Lithium battery pack temperature field prediction method

Abstract

The invention discloses a lithium battery pack temperature field prediction method, and relates to the technical field of batteries, and the method comprises the following steps: S100, data collection and preprocessing; s200, data vector cross coding is carried out; s300, implicit vector propagation is carried out; s400, calculating a total error; s500, training a temperature prediction neural network; and S600, predicting a temperature field. According to the lithium battery pack temperature field prediction method based on the cross attention mechanism coupling operator neural network, time-varying system iteration of a submerged space is introduced, the operation time is shortened, and errors are reduced; according to the method, physical information is introduced into the neural network, parameter optimization is carried out on the deviation of the actual physical scene partial differential equation and the deviation of the data, and the interpretability of the neural network is improved.

Description

Technical Field

[0001] The present invention relates to the field of battery technology, and in particular to a method for predicting a temperature field of a lithium battery pack. Background Art

[0002] As an important modern energy source, lithium batteries have been widely used in electric vehicles and energy storage systems because of their high energy density, long cycle life, low self-discharge, and low maintenance cost. For current lithium battery technology, due to its inherent chemical and thermal properties, studies have shown that its suitable operating temperature should be maintained between 15°C and 45°C. Otherwise, capacity decline, electrode degradation, and even potential safety issues may occur under extreme conditions. Battery temperature state (State of Temperature, SOT) is an important indicator for evaluating battery temperature changes and their impact on battery performance. As the battery usage time increases, the battery temperature state will change, which may affect the battery's charge and discharge efficiency and life. Therefore, accurate and reliable estimation of the battery temperature state is of key significance in battery management technology.

[0003] However, since lithium batteries in electric vehicles or energy storage systems are usually densely connected in series or parallel with hundreds or thousands of battery cells, the overall temperature of the battery is usually difficult to fully monitor, and the temperature changes of the battery under different working conditions and environments will affect the effect of temperature control. Therefore, accurate temperature prediction of the entire lithium battery pack is a prerequisite for battery temperature control.

[0004] Many researchers at home and abroad have conducted many studies on the temperature prediction methods of lithium batteries, such as: 1) Based on the Newman model and the Bernardi heat generation rate model, the instantaneous temperature model of lithium battery cells is established, the battery data is measured and collected, the key parameters are extracted, and then the thermal behavior of the battery cells is simulated to obtain the battery temperature prediction under different working conditions. However, this method requires a long time for simulation calculation, and it is difficult to extract the key parameters of the battery, and it is not applicable to real-time situations; 2) Based on a certain data-driven method (such as an extreme learning machine, etc.), the lithium battery equivalent reduced-order thermal model method uses the Karhunene-Loeve method to decouple and reduce the thermal process of the lithium battery, and then integrates the machine learning method to analyze the data, and finally predicts the battery temperature, but the reduction process may cause information loss, and the obtained model is not very versatile. Moreover, the above methods are all for thermal behavior prediction of lithium battery cells, which may not be applicable to the temperature field formed by the actual lithium battery pack.

[0005] Therefore, technicians in this field are committed to developing a lithium battery pack temperature field prediction method. Summary of the invention

[0006] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is how to use an operator neural network learning method to accurately predict the temperature field of a lithium battery pack.

[0007] The temperature field prediction problem of lithium battery packs is actually to solve the spatiotemporal partial differential equations about temperature conduction, which is equivalent to an operator. The physical information u of the initial conditions and boundary conditions is used as input, and the predicted temperature T(u)(d) of the corresponding point is discretized according to the data coordinate parameter vector d of the target point. The implementation of the operator uses a cross-attention mechanism to cross-encode the information. The input battery pack parameters (voltage, current, initial temperature) are sequentially passed through the encoder, processor, and decoder to finally obtain the battery pack temperature at a predetermined time step. The encoder is used to extract the high-dimensional features of the battery pack parameters and perform cross-encoding, the processor is used to iterate the features in the time step, and the decoder restores the features to the output space to obtain the battery pack temperature.

[0008] In one embodiment of the present invention, a method for predicting the temperature field of a lithium battery pack is provided, comprising the following steps:

[0009] S100, data collection and preprocessing, collecting raw data of lithium battery pack operation, discretizing the temperature field space of the lithium battery pack, randomly collecting temperature data u of sampling points at different heights, and preprocessing the raw data and temperature data u;

[0010] S200, cross-coding the data vector, creating a feature extraction encoder, cross-coding the extracted data vector, approximating the operator, performing high-dimensional mapping and screening, and obtaining a latent vector after residual connection;

[0011] S300, latent vector propagation, creating a time-varying system processor, iteratively propagating the latent vector in the latent space, with the predetermined time step being a fixed time interval until the target number of steps;

[0012] S400, total error calculation, create a decoder, map the latent vector in the potential space back to the output space, obtain a preliminary temperature prediction value, calculate the physical equation error through the partial differential equation, and add it to the prediction error to obtain the total error Loss;

[0013] S500, training the temperature prediction neural network, repeating the steps from S200 to S500, calculating the total error Loss of each round of training, back-propagating the total error Loss, optimizing the network parameters to reduce the total error Loss according to the error gradient obtained by back-propagation, until the total error Loss is less than the set target error value, stopping the training, and obtaining a trained temperature prediction neural network;

[0014] S600, temperature field prediction, using the trained temperature prediction neural network to predict the temperature at any point during the operation of the lithium battery pack.

[0015] Optionally, in the lithium battery pack temperature field prediction method in the above embodiment, step S100 includes:

[0016] S110, collecting raw data of the operation of the lithium battery pack, taking different positions of the lithium battery pack where the temperature needs to be predicted as collection points, and collecting the operation data of the lithium battery pack at a fixed frequency under predetermined charge and discharge conditions;

[0017] S120, collecting temperature data u of sampling points, and randomly collecting temperature data u of sampling points at different heights;

[0018] S130, data preprocessing, preprocessing the original data into a lithium battery pack coordinate parameter vector d, the formula is as follows:

[0019] d=[x,y,z,I t ,U t ,T am ,SOC] T

[0020] Where x, y, z are the coordinates of the temperature field space, I t is the battery current at time t, U t is the battery voltage at time t, T am is the ambient temperature of the battery, SOC is the state of charge of the battery, and t represents time;

[0021] S140, calculate the sampling vector X, pre-process the temperature data u into multiple two-dimensional matrices according to the battery pack in the height direction, each point in the matrix represents a discrete sampling temperature point of a specific height layer, perform multi-layer perceptron encoding on u, and obtain the sampling vector X containing physical information:

[0022] X=L*(x i )=L*f ω (u i ) 1≤i≤N

[0023] where f ω is a multilayer perceptron (MLP), L is the number of layers in the height direction, x i is the sampling vector on different layers, u i is the temperature data of different layers, and N is the number of temperature collection points in different layers.

[0024] Furthermore, in the lithium battery pack temperature field prediction method in the above embodiment, the lithium battery pack coordinate parameter vector d is encoded using a multi-layer perceptron and mapped into a high-dimensional query vector Y:

[0025] Y=(y i ) = f ω (d i ) 1≤i≤L

[0026] Among them, f ω is a multilayer perceptron (MLP), y i is the row vector of Y, d i are the preprocessed coordinate parameters.

[0027] Optionally, in the lithium battery pack temperature field prediction method in any of the above embodiments, the predetermined charge and discharge conditions are 2C charging and 0.5C discharging.

[0028] Optionally, in the lithium battery pack temperature field prediction method in any of the above embodiments, the fixed frequency is once per second.

[0029] Optionally, in the lithium battery pack temperature field prediction method in the above embodiment, the feature extraction encoder is based on a cross attention mechanism.

[0030] Furthermore, in the lithium battery pack temperature field prediction method in the above embodiment, the feature extraction encoder cross-fuses the high-dimensional query vector encoding the lithium battery pack coordinate parameters and the sampling vector containing physical information to obtain a latent vector that can be propagated in the latent space.

[0031] Furthermore, in the lithium battery pack temperature field prediction method in the above embodiment, step S200 includes:

[0032] S210, creating a feature extraction encoder, the feature extraction encoder including a cross attention mechanism layer, a residual connection layer and a point-to-point full connection layer;

[0033] S220, cross encoding, high-dimensional mapping and screening, create parameter matrices Wq, Wk, Wv in the cross attention mechanism layer, multiply Wq by Y to get the query vector Q, multiply Wk and Wv by X to get the key vector K and value vector V, and normalize the query vector Q and key vector K:

[0034]

[0035]

[0036] where q i and k i is the row vector of the query vector Q and the key vector K, n e is the number of columns of the query vector Q and the key vector K;

[0037] S230, calculate the hidden vector, and then obtain the hidden vector z corresponding to the query vector yi through cross multiplication i :

[0038]

[0039] Among them, q t is an input query vector, v i is the row vector of the value vector V, q t Derived from the high-dimensional query vector Y,k containing coordinate parameter information i and v i It comes from the sampling vector X containing physical information. The operator is approximated by cross-multiplication to obtain the implicit vector Z containing the temperature field information of the lithium battery pack. i .

[0040] S240, calculate the actual latent vector. Since the X input has L layers, the feature output corresponds to only one layer. Add all L layers and add the query vector q. t The residual input is q t The corresponding actual latent vector z t :

[0041]

[0042] S250, calculate the hidden vector output, after the cross attention mechanism layer, in order to make the output feature consistent with the input space, then input q t Add a fully connected layer to form a residual connection layer to get the hidden vector output

[0043]

[0044] Optionally, in the lithium battery pack temperature field prediction method in any of the above embodiments, the time-varying system processor is based on a point-type multilayer perceptron.

[0045] Optionally, in the lithium battery pack temperature field prediction method in any of the above embodiments, step S300 includes:

[0046] S310, creating a time-varying system processor;

[0047] S320, hidden vector output propagation, hidden vector Propagation in latent space;

[0048] S330, time iteration, iterate the time to the target number of steps and then output:

[0049]

[0050] After iterative propagation, the target number of steps is reached to obtain the latent vector for output.

[0051] Preferably, in the lithium battery pack temperature field prediction method in the above embodiment, the target number of steps is 5 steps.

[0052] Preferably, in the lithium battery pack temperature field prediction method in the above embodiment, the fixed time interval is 1 second.

[0053] Optionally, in the lithium battery pack temperature field prediction method in any of the above embodiments, the operating data of the lithium battery pack includes voltage, current, and state of charge.

[0054] Optionally, in the lithium battery pack temperature field prediction method in any of the above embodiments, step S400 includes:

[0055] S410, creating a decoder, the decoder is based on a self-attention mechanism;

[0056] S420, calculate the preliminary temperature prediction value, Use the parameter matrix Multiply them to get the query matrix Key Matrix Sum Matrix The parameter matrix used here does not share parameters with those used in S220. The output vector H is calculated by the Softmax function, and then the predicted temperature output is obtained by normalizing the residual connection layer. The formula is as follows:

[0057]

[0058]

[0059] in, To predict the temperature at the query point, FFN is a fully connected layer and does not share parameters with the FFN after the cross attention mechanism layer;

[0060] S430, calculating the physical equation error, the physical equation error L of the network parameter optimization phy is the partial differential equation error, expressed as follows:

[0061]

[0062]

[0063] Among them, ρ, C p , k are the battery density, specific heat capacity and average thermal conductivity in all directions, Q batt is the equivalent heat generated during battery operation, V battis the volume of the temperature field area, E oc is the open circuit voltage of the battery;

[0064] S440, calculate the total error, and obtain the data error L by subtracting the predicted temperature value from the actual temperature value. data , and L phy Add them together to get the total error Loss.

[0065] Optionally, in the lithium battery pack temperature field prediction method in any of the above embodiments, step S500 includes:

[0066] S510, calculating the total error of back propagation, for the total error Loss obtained, through back propagation, the loss value is distributed to each network layer of the neural network encoder, processor, and decoder for calculating the error gradient;

[0067] S520, calculating the error gradient, using the chain rule, taking the partial derivative of the loss value of each network layer of the neural network with respect to the parameters in the neural network, and obtaining the error gradient of each network layer;

[0068] S530, optimize the total error. For the error gradient obtained by back propagation, use the stochastic gradient descent method to optimize the parameters of the neural network, and then repeat steps S200 to S500 until the total error Loss is lower than the set value 1*e -5 So far, a trained temperature prediction neural network is obtained.

[0069] The present invention is based on a lithium battery pack temperature field prediction method of a cross-attention mechanism coupled operator neural network. Through neural network training, the existing temperature information and predicted temperature are used, and the speed is improved compared to the finite element simulation method, and the real-time performance is greatly enhanced. The present invention introduces a time-varying system iteration in the latent space, and iterates on multiple time steps on the basis of only one feature extraction, thereby reducing the calculation time and the error. The present invention introduces physical information into the neural network, and jointly optimizes the parameters of the deviation of the partial differential equation of the actual physical scene and the deviation of the data, thereby improving the interpretability of the neural network.

[0070] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 is a flow chart illustrating a lithium battery pack temperature field prediction method according to an exemplary embodiment. DETAILED DESCRIPTION

[0072] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0073] In the drawings, components with the same structure are indicated by the same numerical reference numerals, and components with similar structures or functions are indicated by similar numerical reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. In order to make the illustration clearer, the thickness of the components is schematically exaggerated in some places in the drawings.

[0074] The inventor provides a method for predicting the temperature field of a lithium battery pack. Figure 1 As shown, the following steps are included:

[0075] S100, data collection and preprocessing, collecting the original data of the lithium battery pack operation, discretizing the temperature field space of the lithium battery pack, randomly collecting the temperature data u of the sampling points at different heights, and preprocessing the original data and temperature data u; including:

[0076] S110, collecting raw data of the lithium battery pack operation, taking different locations in the lithium battery pack where the temperature needs to be predicted as collection points, and collecting the operating data of the lithium battery pack at a fixed frequency under the predetermined charge and discharge conditions, including voltage, current, state of charge, and the predetermined charge and discharge conditions.

[0077] The working condition is 2C charging and 0.5C discharging, and the fixed frequency is once per second;

[0078] S120, collecting temperature data u of sampling points, and randomly collecting temperature data u of sampling points at different heights;

[0079] S130, data preprocessing, preprocessing the original data into a lithium battery pack coordinate parameter vector d, the formula is as follows:

[0080] d=[x,y,z,I t ,U t ,T am ,SOC] T

[0081] Where x, y, z are the coordinates of the temperature field space, I t is the battery current at time t, U t is the battery voltage at time t, T am is the ambient temperature of the battery, SOC is the state of charge of the battery, and t represents time. The lithium battery pack coordinate parameter vector d is encoded using a multilayer perceptron and mapped to a high-dimensional query vector Y:

[0082] Y=(yi ) = f ω (d i ) 1≤i≤L

[0083] Among them, f ω is a multilayer perceptron (MLP), y i is the row vector of Y, d i are the preprocessed coordinate parameters.

[0084] S140, calculate the sampling vector X, pre-process the temperature data u into multiple two-dimensional matrices according to the battery pack in the height direction, each point in the matrix represents a discrete sampling temperature point of a specific height layer, perform multi-layer perceptron encoding on u, and obtain a sampling vector containing physical information

[0085] Quantity X:

[0086] X=L*(x i )=L*f ω (u i ) 1≤i≤N

[0087] where f ω is a multilayer perceptron (MLP), L is the number of layers in the height direction, x i is the sampling vector on different layers, u i is the temperature data of different layers, and N is the number of temperature collection points in different layers.

[0088] S200, cross-coding of data vectors, creating a feature extraction encoder, cross-coding the extracted data vectors based on the cross-attention mechanism, approximating the operators, performing high-dimensional mapping and screening, and obtaining a latent vector after residual connection, that is, the feature extraction encoder cross-fuses the high-dimensional query vector encoding the coordinate parameters of the lithium battery pack and the sampling vector containing physical information to obtain a latent vector that can be propagated in the latent space; specifically including:

[0089] S210, create a feature extraction encoder, the feature extraction encoder includes a cross attention

[0090] A force mechanism layer, a residual connection layer, and a point-wise fully connected layer;

[0091] S220, cross encoding, high-dimensional mapping and screening, create parameter matrices Wq, Wk, Wv in the cross attention mechanism layer, multiply Wq with Y to get the query vector Q, multiply Wk and Wv with X to get the key vector K and value vector V, respectively, and perform cross encoding on the query vector Q and key vector K.

[0092] Vector K is normalized:

[0093]

[0094]

[0095] where q i and k i is the row vector of the query vector Q and the key vector K, n e is the number of columns of the query vector Q and the key vector K;

[0096] S230, calculate the latent vector, and then obtain the latent vector corresponding to the query vector yi through cross multiplication

[0097] Quantity i :

[0098]

[0099] Among them, q t is an input query vector, v i is the row vector of the value vector V, q t Derived from the high-dimensional query vector Y,k containing coordinate parameter information i and v i It comes from the sampling vector X containing physical information. The operator is approximated by cross-multiplication to obtain the implicit vector z containing the temperature field information of the lithium battery pack. i .

[0100] S240, calculate the actual latent vector. Since the X input has L layers, the feature output corresponds to only one layer. Add all L layers and add the query vector q. t The residual input is

[0101] To q t The corresponding actual latent vector z t :

[0102]

[0103] S250, calculate the hidden vector output, after the cross attention mechanism layer, in order to make the output feature consistent with the input space, then input q t Add a fully connected layer (FFN) to form a residual connection layer to get the hidden vector output

[0104]

[0105] S300, hidden vector propagation, creating a time-varying system processor, iteratively propagating the hidden vector in the latent space, with a predetermined time step of a fixed time interval of 1 second until the target number of steps is 5; specifically including:

[0106] S310, creating a time-varying system processor;

[0107] S320, hidden vector output propagation, hidden vector Propagation in latent space;

[0108] S330, time iteration, iterate the time to the target number of steps and then output:

[0109]

[0110] After iterative propagation, the target number of steps is reached to obtain the latent vector for output.

[0111] S400, total error calculation, create a decoder, map the latent vector in the potential space back to the output space, obtain a preliminary temperature prediction value, calculate the physical equation error through the partial differential equation, and add it to the prediction error to obtain the total error Loss; specifically including:

[0112] S410, creating a decoder, the decoder is based on a self-attention mechanism;

[0113] S420, calculate the preliminary temperature prediction value, Use the parameter matrix Multiply them to get the query matrix Key Matrix Sum Matrix The parameter matrix used here does not share parameters with those used in S220. The output vector H is calculated by the Softmax function, and then the predicted temperature output is obtained by normalizing the residual connection layer. The formula is as follows:

[0114]

[0115]

[0116] in, is the predicted temperature at the query point, FFN is the fully connected layer, and

[0117] FFNs with cross-attention layers do not share parameters;

[0118] S430, calculating the physical equation error, the physical equation error L of the network parameter optimization phy for

[0119] The partial differential equation error is expressed as follows:

[0120]

[0121]

[0122] Among them, ρ, C p , k are the battery density, specific heat capacity and average thermal conductivity in all directions, Q batt is the equivalent heat generated during battery operation, V batt Temperature field

[0123] The volume of the domain, E oc is the open circuit voltage of the battery;

[0124] S440, calculate the total error, and obtain the data error by subtracting the predicted temperature value from the actual temperature value

[0125] L data , and L phy Add them together to get the total error Loss.

[0126] S500, training of the temperature prediction neural network, repeating the steps from S200 to S500, calculating the total error Loss of each round of training, back-propagating the total error Loss, optimizing the network parameters to reduce the total error Loss according to the error gradient obtained by back-propagation, until the total error Loss is less than the set target error value, stopping the training, and obtaining a trained temperature prediction neural network; specifically including:

[0127] S510, calculating the total error of back propagation, for the total error Loss obtained, through back propagation, the loss value is distributed to each network layer of the neural network encoder, processor, and decoder for calculating the error gradient;

[0128] S520, calculating the error gradient, using the chain rule, taking the partial derivative of the loss value of each network layer of the neural network with respect to the parameters in the neural network, and obtaining the error gradient of each network layer;

[0129] S530, optimize the total error. For the error gradient obtained by back propagation, use the stochastic gradient descent method to optimize the parameters of the neural network, and then repeat steps S200 to S500 until the total error Loss is lower than the set value 1*e -5 So far, a trained temperature prediction neural network is obtained.

[0130] S600, temperature field prediction, using the trained temperature prediction neural network to predict the temperature at any point during the operation of the lithium battery pack.

[0131] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for predicting the temperature field of a lithium battery pack, characterized in that: The steps include: S100, data collection and preprocessing, collecting raw data of the lithium battery pack operation, discretizing the temperature field space of the lithium battery pack, randomly collecting temperature data u of sampling points at different heights, and preprocessing the raw data and the temperature data u; S200, cross-coding the data vector, creating a feature extraction encoder, cross-coding the extracted data vector, approximating the operator, performing high-dimensional mapping and screening, and obtaining a latent vector after residual connection; S300, latent vector propagation, creating a time-varying system processor, iteratively propagating the latent vector in the latent space, with the predetermined time step being a fixed time interval until the target number of steps; S400, total error calculation, creating a decoder, mapping the latent vector in the latent space back to the output space, obtaining a preliminary temperature prediction value, calculating the physical equation error through a partial differential equation, and adding the error to the prediction error to obtain a total error Loss; S500, temperature prediction neural network training, repeating the steps S200 to S500, calculating the total error Loss of each round of training, back-propagating the total error Loss, optimizing network parameters to reduce the total error Loss according to the error gradient obtained by the back-propagation, until the total error Loss is less than the set target error value, stopping the training, and obtaining a trained temperature prediction neural network; S600, temperature field prediction: using the trained temperature prediction neural network to predict the temperature of any point during the operation of the lithium battery pack.

2. The lithium battery pack temperature field prediction method according to claim 1, characterized in that: The step S100 includes: S110, collecting raw data of the operation of the lithium battery pack, taking different positions of the lithium battery pack where the temperature needs to be predicted as collection points, and collecting the operation data of the lithium battery pack at a fixed frequency under a predetermined charge and discharge condition; S120, collecting temperature data u of sampling points, and randomly collecting temperature data u of sampling points at different heights; S130, data preprocessing, preprocessing the raw data into a lithium battery pack coordinate parameter vector d, the formula is as follows: d=[x,y,z,I t ,U t ,T am ,SOC] T ; S140, calculate the sampling vector X, pre-process the temperature data u into multiple two-dimensional matrices according to the battery pack in the height direction, each point in the matrix represents a discrete sampling temperature point of a specific height layer, perform multi-layer perceptron encoding on u, and obtain the sampling vector X containing physical information: X=L*(x i )=L*f ω (u i ) 1≤i≤N 。 3. The lithium battery pack temperature field prediction method according to claim 2, characterized in that: The lithium battery pack coordinate parameter vector d is encoded using a multi-layer perceptron and mapped into a high-dimensional query vector Y: Y=(y i )=f ω (d i ) 1≤i≤L 。 4. The lithium battery pack temperature field prediction method according to claim 1, characterized in that: The predetermined charge and discharge condition is 2C charging and 0.5C discharging.

5. The lithium battery pack temperature field prediction method according to claim 2, characterized in that: The feature extraction encoder is based on a criss-cross attention mechanism.

6. The lithium battery pack temperature field prediction method according to claim 5, characterized in that: The step S200 includes: S210, creating a feature extraction encoder, wherein the feature extraction encoder includes a cross attention mechanism layer, a residual connection layer, and a point-to-point full connection layer; S220, cross encoding, high-dimensional mapping and screening, create parameter matrices Wq, Wk, Wv in the cross attention mechanism layer, multiply Wq by Y to obtain the query vector Q, multiply Wk and Wv by X to obtain the key vector K and the value vector V, and normalize the query vector Q and the key vector K: S230, calculate the hidden vector, and then obtain the hidden vector z corresponding to the query vector yi through cross multiplication i : S240, calculate the actual latent vector. Since the X input has L layers, the feature output corresponds to only one layer. Add all L layers and add the query vector q. t The residual input is q t The corresponding actual latent vector z t : S250, calculate the latent vector output, and after passing through the cross attention mechanism layer, in order to make the output feature consistent with the input space, then input q t And a fully connected layer is added to form the residual connection layer to obtain the hidden vector output 7. The lithium battery pack temperature field prediction method according to claim 1, characterized in that: The time-varying system processor is based on a point-wise multilayer perceptron.

8. The lithium battery pack temperature field prediction method according to claim 6, characterized in that: The step S300 includes: S310, creating a time-varying system processor; S320, hidden vector output propagation, the hidden vector Propagation in latent space; S330, time iteration, iterate the time to the target number of steps and then output: After iterative propagation, the target number of steps is reached to obtain the latent vector for output.

9. The lithium battery pack temperature field prediction method according to claim 8, characterized in that: The step S400 includes: S410, creating a decoder, wherein the decoder is based on a self-attention mechanism; S420, calculate the preliminary temperature prediction value, Use the parameter matrix Multiply them to get the query matrix Key Matrix Sum Matrix The output vector H is calculated by the Softmax function, and then normalized by the residual connection to obtain the predicted temperature output. The formula is as follows: S430, calculating the physical equation error, the physical equation error L of the network parameter optimization phy is the partial differential equation error, expressed as follows: S440, calculate the total error, and obtain the data error L by subtracting the predicted temperature value from the actual temperature value. data , and L phy Add them together to get the total error Loss.

10. The lithium battery pack temperature field prediction method according to claim 1, characterized in that: The step S500 includes: S510, calculating the total back-propagation error, and for the obtained total error Loss, the loss value is distributed to each network layer of the neural network encoder, processor, and decoder through back-propagation; S520, calculating the error gradient, using the chain rule, taking the partial derivative of the loss value of each network layer of the neural network with respect to the parameters in the neural network, and obtaining the error gradient of each network layer; S530, optimizing the total error, using the stochastic gradient descent method to optimize the parameters of the neural network for the error gradient obtained by back propagation, and then repeating steps S200 to S500 until the total error Loss is lower than the set value 1*e -5 So far, a trained temperature prediction neural network is obtained.

Images