Large-scale lithium battery temperature field space-time modeling method and system

Abstract

The invention discloses a large-scale lithium battery temperature field spatio-temporal modeling method and system, and relates to the technical field of temperature fields of battery systems, and the method comprises the steps: obtaining spatio-temporal data of a battery temperature field, and carrying out the normalization processing and down-sampling processing of the spatio-temporal data, and obtaining sparse sensing data; constructing a distributed thermal model according to the space-time separation network, the space-time synthesis network and the nonlinear time dynamics unit; and inputting the sparse sensing data into the distributed thermal model to obtain temperature field prediction data of the large-scale lithium battery. According to the distributed thermal model constructed by the method, large-scale lithium battery temperature field prediction, total space reconstruction and virtual sensing of non-sensing positions under the condition of limited sensors are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of temperature field of battery system, and particularly relates to a large-scale lithium battery temperature field space-time modeling method and system. BACKGROUND

[0002] Accurate modeling of lithium-ion battery pack thermal processes is particularly important for the energy storage industry. During charge-discharge cycling, internal electrochemical reactions release heat in a spatially inhomogeneous manner. The resulting temperature gradients not only accelerate degradation and reduce cycle life, but can also trigger thermal runaway of the battery system. Therefore, accurate space-time temperature modeling is crucial for reliable battery management systems (BMS), efficient cooling design, and early fault diagnosis.

[0003] Traditional model-based methods attempt to describe the thermal processes of batteries by directly solving the underlying partial differential equations. Representative techniques include finite element analysis, finite difference method, and spectral method. These techniques have high fidelity and interpretability when the governing equations, thermal physical parameters, and boundary conditions are accurately known. However, in practice, the heat generation term in electrochemical kinetics is difficult to parameterize online, the boundary conditions fluctuate with environmental airflow and aging, and the three-dimensional transient calculation burden of partial differential equations (PDE) is extremely challenging for onboard battery management system hardware.

[0004] Data-driven methods circumvent the derivation of explicit PDEs by learning the space-time dynamics directly from sensor measurements. However, the limitations of data-driven methods are: 1) a large number of sensors are usually required to accurately capture the spatial pattern; 2) the resulting spatial basis functions are discrete, so they cannot provide temperature estimates at non-sensor locations; 3) there is currently no unified feature extraction framework that automatically adapts to different battery systems, forcing engineers to resort to laborious trial-and-error design.

[0005] Therefore, in order to solve the above problems, a nonlinear spatial reconstruction method is needed for space-time modeling of large-scale lithium battery systems under limited sensor conditions. SUMMARY

[0006] The purpose of the present application is to provide a large-scale lithium battery temperature field space-time modeling method and system, which realizes large-scale lithium battery temperature field prediction, full-space reconstruction, and virtual sensing at non-sensing locations under limited sensor conditions.

[0007] To achieve the above purpose, the present application provides a large-scale lithium battery temperature field space-time modeling method, comprising the following steps: S1. Acquire spatiotemporal data of the battery temperature field, and perform normalization and downsampling processing on the spatiotemporal data to obtain sparse sensing data. S2. Construct a distributed thermal model based on the spatiotemporal separation network, spatiotemporal synthesis network, and nonlinear time dynamics unit; S3. Input the sparse sensing data into the distributed thermal model to obtain large-scale lithium battery temperature field prediction data.

[0008] Preferably, the specific content of constructing the distributed thermal model in S2 based on the spatiotemporal separation network, the spatiotemporal synthesis network, and the nonlinear time dynamics unit includes: Acquire historical spatiotemporal data and known temperature field data; Normalize the historical spatiotemporal data to obtain normalized historical spatiotemporal data; Normalized historical spatiotemporal data are downsampled to obtain historical sparse sensor data. Training and testing sets are constructed based on historical sparse sensing data and known temperature field data; Based on the convolutional neural network, feature extraction is performed on historical sparse sensor data through convolution and downsampling operations to construct a spatiotemporal separation network and obtain temporal features and spatial basis functions. Based on the deconvolutional neural network, the temporal features and spatial basis functions are deconvolved to construct a spatiotemporal synthesis network, thereby obtaining the reconstructed full-space temperature field. Based on the long short-term memory neural network and the time feature sequence, a nonlinear time dynamics unit is constructed to obtain the predicted time features; Based on the reconstructed full-space temperature field and predicted time characteristics, an initial distributed thermal model is constructed. The training set is input into the initial distributed heat model for training, and the trained initial distributed heat model is obtained. The test set is input into the trained initial distributed heat model, and the trained initial distributed heat model is adjusted to obtain the trained initial distributed heat model. The trained initial distributed thermal model is defined as the distributed thermal model.

[0009] Preferably, the expression for the normalized historical spatiotemporal data is: ; in, To normalize historical spatiotemporal data, i It is a spatial location number. j For timestamp sequence number, For the first The spatial coordinates of each sensor For the first The sampling time corresponding to each sample for Time of the first The sensor is located in space. The collected temperature, The minimum value in the training set. This represents the maximum value in the training set.

[0010] Preferably, the expression for the historical sparse sensing data is: ; in, For timestamp sequence number, For the first The sampling time corresponding to each sample for A full-space snapshot of a moment. for time downsampling snapshot, For downsampling function, This is the downsampling matrix. Represents Kronecker.

[0011] Preferably, the expression for the time feature is: ; in, k For feature map sequence number, j For timestamp sequence number, For the first The sampling time corresponding to each sample for Time of the first k The latent temporal representation corresponding to each feature map It is the Sigmoid activation function. for time downsampling snapshot, for Historical sparse sensor data corresponding to the location of all sensors at any given time. This is a high-dimensional convolution operator. For convolution kernel, For the first k The deviation of each feature map.

[0012] Preferably, the expression for the reconstructed full-space temperature field is: ; in, For timestamp sequence number, For the first The sampling time corresponding to each sample To reconstruct the full-space temperature field, It is the Sigmoid activation function. For feature map sequence number, , The number of feature maps, for Time of the first The latent temporal representation corresponding to each feature map This is a high-dimensional convolution operator. This is a weight flipping operation. represents the bias of each potential graph in the spatiotemporal separation network.

[0013] Preferably, the expression for the predicted time feature is: ; ; in, For timestamp sequence number, For the first The sampling time corresponding to each sample for The first feature map at time step 1. for The second feature map at time step 2. for Time of the first n Each feature map for time A one-dimensional vector composed of the reconstructed vectors corresponding to each feature map. for One-dimensional vector at time The prediction results It is a nonlinear model. A one-dimensional vector The vector corresponding to the previous second, One-dimensional vector forward The vector corresponding to seconds, for The input signal vector of the system at any given time, for The input signal vector corresponding to the previous second, for forward The input signal vector corresponding to a second. The time delay is a time-dependent characteristic. The time delay is the input.

[0014] Preferably, the expression for the distributed thermal model is: ; wherein, is a timestamp index, is a sampling time corresponding to the th sample, is a distributed thermal model of the whole space at the th time, is a Sigmoid activation function, is a feature map index, , is a number of feature maps, is th predicted temporal feature map at the th time, is a high-dimensional convolution operator, is a weight flipping operation, is a bias of each latent map in the spatio-temporal synthesis network.

[0015] Preferably, the loss function of the spatio-temporal synthesis network is: ; wherein, is a timestamp index, is a sampling time corresponding to the th sample, is a whole space snapshot at the th time, is a parameter of the spatio-temporal synthesis network, is a loss function of the spatio-temporal synthesis network, is a number of samples in the training set, is a reconstructed whole space temperature field.

[0016] The application also provides a large-scale lithium battery temperature field spatio-temporal modeling system, comprising: a data acquisition module configured to acquire spatio-temporal data of a battery temperature field, and perform normalization processing on the spatio-temporal data to obtain normalized spatio-temporal data; a model construction module configured to construct a distributed thermal model according to a spatio-temporal separation network, a spatio-temporal synthesis network and a nonlinear time dynamics unit; a data prediction module configured to input the normalized spatio-temporal data into the distributed thermal model to obtain temperature field prediction data of the lithium battery.

[0017] In summary, the large-scale lithium battery temperature field spatio-temporal modeling method and system of the application have the following beneficial effects compared with the prior art: the distributed thermal model constructed by the spatio-temporal separation network, the spatio-temporal synthesis network and the nonlinear time dynamics unit not only realizes large-scale lithium battery temperature field prediction under the condition of limited sensors, but also realizes whole space reconstruction of the large-scale lithium battery temperature field and virtual sensing of non-sensing positions under the condition of limited sensors.

[0018] The technical method of the present application is described in further detail below with reference to the accompanying drawings and examples. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 Flow chart of the large-scale lithium battery temperature field space-time modeling method of the present application; Figure 2 Structural framework chart of the large-scale lithium battery temperature field space-time modeling method of the present application; Figure 3 Three-dimensional structure chart of the lithium battery pack used in the experiment of the present application; Figure 4 Open circuit voltage (OCV) and state of charge (SOC) curve chart of the lithium battery monomer used in the experiment of the present application; Figure 5 Lithium battery pack sensor distribution chart used in the experiment of the present application; Figure 6 Temperature distribution chart of the lithium battery pack at 1000 s under standard conditions of the present application; Figure 7 Temperature prediction output chart of the lithium battery pack at 1500 s under 1C discharge rate without air flow of the present application; Figure 7 (a) in is the predicted temperature at 1500 s without air flow disturbance under 1C discharge rate; Figure 7 (b) in is the absolute error at 1500 s without air flow disturbance under 1C discharge rate; Figure 8 Temperature prediction output of the lithium battery pack at 1500 s under 2C discharge rate without air flow of the present application; Figure 8 (a) in is the predicted temperature at 1500 s without air flow disturbance under 2C discharge rate; Figure 8 (b) in is the absolute error at 1500 s without air flow disturbance under 2C discharge rate; Figure 9 Temperature distribution chart of the lithium battery pack at 1000 s under air flow disturbance of the present application; Figure 10 Temperature prediction output chart of the lithium battery pack at 1500 s under air flow disturbance and 1C discharge rate of the present application; Figure 10 (a) in is the predicted temperature distribution chart of the lithium battery pack at 1500 s under air flow disturbance and 1C discharge rate; Figure 10 (b) in is the absolute prediction error chart of the lithium battery pack at 1500 s under air flow disturbance and 1C discharge rate; Figure 11 Figure (a) in the present application is a predicted temperature distribution map of the lithium battery pack at 1500 s under air flow disturbance and 2C discharge rate; Figure 11 Figure (a) in the present application is a predicted temperature distribution map of the lithium battery pack at 1500 s under air flow disturbance and 2C discharge rate; Figure 11 Figure (b) in the present application is an absolute prediction error map of the lithium battery pack at 1500 s under air flow disturbance and 2C discharge rate; Figure 12 Figure in the present application is a module diagram of a large-scale lithium battery temperature field space-time modeling system. DETAILED DESCRIPTION

[0020] The technical method of the present application is further described below by means of the accompanying drawings and examples. It should be noted that the relative arrangement, numerical expressions and values of the components and steps set forth in these examples do not limit the scope of the present application unless otherwise specifically stated.

[0021] The following description of at least one exemplary embodiment is merely exemplary in nature and is in no way intended to limit the present application or its application or uses.

[0022] Techniques, systems, and devices known to those of ordinary skill in the relevant art can not be discussed in detail herein. However, where appropriate, techniques, systems, and devices should be considered part of the description of the present application.

[0023] In all of the examples shown and discussed herein, any specific values should be interpreted as merely exemplary and not as a limitation. Thus, other examples of the exemplary embodiments can have different values.

[0024] Unless otherwise defined, technical terms or scientific terms used in the present application should be interpreted as having the ordinary meaning understood by one of ordinary skill in the art to which the present application pertains.

[0025] As shown in Figure 1 The present application provides a large-scale lithium battery temperature field space-time modeling method, and a structure framework diagram thereof is shown in Figure 2 Specifically, the present application comprises the following steps: Step S1, obtain the space-time data of the battery temperature field, and perform normalization processing and down-sampling processing on the space-time data to obtain sparse sensing data.

[0026] Step S2, construct a distributed thermal model according to a space-time separation network, a space-time synthesis network and a nonlinear time dynamics unit.

[0027] Further, step S2 can be replaced by the following steps S201-S211: Step S201, obtain historical space-time data and known temperature field data. The measured historical space-time data can be expressed as ,in, i It is a spatial location number. j For timestamp sequence number, For the first The spatial coordinates of each sensor For the first The sampling time corresponding to each sample For historical spatiotemporal data, For the number of sensors, This represents the number of training snapshots.

[0028] Step S202: Normalize the historical spatiotemporal data to obtain normalized historical spatiotemporal data. Specifically, normalize the historical spatiotemporal data to ensure that the values ​​are within the standardized range. Normalizing the historical spatiotemporal data is crucial for improving the performance and convergence of the distributed thermal model during training.

[0029] In an exemplary embodiment of the present invention, minimum-maximum normalization is used to adjust historical spatiotemporal data to a range between 0 and 1. The expression for normalizing historical spatiotemporal data is: ; in, To normalize historical spatiotemporal data, for Time of the first The sensor is located in space. The collected temperature, The minimum value in the training set. This represents the maximum value in the training set.

[0030] This transformation enables distributed thermal models to handle input values ​​with relatively consistent numerical ranges, thereby preventing weight explosion during training and accelerating optimization.

[0031] Step S203: Perform a downsampling operation on the normalized historical spatiotemporal data to obtain historical sparse sensing data. The expression for the historical sparse sensing data is: ; Among them, symbols For sampling data of the entire space, for A full-space snapshot of a moment. for time downsampling snapshot, For downsampling function, The downsampling matrix ( Each element in the table represents the sensor status, where 1 indicates that a sensor is present at that location, and 0 indicates that no sensor is present at that location. representing kroncker product.

[0032] Step S204, constructing a training set and a test set according to historical sparse sensing data and known temperature field data.

[0033] Step S205, according to a convolutional neural network, performing feature extraction on the historical sparse sensing data through convolution operation and down-sampling operation, constructing a space-time separation network, and obtaining time features and spatial basis functions.

[0034] Further, the specific content of step S205 is as follows: Through the convolution operation, the convolution kernel is used to realize the space-time separation and extract the potential time features, and the specific expression is as follows: ; Wherein, k is the feature map sequence number, is the potential time expression corresponding to the k th feature map at the moment, is a Sigmoid activation function, is a high-dimensional convolution operator symbol, is a convolution kernel, is the bias of the k th feature map.

[0035] The specific expression of the convolution kernel is , wherein, is a real set, is a dimension, which is used to capture local patterns in data, such as gradients and textures.

[0036] The feature map generated by the convolution operation will then pass through a series of down-sampling layers, thereby reducing the feature dimension and increasing the number of feature channels. The down-sampling process not only reduces the computational burden, but also enhances the ability of the network to capture global spatial correlation.

[0037] Similar to the one-dimensional space-time transformation neural network, the convolution kernel in the present application is fixed after the distributed thermal model training, and can be regarded as a spatial basis function (SBF) in system modeling. At the same time, the time feature corresponds to the time coefficient.

[0038] Step S206, according to the deconvolutional neural network, performing deconvolution calculation on the time features and the spatial basis functions, constructing a space-time synthesis network, and obtaining a reconstructed full-space temperature field.

[0039] Further, the specific content of step S206 is as follows: The time feature map of sparse temperature field data is extracted by a convolutional neural network, and then a deconvolutional neural network is used for full space reconstruction, thereby solving the limitations of traditional autoencoders in lithium battery temperature field data reconstruction performance. The traditional autoencoder mainly focuses on reconstructing data from measurement points, which limits its ability to capture global spatial information. To overcome this challenge, the present application first performs downsampling operation on the full space temperature field data to obtain sparse temperature field data, then uses a convolutional neural network to extract the time feature map of the sparse temperature field data, and finally effectively improves the data resolution by transposed convolution operation, thereby realizing full space reconstruction.

[0040] In order to realize full space reconstruction, it is necessary to downsample the normalized historical spatio-temporal data to enhance the generalization ability of the model. This process is crucial for full space reconstruction, because the number of sensors is often small during online prediction.

[0041] Downsampling the normalized historical spatio-temporal data ensures that sparse sensing data can be used to effectively reconstruct the entire spatial domain during the online prediction process of the distributed thermal model, without the need for a large number of sensors.

[0042] The spatio-temporal synthesis network combines spatial and temporal features into a comprehensive representation. It uses deconvolution operation to expand the feature map from low-dimensional representation to target size. By setting appropriate step size and padding, it ensures that the dimension of the reconstructed output is greater than the original dimension. The expression of the reconstructed full space temperature field obtained by deconvolution is: ; where, is the flip operation of the weight, is the bias of each latent map in the spatio-temporal separation network, so that each filter focuses on the features of the entire input rather than a single pixel.

[0043] The loss function between the input after upsampling the normalized historical spatio-temporal data and the reconstructed full space temperature field, i.e. the loss function of the spatio-temporal synthesis network, is: ; where, is the parameter of the spatio-temporal synthesis network, is the loss function of the spatio-temporal synthesis network.

[0044] The optimization goal is to minimize the reconstruction error, which can be achieved by adjusting the network weights through backpropagation algorithm.

[0045] By introducing the spatio-temporal synthesis network and using deconvolution operation, the present application can realize full space data reconstruction and improve data resolution, not only making up for the shortcomings of traditional autoencoders, but also providing a more comprehensive perspective for analyzing complex systems.

[0046] Step S207, constructing a nonlinear time dynamics unit according to the long short-term memory neural network and the time feature sequence to obtain a predicted time feature. An expression of the predicted time feature is as follows: ; ; wherein, is a first feature map at the moment, is a second feature map at the moment, is an mth feature map at the moment, n is a one-dimensional vector composed of a reconstruction vector corresponding to the mth feature map at the moment, is a predicted result of the one-dimensional vector at the moment, is a nonlinear model, is a vector corresponding to the one-dimensional vector at the first second, is a vector corresponding to the one-dimensional vector at the second second, is an input signal vector of the system at the moment, is an input signal vector corresponding to the first second, is an input signal vector corresponding to the second second, is a time lag of the time feature, is a time lag of the input.

[0047] Further, the specific content of step S207 is as follows: In order to realize large-scale lithium battery temperature field prediction, it is necessary to deduce a time dynamics equation based on the extracted time features . Taking a two-dimensional space as an example: first, the matrix is flattened into a vector form . The time feature vector at the moment can be expressed as: .

[0048] For the above nonlinear identification problem, the time dynamics equation in the nonlinear time dynamics unit can be expressed as follows: .

[0049] ​​​​​​​​​​​​​​​Since Long Short-Term Memory (LSTM) neural networks can handle long-term dependencies and model complex nonlinear relationships, and temperature data from distributed parameter systems often contain complex nonlinear dynamic information, this paper utilizes LSTM networks to identify nonlinear time dynamic units. The architecture of an LSTM network consists of memory cells, input gates, forget gates, and output gates. These gates regulate the information flow, enabling the network to selectively retain or forget information over time. The cell state update within an LSTM cell is as follows: ; in, for t The cell state at time 10:00. The activation value for the forget gate. for The cell state at time -1 The input gate activation value, This is the candidate cell state. The hidden state is updated based on the output gate: ; in, In hidden state, This is the output gate.

[0050] Finally, train the LSTM network to minimize The mean square error between the predicted and actual values: ; in, For the parameters of the LSTM network, To minimize The mean squared error between the predicted and actual values ​​is considered. The LSTM network was chosen over other network architectures because it prevents overfitting and has strong generalization ability to unknown data. Furthermore, LSTM networks perform exceptionally well in sequence prediction tasks, which are crucial for capturing the temporal dynamics of distributed thermal processes.

[0051] Step S208: Construct an initial distributed thermal model based on the reconstructed full-space temperature field and predicted time characteristics.

[0052] Furthermore, the specific content of step S208 is as follows: After obtaining the nonlinear time-dynamic unit, a distributed thermal model can be constructed, as shown in the following expression: ; in, for A time-space distributed heat model. a flipping operation for the weights, Bias for each latent graph in the spatio-temporal synthesis network. The distributed thermal model not only realizes the prediction of lithium battery temperature field, but also completes the full space reconstruction, realizing the virtual perception of non-perception position.

[0053] Step S209, input the training set into the initial distributed thermal model for training, and obtain the trained initial distributed thermal model.

[0054] Step S210, input the test set into the trained initial distributed thermal model, adjust the trained initial distributed thermal model, and obtain the trained initial distributed thermal model.

[0055] Step S211, determine the trained initial distributed thermal model as the distributed thermal model.

[0056] S3, input the sparse sensing data into the distributed thermal model to obtain the temperature field prediction data of the large-scale lithium battery.

[0057] Experimental verification (1) Experimental setup As shown in Figure 3 , a lithium battery pack composed of 121 lithium battery monomers is used to verify the effectiveness of the provided large-scale lithium battery temperature field spatio-temporal modeling method. In the lithium battery pack, 11 lithium battery monomers are first connected in parallel, and then 11 parallel connections are connected in series. The lithium battery monomer used is 21700 type. Air flow facing the left side of the lithium battery pack is used to simulate heat dissipation, where v represents the air flow velocity. The main lithium battery pack parameters are shown in Table 1.

[0058] Table 1 Main lithium battery pack parameter table

[0059] Due to the limitations of experimental equipment and conditions, the large-scale lithium battery temperature field spatio-temporal modeling method is used to simulate the thermal evolution process of the lithium battery pack. Before simulating the thermal process of the lithium battery pack, it is necessary to identify the relationship function between the open circuit voltage (OCV) and the state of charge (SOC) of the lithium battery monomer. The OCV-SOC curve of the lithium battery monomer is determined by using the parameter identification method, and the specific expression after identification is as follows: ; where, is the state of charge SOC of the lithium battery. The identified OCV-SOC curve is shown in Figure 4 .

[0060] During offline training, a temperature sensor is placed on the top of each lithium battery unit. As shown inFigure 5 As shown, spatial and temporal prediction of the entire lithium-ion battery pack can be achieved using only 36 sparse sensors during the online process. Temperature data for each lithium-ion cell in the battery pack can be collected under different test conditions using simulation methods. The sampling frequency is 1 Hz. A total of 2,000 snapshots are collected. The first 800 snapshots were used for model training, and the last 1200 were used for testing. Each snapshot contains temperature data for 121 lithium battery cells.

[0061] (2) Experimental verification under standard conditions Under standard conditions, without airflow interference, the system is in a state of natural convection. Figure 5 This represents the temperature distribution of a lithium-ion battery pack under standard conditions for 1000 seconds. From... Figure 6 It can be seen that, in the absence of airflow interference, the battery cells near the edge have better heat dissipation conditions, and therefore their temperature is lower than that of the battery cells surrounding the center.

[0062] Distributed temperature modeling results of a large-scale lithium battery temperature field spatiotemporal modeling method are as follows: Figure 7 and Figure 8 As shown. Figure 7 (a) shows the predicted temperature at 1500 s using a large-scale lithium battery temperature field spatiotemporal modeling method without airflow interference at a 1C discharge rate. Figure 7 In the figure, (b) represents the absolute error of the large-scale lithium battery temperature field spatiotemporal modeling method at 1C discharge rate without airflow interference, at 1500 s. Figure 7 As shown in (a) of the prediction results, the temperature of most lithium battery cells is concentrated around 26.5℃, the temperature of the edge lithium battery cells is concentrated around 25.5℃, and the temperature of the lithium battery cells in the four corners is the lowest, around 25℃. Figure 7 As shown in (b), the absolute prediction error at all locations is less than 0.6℃.

[0063] Figure 8 (a) shows the predicted temperature at 1500 s under 2C discharge rate and no airflow interference, using a large-scale spatiotemporal modeling method for the temperature field of lithium batteries. Figure 8 In the figure, (b) represents the absolute error of the large-scale lithium battery temperature field spatiotemporal modeling method at 1500 s under 2C discharge rate and without airflow interference. Figure 8 From (a) we can see that, with Figure 7 Compared to the 1C discharge rate in (a) of the above, the temperature of each individual cell increased at the 2C discharge rate, but the overall temperature distribution trend was similar to that at 1C, with higher temperatures at the center and lower temperatures at the edges. Figure 8As shown in (b), the absolute prediction error at most locations is less than 0.5 ℃, while the maximum absolute prediction error occurs at the four corners, which is about 0.8 ℃.

[0064] The performance of the large-scale lithium battery temperature field spatiotemporal modeling method provided in this invention was compared with that of the traditional Karhunen-Loève (KL) method, incremental learning method, and spatial construction method. Given that the traditional KL method and the incremental learning-based method cannot achieve full-space modeling under limited sensing conditions, both methods use linear interpolation to approximate temperature data at non-sensor locations.

[0065] Table 2 lists the comparison results of different methods under 1C discharge rate and no airflow interference conditions, with the best-performing method highlighted in bold. The root mean square error (RMSE) of the large-scale lithium battery temperature field spatiotemporal modeling method provided by this invention on the training set is 0.0443, slightly higher than the training set RMSE of the traditional KL method (0.0404). However, compared with all other methods (KL method, incremental learning method, and spatial construction method), the large-scale lithium battery temperature field spatiotemporal modeling method provided by this invention exhibits the smallest test set RMSE, indicating better online performance. Therefore, in comprehensive evaluation, the test set RMSE of this invention demonstrates better modeling performance.

[0066] Table 2. Performance Comparison of Different Methods under 1C Discharge Rate and No Airflow Interference Conditions

[0067] (3) Experimental verification under airflow interference conditions airflow velocity under airflow disturbance conditions m / s. The temperature distribution of the lithium battery pack at 1000 s is as follows. Figure 9 As shown in the diagram, it can be observed that under airflow interference conditions, the temperature of the lithium battery cells directly facing the airflow is relatively low, approximately 21 °C. In contrast, the temperature of lithium battery cells farther from the airflow inlet is higher, with the highest temperature exceeding 30 °C. Figure 6 Compared to scenarios without airflow interference, the temperature difference under airflow conditions is significantly larger, with the difference between the highest and lowest temperatures exceeding 10 °C. Therefore, accurate modeling of the thermal process of lithium battery packs is necessary to better design battery cooling systems.

[0068] Under airflow disturbance conditions Figure 10 and Figure 11 The results of distributed temperature modeling at 1C and 2C rates are presented using the spatiotemporal modeling method for the temperature field of large-scale lithium batteries. Figure 10(b) in FIG. 4 is an absolute prediction error map of the lithium battery pack at 1500 s under air flow disturbance and 1C discharge rate, Figure 10 (b) in FIG. 4 is an absolute prediction error map of the lithium battery pack at 1500 s under air flow disturbance and 1C discharge rate, Figure 11 (a) in FIG. 5 is a predicted temperature distribution map of the lithium battery pack at 1500 s under air flow disturbance and 2C discharge rate, Figure 11 (b) in FIG. 5 is an absolute prediction error map of the lithium battery pack at 1500 s under air flow disturbance and 2C discharge rate. The temperature distribution of the lithium battery pack under two different discharge rates shows similar trends. According to Figure 10 (a) in FIG. 4 and Figure 11 According to the modeling results at 1500 s in (a) in FIG. 4, the lithium battery cells close to the air inlet have temperatures close to 20 ℃ at 1C and 2C rates. In contrast, the lithium battery cells far from the air inlet have higher temperatures at 2C discharge rate, exceeding 30 ℃, while the temperatures at these locations are only about 25 ℃ at 1C rate. At 1C discharge rate, the absolute prediction error of all lithium battery cells using the large-scale lithium battery temperature field spatiotemporal modeling method is less than 0.4 ℃. In contrast, the error is slightly higher at 2C discharge rate, but still remains below 0.6 ℃.

[0069] Similar to the case without air flow disturbance, at 1C discharge rate with air flow disturbance, the performance of the large-scale lithium battery temperature field spatiotemporal modeling method was compared with that of the traditional KL method, the incremental learning method and the spatial construction method, and the results are shown in Table 3. Under the condition of air flow disturbance, the test performance of the large-scale lithium battery temperature field spatiotemporal modeling method is slightly inferior to that of the traditional KL method, but compared with similar methods, the large-scale lithium battery temperature field spatiotemporal modeling method still shows excellent test performance, with a test RMSE of 0.1903.

[0070] Table 3 Performance comparison between different methods under 2C discharge rate and air flow disturbance

[0071] The present application also provides a large-scale lithium battery temperature field spatiotemporal modeling system, as shown in FIG. 6, comprising: Figure 12 a data acquisition module for acquiring spatiotemporal data of the battery temperature field and performing normalization processing on the spatiotemporal data to obtain normalized spatiotemporal data. a model construction module for constructing a distributed thermal model according to a spatiotemporal separation network, a spatiotemporal synthesis network and a nonlinear time dynamics unit.

[0072] a data prediction module for inputting the normalized spatiotemporal data into the distributed thermal model to obtain temperature field prediction data of the lithium battery.

[0073] a data prediction module for inputting the normalized spatiotemporal data into the distributed thermal model to obtain temperature field prediction data of the lithium battery.

[0074] Finally, it should be noted that the above examples are intended to illustrate the technical method of the present application, but not to limit it. Although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical method of the present application can be modified or equivalently replaced, and these modifications or equivalent replacements should not make the modified technical method deviate from the spirit and scope of the technical method of the present application.

Claims

1. A method for spatiotemporal modeling of temperature field in large lithium batteries, characterized in that, The large-scale lithium battery space-time modeling method comprises the following steps: S1, obtaining the space-time data of the battery temperature field, and performing normalization processing and down-sampling processing on the space-time data to obtain sparse sensing data; S2, constructing a distributed thermal model according to a space-time separation network, a space-time synthesis network and a nonlinear time dynamics unit; S3, inputting the sparse sensing data into the distributed thermal model to obtain temperature field prediction data of the large-scale lithium battery.

2. The method according to claim 1, wherein, The specific content of constructing the distributed thermal model according to the space-time separation network, the space-time synthesis network and the nonlinear time dynamics unit in S2 comprises: obtaining historical space-time data and known temperature field data; performing normalization processing on the historical space-time data to obtain normalized historical space-time data; performing down-sampling operation on the normalized historical space-time data to obtain historical sparse sensing data; constructing a training set and a test set according to the historical sparse sensing data and the known temperature field data; extracting features of the historical sparse sensing data by convolution operation and down-sampling operation according to a convolutional neural network to construct a space-time separation network, and obtaining time characteristics and spatial basis functions; performing deconvolution calculation on the time characteristics and the spatial basis functions according to a deconvolutional neural network to construct a space-time synthesis network, and obtaining a reconstructed full-space temperature field; constructing a nonlinear time dynamics unit according to a long short-term memory neural network and a time characteristic sequence to obtain a predicted time characteristic; constructing an initial distributed thermal model according to the reconstructed full-space temperature field and the predicted time characteristic; inputting the training set into the initial distributed thermal model for training to obtain a trained initial distributed thermal model; inputting the test set into the trained initial distributed thermal model to adjust the trained initial distributed thermal model to obtain a trained initial distributed thermal model; determining the trained initial distributed thermal model as the distributed thermal model.

3. The method according to claim 2, wherein, The expression of the normalized historical space-time data is: ； wherein, is the normalized historical spatiotemporal data, i is the spatial position sequence number, j is the timestamp sequence number, is the spatial position coordinate of the sensor, is the sampling time corresponding to the sample, is the time the sensor at the spatial position collected the temperature, is the minimum value of the training set, is the maximum value of the training set.

4. The method according to claim 2, wherein, The expression of the historical sparse sensing data is: ； wherein, is the timestamp number, is the sampling time corresponding to the th sample, is the full space snapshot at is the down-sampled snapshot at is the down-sampling function, is the down-sampling matrix, is the down-sampled snapshot at is the down-sampling function, is the down-sampling matrix, each element in the matrix represents the sensor status, 1 represents that there is a sensor at this position, and 0 represents that there is no sensor at this position, represents the Kronecker product.

5. The method of claim 2, wherein, The expression of the time characteristics is: ； wherein, k is a feature map sequence number, j is a timestamp sequence number, is a sampling time corresponding to the th sample, is a latent time expression corresponding to the th feature map at the sampling time, k is a Sigmoid activation function, is a down-sampling snapshot at the th sampling time, is historical sparse sensor data corresponding to the positions of all sensors at the th sampling time, is a high-dimensional convolution operator symbol, is a convolution kernel, is a bias of the k th feature map.​​ 6. The method of claim 2, wherein, The expression of the reconstructed full-space temperature field is: ； wherein, is a timestamp number, is a sampling time corresponding to the th sample, is a reconstructed full-space temperature field, is a Sigmoid activation function, is a feature map number, , is a number of feature maps, is a latent temporal expression corresponding to the th feature map at the time, is a high-dimensional convolution operator symbol, is a weight flipping operation, is a bias of each latent map in the space-time separation network.

7. The method according to claim 2, wherein, The expression of the predicted time characteristic is: ； ； wherein, is a timestamp number, is a sampling time corresponding to the th sample, is a first feature map at the time t, is a second feature map at the time t, is a th feature map at the time t, n is a th feature map at the time t, is a one-dimensional vector composed of the reconstruction vectors corresponding to the th feature map at the time t, is a prediction result of the one-dimensional vector is a nonlinear model, is a vector corresponding to the one-dimensional vector in the previous 1 second, is a vector corresponding to the one-dimensional vector in the previous seconds, is a input signal vector of the system at the time t, is a input signal vector corresponding to the previous 1 second, is a input signal vector corresponding to the previous seconds, is a time lag of the time characteristic, is a time lag of the input.

8. The method of claim 2, wherein, The expression of the distributed thermal model is: ； wherein, is a timestamp number, is a sampling time corresponding to the th sample, is a distributed thermal model of the whole space at the th time, is a Sigmoid activation function, is a feature map number, , is a number of feature maps, is th predicted temporal feature map at the th time, is a high-dimensional convolution operator symbol, is a weight flipping operation, is a bias of each latent map in the spatio-temporal synthesis network.

9. The method of claim 2, wherein, The loss function of the space-time synthesis network is: ； in, For timestamp sequence number, For the first The sampling time corresponding to each sample for A full-space snapshot of a moment. These are the parameters of the spatiotemporal synthesis network. Let be the loss function of the spatiotemporal synthesis network. The number of samples in the training set. For the reconstructed full-space temperature field.

10. A large scale lithium battery temperature field space-time modeling system, characterized in that, comprises: a data acquisition module configured to obtain space-time data of a battery temperature field, and perform normalization processing on the space-time data to obtain normalized space-time data; a model construction module configured to construct a distributed thermal model according to a space-time separation network, a space-time synthesis network and a nonlinear time dynamics unit; a data prediction module configured to input the normalized space-time data into the distributed thermal model to obtain temperature field prediction data of the lithium battery.

Images