterizes the function of the change of the battery temperature state with time by a neural network to form a neural ordinary differential equation model.

[0010] As a further improvement of the above scheme, the time interpolation processing uses an interpolation algorithm to map the original discrete sampling data to a continuous time sequence to eliminate the influence of inconsistent sampling periods of different sensors on the modeling accuracy.

[0011] As a further improvement of the above scheme, the thermal dynamics model includes a heating model and a heat dissipation model; The heating model is used to describe the internal heating effect of the battery, and the heat dissipation model is used to describe the influence of external cooling conditions.

[0012] As a further improvement of the above scheme, the outputs of the heating model and the heat dissipation model are fused by a weight parameter to represent the temperature change trend under different operating conditions; Specifically, the temperature prediction formula is as follows: ; wherein, is a temperature prediction output, representing the predicted temperature distribution of all monomers in the battery pack at a certain time; is a dynamic weight parameter, ∈[0,1], which is automatically learned by a neural network; is a heating function, used to simulate the heating effect of the internal heat source of the battery, is a heat dissipation function, used to simulate the heat dissipation of the battery system to the external environment; is an input comprehensive feature.

[0013] As a further improvement of the above scheme, the neural ordinary differential equation model is solved by a numerical integration algorithm to obtain the continuous change result of the battery temperature with time; Preferably, the fourth-order Runge-Kutta method (RK4) is used for numerical integration of Neural ODE; the adjoint sensitivity method is used to calculate the gradient in the training stage to reduce the memory occupation and support long sequence modeling.

[0014] In a second aspect, the application further provides a battery temperature prediction system based on neural ordinary differential equations, which adopts the method provided in the first aspect, and the system comprises: a data acquisition module for acquiring multi-source operation data of the battery system; a data preprocessing module for time interpolation processing of the multi-source operation data to construct continuous time input features; a feature extraction module including a global feature extraction submodule and a discriminative feature extraction submodule for generating a feature vector for representing the thermal state of the battery; a thermal dynamics modeling module for constructing a battery thermal dynamics model based on neural ordinary differential equations and continuously predicting the battery temperature; an output module for outputting the temperature prediction results of each monomer of the battery system at a target time.

[0015] As a further improvement of the above scheme, the thermal dynamics modeling module includes a heating module, a heat dissipation module and a fusion module for modeling the internal heating effect and external cooling effect of the battery and performing weighted fusion.

[0016] As a further improvement of the above scheme, it further includes a model training module for pre-training the battery thermal dynamics model based on neural ordinary differential equations using historical data or simulation data.

[0017] Due to the above technical solutions, the present application has the following beneficial effects: The application provides a battery temperature prediction method based on a neural ordinary differential equation. Firstly, a thermal dynamic process is modeled in a data-driven manner to reduce the dependence on explicit modeling of thermal parameters. Specifically, the battery thermal dynamic process is parameterized and modeled by a neural ordinary differential equation, which realizes temperature prediction without relying on explicit modeling of physical parameters such as thermal capacity and thermal resistance, thereby reducing the complexity of model construction and improving the applicability in actual engineering scenarios. Secondly, the application is beneficial to improve the adaptability of the model under different operating conditions. Specifically, since the application adopts a continuous-time modeling method and combines time interpolation processing and feature separation and fusion structure, when the operating condition changes, the model can naturally reflect the thermal behavior characteristics under different conditions through input feature changes, thereby to some extent, alleviating the problem of insufficient generalization ability under multiple operating conditions in the prior art.

[0018] In addition, in step S2, time interpolation processing is performed on the multi-source operating data with different sampling frequencies and inconsistent time intervals, so that various input parameters participate in modeling under the same continuous time axis, thereby avoiding modeling bias introduced by asynchronous sampling time and improving the continuity and stability of the description of the battery temperature change process. Through time interpolation processing, data with different sampling frequencies are unified to the continuous time dimension, eliminating the cumulative error caused by inconsistent sampling in traditional discrete models. Combined with the continuous-time modeling characteristics of the neural ordinary differential equation, the solution of the high-dimensional temperature state is converted into a numerical integration problem of a continuous dynamic system, avoiding the calculation complexity of traditional methods for independent calculation at each time step.

[0019] In step S3, global features and discriminative features are respectively constructed, wherein the discriminative features are mainly modeled for parameters such as battery heat generation and cooling medium flow, which have a high correlation with thermal behavior, which helps to strengthen the description of key thermal factors under high-dimensional input conditions and avoid weakening important thermal information in the overall feature modeling process. By respectively constructing global features and discriminative features, the coupling relationship between multiple parameters is retained, and the special representation of key thermal parameters is strengthened. The feature fusion mechanism combines physical prior knowledge with data-driven modeling, effectively capturing the spatial thermal coupling relationship between battery monomers. In step S4, the global features, discriminative features and the current temperature state of the battery are jointly constructed into comprehensive features, so that the model can consider the changes in operating conditions and historical state information when performing temperature prediction, thereby helping to maintain the consistency of the temperature evolution process in time. In step S5, the neural ordinary differential equation is used to describe the change relationship of the battery temperature with time, and the temperature trajectory is solved by numerical integration. Compared with traditional discrete-time prediction models, the neural ordinary differential equation can directly obtain the temperature prediction results at any time point, and is more suitable for describing the continuous change characteristics of the battery thermal state under complex conditions. BRIEF DESCRIPTION OF DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only show some of the embodiments of the present application, and for those skilled in the art, other drawings can also be obtained from the structures shown in the drawings without creative labor.

[0021] Figure 1 A flowchart of a battery temperature prediction method based on a neural ordinary differential equation disclosed in Embodiment 1 of the present application. Figure 2 A partial structure diagram of a battery temperature prediction system based on a neural ordinary differential equation disclosed in Embodiment 2 of the present application.

[0022] The purposes, functional features and advantages of the present application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0023] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0024] It should be noted that the technical solutions of the various embodiments of the present application can be combined with each other, but it must be based on the fact that a person skilled in the art can realize it. When the combination of technical solutions appears contradictory or unachievable, it should be considered that the combination of technical solutions does not exist and is not within the scope of protection required by the present application.

[0025] Embodiment 1

[0026] Referring to Figure 1 The present application provides a battery temperature prediction method based on a neural ordinary differential equation, which is based on the temperature prediction demand of a battery system in actual operation process, and is used to solve the problem of unstable temperature prediction accuracy caused by inconsistent sampling, high-dimensional state complexity and insufficient time-space coupling description in the prior art.

[0027] Specifically, the steps of the battery temperature prediction method include: S1, obtaining multi-source operation data: First, obtain the multi-source operation data of the battery system in the operation process. The multi-source operation data includes but is not limited to sampling time, environment temperature, battery monomer heat generation, battery working current, cooling medium flow and cooling medium inlet temperature and other parameters.

[0028] By simultaneously collecting operation data reflecting the battery internal heat generation characteristics and external cooling conditions, the subsequent temperature prediction process can simultaneously consider the battery heat source changes and heat dissipation condition changes, thereby providing a relatively complete data basis for thermodynamic modeling.

[0029] S2, time interpolation processing of multi-source data: Since the above multi-source operation data often have different sampling frequencies and inconsistent sampling times in the actual collection process, directly using them for modeling can easily introduce time alignment errors. In this embodiment, time interpolation processing is performed on the obtained multi-source operation data to uniformly map data with different sampling frequencies and unequal time intervals to a continuous time dimension, thereby obtaining continuous time input features.

[0030] Through time interpolation processing, various operation parameters are involved in modeling at the same time scale, which is beneficial to reduce the modeling deviation caused by asynchronous sampling, thereby improving the continuity and stability of the battery temperature change process description.

[0031] S3, construction of global features and discriminative features: After completing the time interpolation processing, global features and discriminative features are constructed based on the continuous time input features. The global features are used to represent the overall change trend of the multi-source operation data in the continuous time, and the discriminative features are used to represent at least the thermal influence factors corresponding to the battery heat generation and the cooling medium flow.

[0032] Since the battery heat generation and the cooling medium flow have high correlation with the temperature change, targeted feature modeling of the above parameters helps to strengthen the description of key thermal influence factors under high-dimensional input conditions and avoid weakening important thermal information in the overall feature modeling process.

[0033] S4, construction of comprehensive features: In this embodiment, the constructed global features, discriminative features, and the current temperature state of the battery are fused to form comprehensive features for describing the thermodynamic state of the battery.

[0034] By introducing the current temperature state of the battery in the feature construction stage, the model can simultaneously consider the operation condition changes and historical temperature state information when predicting the subsequent temperature changes, which is beneficial to maintain the continuity and consistency of the battery temperature evolution process in the time dimension.

[0035] S5, temperature prediction based on neural ordinary differential equation: The comprehensive features are input into the thermodynamic model constructed based on the neural ordinary differential equation, and the continuous evolution process of the battery temperature with time is solved by numerical integration, thereby obtaining the single-cell temperature prediction results of the battery system at the target time.

[0036] Compared with the traditional discrete-time prediction model, the modeling method directly describes the continuous change relationship of the temperature state with respect to time through the neural ordinary differential equation, can realize the prediction of the battery temperature at any time point without relying on the modeling of the explicit heat capacity, thermal resistance and other physical parameters, and is more suitable for describing the continuous change characteristics of the battery thermal state under complex operating conditions.

[0037] As a preferred embodiment, the thermodynamic model parameterizes the change function of the battery temperature state with respect to time through a neural network, thereby constructing a neural ordinary differential equation model for describing the continuous evolution process of the battery temperature with time.

[0038] During the operation of the battery system, the change of the single battery temperature conforms to the law of conservation of energy, and the thermal balance relationship can be expressed as: ; wherein, denotes the temperature state of the battery at time , is the equivalent heat capacity of the battery, denotes the heat generated by the battery during operation, denotes the heat dissipated through convection, conduction and radiation and the like. In the traditional thermal model, the above-mentioned heat capacity, thermal resistance and other parameters usually need to be obtained through experimental calibration, and it is difficult to accurately determine them in a complex operating condition or a large-scale battery system.

[0039] To avoid explicit modeling of the above-mentioned physical parameters, the embodiment introduces a neural network to parameterize the change rate of the battery temperature. Specifically, the per-single battery temperature state of the battery system at time is expressed as a state vector: ; wherein, denotes the number of battery monomers. The input variables reflecting the operating conditions are expressed as: ; correspond to the sampling time, the ambient temperature, the heat generated by the battery monomer, the working current of the battery, the flow of the cooling medium and the inlet temperature of the cooling medium, respectively.

[0040] On this basis, the change function of the temperature state with respect to time is parameterized through a neural network, and a neural ordinary differential equation model of the following form is constructed: ; wherein, is a continuous-time dynamic function represented by a neural network, are learnable parameters of the network. In this way, the rate of change of the battery temperature is no longer dependent on explicitly given heat capacity or thermal resistance parameters, but is automatically learned by the neural network from historical operating data during the training process.

[0041] In some preferred embodiments, the neural network is further divided into a heating sub-network for describing the internal heating effect of the battery and a cooling sub-network for describing the influence of external cooling conditions, so that the above neural ordinary differential equation is represented as: ; wherein, is used to characterize the temperature change trend caused by the internal heating of the battery, is used to characterize the influence of cooling medium flow, inlet temperature, environmental conditions and other factors on temperature change; and are weight parameters learned by the network, used to adjust the relative contribution of heating effect and cooling effect to temperature change under different operating conditions.

[0042] In the model inference stage, the above neural ordinary differential equation is solved by a numerical integration method to obtain the evolution trajectory of the battery temperature in any time interval, and the prediction process can be represented as: ; By using the above modeling method based on neural ordinary differential equation, the embodiment can describe the dynamic change process of the battery temperature in the continuous time dimension, so that the model can still maintain good time consistency when facing unequal time sampling and operating condition changes. At the same time, by implicitly mapping the complex thermal mechanism to the neural network parameters, it is beneficial to reduce the dependence on explicit thermal parameter modeling and improve the engineering applicability of the model in large-scale battery systems.

[0043] As a preferred embodiment, the time interpolation process uses an interpolation algorithm to map the original discrete sampling data to a continuous time sequence to eliminate the influence of inconsistent sampling periods of different sensors on the modeling accuracy of thermal dynamics.

[0044] In the actual operation process of the battery system, different operating parameters are usually collected by different types of sensors, and the sampling period and sampling time of each sensor are often inconsistent. For example, the battery current, cooling medium flow and other parameters may be sampled at a high frequency, while the sampling frequency of the environmental temperature, cooling medium inlet temperature and other parameters is relatively low; at the same time, the sampling time points of various sensors may also have a time offset. If the above discrete sampling data is directly used as the model input, it is easy to introduce discontinuity or misalignment in the time dimension, thereby affecting the stability and consistency of temperature prediction.

[0045] To this end, time interpolation is performed on the multi-source discrete sampling data before subsequent feature construction and thermodynamic modeling. Specifically, let a certain operating parameter be at discrete sampling time points: ; The corresponding sampling values are: ; The discrete sequence is mapped to a continuous function defined in a continuous time interval by an interpolation algorithm .

[0046] In an embodiment, the interpolation process can adopt a linear interpolation method, which is expressed as: ; Through the above method, the estimated value of the operating parameter at any target time point can be obtained. For other operating parameters, the corresponding interpolation method is also used for processing, so as to uniformly map various types of original discrete sampling data to the same continuous time axis to form a continuous time input feature sequence. Through the above time interpolation process, the embodiment can effectively reduce the influence of inconsistent sampling periods of multi-source sensors on the modeling process, which is beneficial to improve the ability to maintain time continuity in the battery temperature prediction process, and further improve the stability and engineering applicability of the thermodynamic model under complex operating conditions.

[0047] As a preferred embodiment, the thermodynamic model further includes a heating model and a heat dissipation model, wherein the heating model is used to describe the internal heating effect of the battery, and the heat dissipation model is used to describe the influence of external cooling conditions on the temperature change of the battery.

[0048] During the operation of the battery system, the change of the battery temperature is usually determined by two types of thermal effects, i.e., internal heat generation and external heat dissipation. The internal heat generation is mainly caused by factors such as electrochemical reaction heat and ohmic heat, while the external heat dissipation is closely related to factors such as cooling medium flow, cooling medium inlet temperature and environmental conditions. If the above different thermal effects are modeled uniformly, the relative influence of various thermal effects under different operating conditions is easily obscured, which is not conducive to accurately depicting the temperature change trend.

[0049] Therefore, in the present embodiment, the thermodynamic model is divided into a heating model and a heat dissipation model. The heating model is used to model the temperature rise effect caused by the internal heat source of the battery, and the heat dissipation model is used to model the temperature drop effect caused by the heat dissipation of the battery system to the external environment. Through the above method, the internal heating mechanism and the external cooling mechanism are distinguished at the model structure level, which is conducive to depicting the contribution of different thermal effects to the temperature change.

[0050] On this basis, the dynamic function in the neural ordinary differential equation can be further expressed as a weighted combination of the heating model and the heat dissipation model. The temperature change rate calculated through the weighted combination is further used to obtain the temperature prediction result at the corresponding time through the numerical integration process of the neural ordinary differential equation, so as to represent the comprehensive effect of the temperature change trend under different operating conditions. Specifically, the predicted output of the battery temperature can be expressed as: ; wherein, is the temperature prediction output, representing the predicted temperature distribution of all single cells in the battery pack at a specific time; is a dynamic weight parameter, which is automatically learned through a neural network, and is used to adjust the relative weight of the heating effect and the heat dissipation effect under different operating conditions; is a heating function, used to simulate the heating effect of the internal heat source of the battery; is a heat dissipation function, used to simulate the heat loss of the battery system to the external environment; is an input comprehensive feature, which is composed of operating condition features and current battery temperature state.

[0051] By introducing the dynamic weight parameter to fuse the heating model and the heat dissipation model, when the internal heating of the battery dominates, the model can correspondingly increase the weight of the heating model; when the cooling condition changes have a more significant impact on the temperature, the model can correspondingly increase the weight of the heat dissipation model. This structure enables the model to adaptively adjust the contribution proportion of different heat effects according to the change of the operating condition, so as to more reasonably describe the temperature change process.

[0052] As a preferred embodiment, the neural ordinary differential equation model is solved by a numerical integration algorithm to obtain the continuous evolution result of the battery temperature with time.

[0053] In the process of constructing the battery thermal dynamics model based on the neural ordinary differential equation, the change relationship of the battery temperature state with respect to time can be expressed as: ; wherein, represents the single-cell temperature state vector of the battery system at time ; represents the operating condition input, is a temperature change rate function parameterized by a neural network. To obtain the change trajectory of the battery temperature within a given time interval, the above neural ordinary differential equation needs to be numerically solved.

[0054] In this embodiment, the fourth-order Runge-Kutta method (RK4) is preferably used for numerical integration of the neural ordinary differential equation. Let the initial value problem be expressed as: ; Then at a time step of Under the given conditions, the iterative process of RK4 can be expressed as: ; ; ; ; The battery temperature status update result at the next time step is as follows: ; Through the numerical integration process described above, the evolution trajectory of battery temperature over time can be gradually obtained in a continuous time dimension. Compared with low-order numerical integration methods, the fourth-order Runge-Kutta method has higher integration accuracy while ensuring numerical stability, which is beneficial for accurately characterizing the dynamic change process of battery temperature under large time steps, thereby improving the reliability of continuous-time prediction results.

[0055] During the model training phase, in order to optimize the parameters in the neural network's ordinary differential equation model, it is necessary to calculate the gradient of the prediction error with respect to the network parameters. In this embodiment, the adjoint method is preferably used for gradient calculation.

[0056] Specifically, the loss function is expressed as: ; in, This is the temperature prediction result at the termination time obtained through numerical integration. The adjoint sensitivity method introduces adjoint variables. The corresponding adjoint equation is solved in reverse time, thereby enabling the calculation of the gradient of the model parameters without having to save all intermediate states during the forward integration process.

[0057] By employing the adjoint sensitivity method, this implementation can significantly reduce the memory requirements for storing intermediate states during training while ensuring the correctness of gradient calculation. This enables the model to support modeling of longer time series or higher-dimensional battery temperature states, which is beneficial to improving the trainability and engineering applicability of the model in large-scale battery systems.

[0058] Example 2

[0059] See Figure 2The application also provides a battery temperature prediction system based on neural ordinary differential equation, which is used to realize continuous time prediction of the temperature of each battery cell in the battery system by using the method of embodiment 1. Specifically, the prediction system comprises: a data acquisition module, which is used to acquire multi-source operation data of the battery system during operation. The multi-source operation data includes but is not limited to sampling time, ambient temperature, battery cell heat generation, battery working current, cooling medium flow rate and cooling medium inlet temperature and the like. By acquiring the operation data reflecting the internal heat generation characteristics and external cooling conditions of the battery at the same time, the system can comprehensively consider different thermal influence factors in the subsequent modeling process, and provide a relatively complete input information basis for battery temperature prediction.

[0060] a data preprocessing module, which is used to perform time interpolation processing on the multi-source operation data acquired by the data acquisition module, so as to construct continuous time input features. Through the interpolation operation, all input variables (including time, tamb, cellHeat, current, flowrate, inletTemp) and the 4160 battery temperature sequences of the output can be uniformly mapped to continuous time points, so as to meet the requirement of neural ordinary differential equation for continuous time dynamic characteristics. After the interpolation is completed, the input matrix X and the output vector Y in the continuous time sense are obtained, which provide a consistent data basis for subsequent feature extraction and dynamic modeling. Since different operation parameters are often collected by different sensors, the sampling period and sampling time are inconsistent. The data preprocessing module uniformly maps the original discrete sampling data to the continuous time dimension through the time interpolation algorithm, so that various operation parameters participate in the subsequent processing under the same time scale. Through the above mode, it is favorable to reduce the influence of inconsistent sampling time on the modeling process, thereby improving the continuity and consistency of the battery temperature change process in the time dimension.

[0061] The feature extraction module includes a global feature extraction submodule and a discriminative feature extraction submodule, and is configured to generate a feature vector representing the thermal state of the battery from continuous-time input features. The global feature extraction submodule is configured to model the overall trend of the multi-source operating data at continuous time to obtain global features reflecting the overall thermal state of the battery system. The discriminative feature extraction submodule is configured to extract features from operating parameters such as battery heat and cooling medium flow rate, which are highly correlated with thermal behavior. Specifically, the global feature extraction submodule receives the interpolated complete input variable X. The module uses a deep neural network structure to model and abstract the coupling relationship between the six types of input parameters, i.e., time, tamb, cellHeat, current, flowrate, and inletTemp, at continuous time. The module outputs a 128-dimensional global feature vector, which is used to reflect the overall trend of the battery thermal state over time.

[0062] To further strengthen the characterization of the key parameters that most affect the thermal dynamics, a discriminative feature extraction submodule is provided to specially process only the two variables, i.e., flowrate and cellHeat, which have the strongest physical correlation. The discriminative module aims to strengthen the sensitivity of the model to changes in cooling capacity (flow rate) and the battery's own heat generation capacity, thereby improving the ability to capture nonlinear thermal behaviors such as thermal mutations and temperature rise rates. The module outputs another set of features and participates in subsequent modeling together with the global features.

[0063] The features from the global feature extraction submodule and the discriminative feature extraction submodule are fused and then concatenated with the interpolated 4160-dimensional output temperature vector to form a comprehensive feature vector containing global information of the battery state, discriminative information of key parameters, and information of adjacent states, thereby providing sufficient information input for the subsequent thermal effect network. By combining global features with discriminative features, it is beneficial to highlight key thermal factors under high-dimensional input conditions and avoid weakening important thermal information during feature modeling.

[0064] The thermal dynamics modeling module is configured to build a neural ordinary differential equation-based battery thermal dynamics model and to perform continuous-time prediction of the battery temperature based on the model. In this embodiment, the thermal dynamics model parameterizes the function of the battery temperature state with respect to time through a neural network to form a neural ordinary differential equation model, and solves the evolution process of the temperature with respect to time through numerical integration. Through continuous-time modeling, the system can smoothly describe the dynamic change process of the battery temperature under unequal time sampling and varying operating conditions, thereby improving the stability of the temperature prediction results.

[0065] An output module is configured to output a temperature prediction result of each single battery in the battery system at a target time. The temperature prediction result can be a transient temperature distribution of the single battery, which is used for subsequent thermal management strategy making, abnormal temperature rise monitoring, or safety control decision making.

[0066] With the modular structure, the system can effectively organize the battery temperature prediction process while ensuring clear overall structure, which is conducive to the deployment and maintenance of the system in actual engineering environment.

[0067] As a preferred embodiment, the thermodynamic modeling module further comprises a heating module, a heat dissipation module, and a fusion module, which are configured to model the internal heat generation effect and the external cooling effect of the battery respectively, and to fuse the modeling results of the two effects to form a thermodynamic expression for battery temperature prediction.

[0068] During the operation of the battery system, the battery temperature change is usually affected by both internal heat generation and external heat dissipation. The internal heat generation is mainly caused by factors such as battery electrochemical reaction and conductive loss, while the external heat dissipation is related to factors such as cooling medium flow, cooling medium inlet temperature, and environmental conditions. If the above different thermal effects are not distinguished for modeling, the ability to describe the temperature change mechanism may be weakened under complex working conditions.

[0069] Therefore, in the present embodiment, the heating module is configured to model the internal heat generation effect of the battery. The heating module takes the comprehensive features output by the feature extraction module as input, models the temperature rise trend reflecting the change of the internal heat source of the battery, and thus obtains the corresponding heating effect output. By modeling the internal heat generation effect independently, the response ability of the model to the change of the battery's own heat generation can be enhanced.

[0070] The heat dissipation module is configured to model the influence of external cooling conditions on the temperature change of the battery. The heat dissipation module also takes the comprehensive features as input, models the heat dissipation effect caused by factors such as cooling medium flow, cooling medium inlet temperature, and environmental temperature, and thus obtains the corresponding heat dissipation effect output. By modeling the external heat dissipation effect independently, the influence of the change of the cooling condition on the temperature change process can be accurately reflected.

[0071] The fusion module is configured to fuse the outputs of the heating module and the heat dissipation module. In one embodiment, the fusion module introduces a dynamic weight parameter to combine the heating effect and the heat dissipation effect, and the temperature prediction expression thereof can be expressed as: wherein, is the temperature prediction output, representing the predicted temperature distribution of each single battery in the battery system at the target time; is a dynamic weight parameter,​ automatically learned by the neural network, for adjusting the relative weights of the heating effect and the heat dissipation effect under different operating conditions; represents an output function of the heating module, represents an output function of the heat dissipation module; is an input comprehensive feature.

[0072] By introducing the fusion module and weighting the heating effect and the heat dissipation effect, when the internal heating effect of the battery dominates, the fusion module can correspondingly increase the weight of the output of the heating module; when the external cooling condition has a more significant influence on temperature change, the fusion module can correspondingly increase the weight of the output of the heat dissipation module. This structure enables the thermodynamic modeling module to adaptively adjust the contribution proportion of different thermal effects to temperature change according to the change of operating conditions, so as to more reasonably represent the change trend of battery temperature.

[0073] By setting the heating module, the heat dissipation module and the fusion module, the embodiment structurally models the internal heating effect and the external cooling effect of the battery while keeping the model structure clear, which is conducive to improving the adaptability of the thermodynamic model to complex operating conditions and providing a more stable and consistent dynamic basis for continuous-time temperature prediction based on neural ordinary differential equations.

[0074] As a preferred embodiment, the system further includes a model training module for pre-training the battery thermodynamic model based on neural ordinary differential equations using historical data or simulation data.

[0075] In battery temperature prediction applications, due to the limitations of operating environment, operating condition combination and data acquisition conditions, the number of high-quality labeled data that can be directly obtained in actual systems is usually limited. If only a small amount of real running data is used for model training, it is easy to lead to insufficient convergence of model parameters or insufficient generalization ability. Therefore, the present embodiment introduces a model training module in the system structure to pre-train the neural ordinary differential equation model in the thermodynamic modeling module.

[0076] The model training module first obtains historical running data or simulation data generated based on a battery thermal model. The simulation data can cover different load conditions, cooling conditions and environmental temperature ranges, and is used to reflect the temperature change characteristics of the battery system under various operating scenarios. By introducing simulation data, the model can be exposed to more rich operating condition combinations at the early stage of training, which is conducive to forming a stable thermodynamic representation.

[0077] In a certain embodiment, the model training module performs parameter training on the battery thermal dynamics model based on the historical data or the simulation data described above. During the training process, the model calculates a predicted temperature trajectory through numerical integration, and compares the predicted result with the corresponding real temperature or simulation temperature to build a loss function and update the model parameters. Through the pre-training process described above, the model can learn the basic evolution law of the battery temperature change over time.

[0078] By using the historical data or the simulation data for pre-training, the model training module can provide a more reasonable initial state of the parameters for the battery thermal dynamics model before model deployment, thereby reducing the dependence of the model on a large amount of real data in the actual application scenario, and being conducive to improving the stability and engineering applicability of the model under different operating conditions. At the same time, this pre-training method provides a basis for subsequent model adjustment based on a small amount of real data, which helps to shorten the model training period and improve the overall deployment efficiency of the system.

[0079] The above is only a preferred embodiment of the present application, and does not limit the patent scope of the present application. Any equivalent structural transformation made under the inventive concept of the present application, using the content of the present application specification and drawings, or direct / indirect application in other related technical fields is included in the patent protection scope of the present application.

Claims

1. A battery temperature prediction method based on the ordinary differential equation, characterized in that, The steps include: S1. Acquire multi-source operational data during the operation of the battery system; S2. Perform time interpolation processing on the multi-source operating data to uniformly map data with different sampling frequencies and unequal time intervals to the continuous time dimension to obtain continuous time input features; S3. Based on the continuous time input features, construct global features and discriminant features respectively, wherein the discriminant features are used at least to characterize the thermal influencing factors corresponding to the battery heat generation and the cooling medium flow rate; S4. The global features, discrimination features and the current temperature state of the battery are fused together to construct a comprehensive feature for describing the thermodynamic state of the battery; S5. Input the comprehensive features into the thermodynamic model constructed based on the neural network differential equation, and solve the continuous evolution process of battery temperature over time by numerical integration to obtain the cell-by-cell temperature prediction results of the battery system at the target time.

2. The battery temperature prediction method based on the neural ordinary differential equation according to claim 1, characterized in that, The thermodynamic model uses a neural network to parameterize the change function of battery temperature state with respect to time, thereby forming a neural ordinary differential equation model.

3. The battery temperature prediction method based on the neural ordinary differential equation according to claim 1, characterized in that, The time interpolation process uses an interpolation algorithm to map the original discrete sampled data into a continuous time series, so as to eliminate the impact of inconsistent sampling periods of different sensors on the modeling accuracy.

4. The battery temperature prediction method based on the neural ordinary differential equation according to claim 1, characterized in that, The thermodynamic model includes a heating model and a heat dissipation model; The heating model is used to describe the internal heating effect of the battery, while the heat dissipation model is used to describe the influence of external cooling conditions.

5. The battery temperature prediction method based on the neural ordinary differential equation according to claim 4, characterized in that, In step S5, the outputs of the heating model and the heat dissipation model are fused using weighting parameters to characterize the temperature change trend under different operating conditions.

6. The battery temperature prediction method based on the neural ordinary differential equation according to claim 5, characterized in that, The temperature prediction formula is shown below: ； in, Output for temperature prediction; For dynamic weight parameters, ∈[0,1], automatically learned through a neural network; For heating function, This is the heat dissipation function; The comprehensive features are the input.

7. The battery temperature prediction method based on the neural ordinary differential equation according to claim 2, characterized in that, The aforementioned constant differential equation model is solved using a numerical integration algorithm to obtain the continuous change of battery temperature over time.

8. A battery temperature prediction system based on the constant differential equation, employing the method described in any one of claims 1-7, characterized in that, The system includes: The data acquisition module is used to collect multi-source operating data of the battery system; The data preprocessing module is used to perform time interpolation processing on the multi-source operating data to construct continuous time input features; The feature extraction module includes a global feature extraction submodule and a discriminative feature extraction submodule, which are used to generate feature vectors that characterize the thermal state of the battery; The thermodynamic modeling module is used to construct a battery thermodynamic model based on the neural network differential equation and to make continuous-time predictions of battery temperature. The output module is used to output the predicted temperature of each cell in the battery system at the target time.

9. A battery temperature prediction system based on a neural ordinary differential equation according to claim 8, characterized in that, The thermodynamic modeling module includes a heating module, a heat dissipation module, and a fusion module, which are used to model the internal heating effect and the external cooling effect of the battery respectively and perform weighted fusion.

10. A battery temperature prediction system based on a neural ordinary differential equation according to claim 8, characterized in that, It also includes a model training module for pre-training the battery thermodynamics model based on the neural ordinary differential equation using historical data or simulation data.

Images