A battery module temperature prediction method based on physical information neural network and SSA-ELM

Abstract

The present invention discloses a battery module temperature prediction method based on physical information neural network and SSA‑ELM. The mutual information method is used to calculate the estimated value of the mutual information between the characteristic factors and the battery module temperature, and the types of characteristic factors screened by setting thresholds are placed in a preset set; the SSA‑ELM model is trained by the historical data of the characteristic factors in the preset set, and the predicted value of the battery module temperature at a certain time is output; PINN uses the battery module thermal equation as a physical condition constraint, and takes the predicted value and parameter value of the battery module temperature as input, and obtains the convergence value of the battery module temperature by updating the variables until convergence. The present invention trains the SSA‑ELM model with a small-scale sample to reduce the cost of collecting data and speed up the prediction speed. The loss function is introduced into the thermal balance equation using PINN, and the loss function is continuously reduced until convergence by updating the convergence value of the battery module temperature, thereby improving the prediction accuracy.

Description

Technical field:

[0001] The present invention belongs to the technical field of neural network deep learning, and in particular relates to a battery module temperature prediction method based on physical information neural network and SSA-ELM. Background technology:

[0002] With the development of science and technology, batteries have been widely used in industry, among which lithium battery packs are the most common. In order to meet the needs of reliability, efficiency and safety, lithium-ion battery technology needs to be further developed. The performance of batteries is affected by temperature changes. Too high or too low temperature will affect the chemical reaction inside the battery, thereby affecting the operation of industrial devices. In order to solve these problems, scientific methods are used to predict the internal temperature of the battery module at a certain time, so as to maintain good battery performance and ensure stable industrial development.

[0003] Since the concept of ELM (extreme learning machine) was proposed, it has been welcomed by many researchers. It subverts the traditional neural network based on gradient calculation, combines the knowledge of matrix theory with the least squares method, and has a fast calculation speed. However, the parameters of the input layer and hidden layer of ELM are randomly generated, and this randomness will affect the accuracy of prediction.

[0004] The biggest feature of the PINN physical information neural network is that it adds prior knowledge to the neural network layer. By adding the residual of the physical equation as a loss term to the loss function, it is obvious that most cases that satisfy the laws of physics are real values. This feature can be used to give only network inputs without providing outputs, allowing the network to learn these input parameters until it converges to a value. However, the introduction of the residual term in the PINN physical information neural network will consume a lot of prediction time.

[0005] Patent ZL202211119179.3 discloses "a battery temperature prediction method, device, vehicle and storage medium". This method estimates the future predicted current of the battery by distinguishing the vehicle's operating scenarios, thereby improving the accuracy of the battery's future current prediction. Furthermore, based on the predicted current at the target time, the predicted temperature value at the target time is predicted, thereby improving the accuracy of the future temperature prediction. However, there is no applicable criterion for judging the predicted temperature.

[0006] Patent ZL202211198217.9 discloses "a temperature prediction method, system, device and storage medium for lithium batteries". This method does not require equation discretization and gridding, has good portability, and is easier to modify the target temperature equation. However, it takes a certain amount of time to train multiple loss functions together.

[0007] Therefore, there is an urgent need for a prediction model that combines data-driven and physical-driven methods to accurately predict the battery module temperature at a certain time, while speeding up the prediction process and reducing the cost of data.

[0008] The information disclosed in this background technology section is only intended to enhance the understanding of the overall background of the invention and should not be regarded as an acknowledgment or any form of suggestion that the information constitutes the prior art already known to a person skilled in the art. Summary of the invention:

[0009] The purpose of the present invention is to provide a battery module temperature prediction method based on physical information neural network and SSA-ELM. The main prediction target is the battery module temperature. The SSA algorithm is used to optimize the weights and thresholds of the input layer and the hidden layer in the ELM model. Combining ELM with PINN can not only improve the prediction accuracy, but also neutralize the prediction time, thereby overcoming the defects in the above-mentioned prior art.

[0010] To achieve the above object, the present invention provides a battery module temperature prediction method based on physical information neural network and SSA-ELM, the steps of which are:

[0011] S01. Select a battery module, and calculate the mutual information estimation value of each characteristic factor and the battery module temperature respectively by using the mutual information method according to the characteristic factor historical data. The magnitude of the mutual information estimation value reflects the correlation between the characteristic factor and the battery module temperature. By setting a threshold for the mutual information estimation value, the types of characteristic factors whose mutual information estimation value exceeds the threshold are put into a preset set; the characteristic factors include chamber temperature, voltage, current, and state of charge;

[0012] S02. The characteristic factor historical data and the corresponding battery module temperature historical data are combined into a pre-training data set, wherein the characteristic factor historical data and the corresponding battery module temperature historical data are data before time t. The SSA algorithm is used to optimize the ELM model to obtain the SSA-ELM model. The pre-training data set is brought into the SSA-ELM model for training. The training content is to obtain the neuron weights and thresholds in the SSA-ELM model by the matrix inversion method. After the training is completed, the characteristic factors in the preset set are input into the SSA-ELM model, and the SSA-ELM model outputs the battery module temperature prediction value Tem t ;

[0013] S03. Construct a physical information neural network PINN. The physical information neural network PINN uses the residual of the thermal balance equation of the battery module as a loss function. There are two unknown variables in the loss function, namely, the temperature convergence value of the battery module and the temperature variation of the battery module over time. The parameter values ​​in the loss function are measured and the predicted value of the battery module temperature Tem is used. tAs the initial value of the battery module temperature convergence value, the initial value of the battery module temperature convergence value and the measured parameter value are input into the loss function to obtain the variable of the battery module temperature changing with time, and then the variable of the battery module temperature changing with time is input into the loss function. In this process, the result of the battery module temperature changing with time is generated, and the convergence value of the battery module temperature at that time is obtained; by updating the battery module temperature convergence value and the variable of the battery module temperature changing with time, the value of the loss function is continuously reduced until the loss function converges, and finally the optimal solution of the battery module temperature convergence value at that time is obtained.

[0014] Preferably, in the technical solution, in step S01, the mutual information method represents the correlation between the battery module temperature and the characteristic factor; the characteristic factor is represented as in represents i (i=1, 2, 3...N) characteristic factors at time t, where N represents the number of characteristic factors;

[0015] Since the multiple historical data of each characteristic factor are in different quantity ranges, each characteristic factor historical data is normalized, and the data normalization is expressed as:

[0016]

[0017] in represents the maximum value among the i characteristic factors at time t, represents the minimum value of the i-th characteristic factor at time t, represents the average value of i characteristic factors at time t; represents the i characteristic factors after normalization;

[0018] At time tL (L>0), calculate the battery module temperature history data Ta t-L Estimated mutual information with each normalized feature factor The expression of mutual information is expressed as:

[0019]

[0020] Where x represents the variable distribution of the battery module temperature, y represents the variable distribution of the characteristic factor, p(x) and p(y) are the marginal probability distribution functions of x and y respectively, and p(x, y) is the joint probability distribution function of x and y. The mutual information estimation value in formula (2) is placed in a set Q1,

[0021] A threshold β is set, and the mutual information estimation value in Q1 is compared with the threshold β. The mutual information estimation value in Q1 that is greater than the threshold β is placed in the preset set Q2. Since the mutual information estimation value affects the prediction of the battery module temperature, and the mutual information estimation value is positively correlated with the influence of the characteristic factor on the battery module temperature, the characteristic factors with the mutual information estimation value above the threshold are selected. The preset set Q2 is expressed as:

[0022]

[0023] Preferably, in the technical solution, in step S02, the ELM model is a single hidden layer neural network, and the network structure consists of an input layer of d neurons, a hidden layer of m neurons, and an output layer of v neurons, where v is 1, representing the output battery module temperature prediction value, wherein the length d of the input preset set Q2 corresponds to d neurons; the weights and thresholds of the input layer and the hidden layer are randomly generated during the training process, and the weight matrix w of the input layer j =[w1 w2 … w d ] T , the threshold matrix b of the input layer k =[b1 b2 … b d ] T , the weight matrix c of the hidden layer g =[c1 c2 … c m ] T , the vector matrix out output by the output layer index =[out1 out2 … out v ] T , where T represents the transpose operation of the matrix, w d represents the weight of the jth (j=1,2,3…d) neuron in the input layer, b d represents the threshold of the kth (k=1,2,3…d)th neuron in the input layer, c m represents the weight of the gth (g=1,2,3...m)th neuron in the hidden layer, out v Represents the output of the index(index=1,2,3……v)th neuron in the output layer.

[0024] Preferably, in the technical solution, in step S02, the SSA algorithm first transforms the weight matrix w of the input layer into j With the threshold matrix bk of the input layer = [b1 b2 … b d ] T Concatenate into a weight threshold matrix Link d*2 ,

[0025]

[0026] In the weight threshold matrix of formula (4), each row represents the weight and threshold of a neuron. The fitness value of each neuron is calculated respectively. The fitness value represents the quality of the data. The higher the fitness value, the more conducive the data generated by the neuron is to training. The calculation process of the fitness value is expressed as:

[0027]

[0028] ADP good Represents a matrix composed of fitness values, adp(w d , b d ) indicates that each row of the matrix represents the fitness value of different neurons, and adp(·) indicates the mapping function for calculating the fitness value;

[0029] The SSA algorithm includes explorers, trackers, and scouts, where the explorer iteration process is expressed as:

[0030]

[0031] The Link matrix is ​​a weight threshold matrix with d rows and 2 columns. represents the data of the pth dimension in the oth row in the cntth iteration, p=1 represents the weight, and p=2 represents the threshold; iter max represents the preset maximum number of iterations, α is a random number in the interval [0,1]; R2 represents the warning value, ST represents the data safety value; ram is a random number that obeys the normal distribution, and L is a 1*d matrix of all 1s;

[0032] Based on the explorer's optimization, the iterative process of the tracker is expressed as:

[0033]

[0034] in represents the worst data in the cnt-th iteration, represents the best position of the tracker at the cnt+1th time, A represents a 1*d matrix, and the values ​​in the matrix are randomly selected from -1 or 1, and re represents the number of trackers;

[0035] Based on the optimization of the explorer and the tracker, the weight threshold matrix is ​​further optimized, and the iterative process of the scout is expressed as:

[0036]

[0037] in represents the global optimal position at the cnt-th iteration, ρ represents a random number that follows a normal distribution with a mean of 0 and a variance of 1, and f nowrepresents the fitness value of the current individual, f minf Represents the minimum fitness value among all individuals, f maxf represents the maximum fitness value among all individuals, ε represents a non-zero minimum, and μ represents a random number in the range of [-1,1];

[0038] Through multiple iterations of explorers, trackers, and scouts, the weights and thresholds of each neuron are continuously updated in a direction suitable for training, and the updated results are returned to Link d*2 In the matrix.

[0039] Preferably, in the technical solution, after the weights and thresholds of the input layer and the hidden layer of the SSA-ELM model are optimized, when the weights and thresholds of each neuron in the input layer and the hidden layer are known, the preset set Q2 is used as the input of the SSA-ELM model. The output of the output layer of the SSA-ELM model is expressed as:

[0040]

[0041] Where SC represents the output value of the output layer, c lr represents the weight of the hidden layer output, w lr represents the weight of the input layer, b lr represents the threshold of the input layer, lr = 1, 2, 3 ... m, input = 1, 2, 3 ... d, Q 2，input represents the input feature factor in the preset set Q2; result represents a function, which is expressed as:

[0042] result(w lr , b lr , Q 2，input )=Sigmoid(w lr Q 2，input +b lr ) (10);

[0043] Where Sigmoid(·) represents the activation function;

[0044] The result(w lr , b lr , Q 2，input ) function as a whole H, H is regarded as a hidden layer output coefficient matrix, then formula (9) is simplified to:

[0045]

[0046] The H matrix is ​​represented as:

[0047]

[0048] By using the matrix inversion method, we can solve the problem of c in equation (11) lr Solving for:

[0049] c lr =SC·H + (13);

[0050] Among them, H + is the pseudo-inverse form of the H matrix; respectively, the obtained c lr Substitute it into the SSA-ELM model to find the predicted value of the battery module temperature Tem at time t t .

[0051] Preferably, in the technical solution, in step S03, the physical information neural network PINN physically constrains the temperature of the battery module through the thermal balance equation of the battery module, so that the temperature of the battery module follows the laws of physics, and the introduction of the thermal balance equation injects physical prior knowledge in the battery module into the physical information neural network PINN, and the thermal balance equation of the battery module is:

[0052]

[0053] Where mass is the mass of the battery module, cap is the heat capacity of the battery module, T n Indicates the temperature of the outer metal shell of the battery module, T e Indicates the temperature of the metal shell inside the battery module, R c，n Represents the resistance value of the outer metal shell of the battery module, R c，e represents the internal metal shell resistance value, t represents time, ω represents the convection heat transfer coefficient of the battery module, area represents the cross-sectional area of ​​the battery module, T r Indicates the ambient temperature of the battery module, and heat indicates the heat generated by the battery module;

[0054] The Loss loss function is introduced, and the residual of the heat balance equation is regarded as the loss term, which is expressed as:

[0055]

[0056] Formula (15) is the loss function composed of the residual of the thermal balance equation of the battery module. The unknown variables in the loss function include the variables of the battery module temperature changing with time: Battery module temperature convergence value T, is a function of T with respect to time; measurement parameters are mass, cap, heat, T n , R c，n , T e , R c，e ,ω,area,T r The data obtained in step S02 is tAs the initial value of T, the initial value of T and the parameters mass, cap, heat, T n , R c，n , T e , R c，e ,ω,area,T r The data is input into the loss function to obtain Then the obtained Input the loss function and find T at time t; in the training process of the physical information neural network PINN, as With the continuous updating of and T, the loss function gradually decreases, and the optimal solution of T at time t is solved.

[0057] Preferably, in the technical solution, in step S03, the battery module temperature convergence value T obtained by the physical information neural network PINN is detected, and a threshold δ is set to determine the battery module temperature prediction value Tem output by the SSA-ELM model. t The difference between the predicted value Tem of the battery module temperature and the convergence value T of the battery module temperature output by the physical information neural network PINN is expressed as follows:

[0058] |T-Tem t |<δ(16);

[0059] When the battery module temperature prediction value Tem t If the battery module temperature convergence value T satisfies the condition of formula (16), the battery module temperature convergence value T is used as the result of the battery module temperature to be predicted at time t; if the condition of formula (16) is not satisfied, steps S01-S03 are repeated.

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] 1. The ELM model used in the present invention is composed of a single hidden layer neural network, and unlike the traditional method of using gradient to calculate parameters, the output layer parameters are calculated by matrix inversion, which can significantly reduce the prediction time;

[0062] 2. The SSA algorithm can optimize the process of randomly generating parameters in the input layer and hidden layer of the ELM model, and improve the accuracy of prediction by iteratively calculating the weight and threshold of each neuron. The output layer weight of the SSA-ELM model is obtained by matrix inversion, so it has the characteristics of fast solution. The SSA algorithm optimizes the weight and threshold, and improves the uncertainty caused by the randomly generated weight and threshold. The SSA-ELM model uses small-scale samples for training, which reduces the cost of data collection.

[0063] 3. Using the physical information neural network PINN, the physical equation that satisfies the internal thermal balance of the battery module can be introduced into the loss function, and the loss function can be continuously reduced until convergence by updating the battery module temperature convergence value T. PINN can judge whether the temperature predicted by the SSA-ELM model is accurate through indirect supervision, which can reduce the cost of data training. Description of the drawings:

[0064] Figure 1 It is a flow chart of a battery module temperature prediction method based on physical information neural network and SSA-ELM of the present invention;

[0065] Figure 2 This is a neural network architecture diagram of the SSA-ELM model of the present invention;

[0066] Figure 3 It is a flow chart of the physical information neural network PINN of the present invention. Specific implementation method:

[0067] The specific embodiments of the present invention are described in detail below, but it should be understood that the protection scope of the present invention is not limited by the specific embodiments.

[0068] Unless explicitly stated otherwise, throughout the specification and claims, the term “comprise” or variations such as “include” or “comprising”, etc., will be understood to include the stated elements or components but not to exclude other elements or components.

[0069] like Figure 1 As shown, the present invention provides a battery module temperature prediction method based on physical information neural network and SSA-ELM, the steps of which are:

[0070] S01. Select a battery module, and use the mutual information method to calculate the mutual information estimation value of each characteristic factor and the battery module temperature according to the historical data of the characteristic factors. The size of the mutual information estimation value reflects the correlation between the characteristic factor and the battery module temperature. By setting a threshold for the mutual information estimation value, the types of characteristic factors whose mutual information estimation value exceeds the threshold are placed in a preset set; the characteristic factors include chamber temperature, voltage, current, and charge state.

[0071] The mutual information method represents the correlation between the battery module temperature and the characteristic factors; the characteristic factors are expressed as in represents i (i=1, 2, 3...N) characteristic factors at time t, where N represents the number of characteristic factors;

[0072] Since the multiple historical data of each characteristic factor are in different quantity ranges, each characteristic factor historical data is normalized, and the data normalization is expressed as:

[0073]

[0074] in represents the maximum value among the i characteristic factors at time t, represents the minimum value of the i-th characteristic factor at time t, represents the average value of i characteristic factors at time t; represents the i characteristic factors after normalization;

[0075] At time tL (L>0), calculate the battery module temperature history data Ta t-L Estimated mutual information with each normalized feature factor The expression of mutual information is expressed as:

[0076]

[0077] Where x represents the variable distribution of the battery module temperature, y represents the variable distribution of the characteristic factor, p(x) and p(y) are the marginal probability distribution functions of x and y respectively, and p(x, y) is the joint probability distribution function of x and y. The mutual information estimation value in formula (2) is placed in a set Q1,

[0078] A threshold β is set, and the mutual information estimation value in Q1 is compared with the threshold β. The mutual information estimation value in Q1 that is greater than the threshold β is placed in the preset set Q2. Since the mutual information estimation value affects the prediction of the battery module temperature, and the mutual information estimation value is positively correlated with the influence of the characteristic factor on the battery module temperature, the characteristic factors with the mutual information estimation value above the threshold are selected. The preset set Q2 is expressed as:

[0079]

[0080] S02, such as Figure 2As shown, the characteristic factor historical data and the corresponding battery module temperature historical data are combined into a pre-training data set, wherein the characteristic factor historical data and the corresponding battery module temperature historical data are data before time t, and the SSA algorithm is used to optimize the ELM model to obtain the SSA-ELM model, and the pre-training data set is brought into the SSA-ELM model for training. The training content is to obtain the neuron weights and thresholds in the SSA-ELM model by the matrix inversion method. After the training is completed, the characteristic factors in the preset set are input into the SSA-ELM model, and the SSA-ELM model outputs the battery module temperature prediction value Tem t ;

[0081] The ELM model is a single hidden layer neural network. The network structure consists of an input layer of d neurons, a hidden layer of m neurons, and an output layer of v neurons. Here, v is 1, representing the output battery module temperature prediction value. The length d of the input preset set Q2 corresponds to d neurons. The weights and thresholds of the input layer and the hidden layer are randomly generated during the training process. The weight matrix w of the input layer j =[w1 w2 … w d ] T , the threshold matrix b of the input layer k =[b1 b2 … b d ] T , the weight matrix c of the hidden layer g =[c1 c2 … c m ] T , the vector matrix out output by the output layer index =[out1 out2 … out v ] T , where T represents the transpose operation of the matrix, w d represents the weight of the jth (j=1,2,3…d) neuron in the input layer, b d represents the threshold of the kth (k=1,2,3…d)th neuron in the input layer, c m represents the weight of the gth (g=1,2,3...m)th neuron in the hidden layer, out v Represents the output of the index(index=1,2,3……v)th neuron in the output layer.

[0082] In the process of optimizing the ELM model, the SSA algorithm first converts the weight matrix w of the input layer into j and the threshold matrix b of the input layer k =[b1 b2 … b d ] T Concatenate into a weight threshold matrix Link d*2 ,

[0083]

[0084] In the weight threshold matrix of formula (4), each row represents the weight and threshold of a neuron. The fitness value of each neuron is calculated respectively. The fitness value represents the quality of the data. The higher the fitness value, the more conducive the data generated by the neuron is to training. The calculation process of the fitness value is expressed as:

[0085]

[0086] ADP good Represents a matrix composed of fitness values, adp(w d , b d ) indicates that each row of the matrix represents the fitness value of different neurons, and adp(·) indicates the mapping function for calculating the fitness value;

[0087] The SSA algorithm includes explorers, trackers, and scouts, where the explorer iteration process is expressed as:

[0088]

[0089] The Link matrix is ​​a weight threshold matrix with d rows and 2 columns. represents the data of the pth dimension in the oth row in the cntth iteration, p=1 represents the weight, and p=2 represents the threshold; iter max represents the preset maximum number of iterations, α is a random number in the interval [0,1]; R2 represents the warning value, ST represents the data safety value; ram is a random number that obeys the normal distribution, and L is a 1*d matrix of all 1s;

[0090] Based on the explorer's optimization, the iterative process of the tracker is expressed as:

[0091]

[0092] in represents the worst data in the cnt-th iteration, represents the best position of the tracker at the cnt+1th time, A represents a 1*d matrix, and the values ​​in the matrix are randomly selected from -1 or 1, and re represents the number of trackers;

[0093] Based on the optimization of the explorer and the tracker, the weight threshold matrix is ​​further optimized, and the iterative process of the scout is expressed as:

[0094]

[0095] in represents the global optimal position at the cnt-th iteration, ρ represents a random number that follows a normal distribution with a mean of 0 and a variance of 1, and f now represents the fitness value of the current individual, f minf Represents the minimum fitness value among all individuals, f maxf represents the maximum fitness value among all individuals, ε represents a non-zero minimum, and μ represents a random number in the range of [-1,1];

[0096] Through multiple iterations of explorers, trackers, and scouts, the weights and thresholds of each neuron are continuously updated in a direction suitable for training, and the updated results are returned to Link d*2 In the matrix.

[0097] After the weights and thresholds of the input layer and hidden layer of the SSA-ELM model are optimized, when the weights and thresholds of each neuron in the input layer and hidden layer are known, the preset set Q2 is used as the input of the SSA-ELM model. The output of the output layer of the SSA-ELM model is expressed as:

[0098]

[0099] Where SC represents the output value of the output layer, c lr represents the weight of the hidden layer output, w lr represents the weight of the input layer, b lr represents the threshold of the input layer, lr = 1, 2, 3 ... m, input = 1, 2, 3 ... d, Q 2，input represents the input feature factor in the preset set Q2; result represents a function, which is expressed as:

[0100] result(w lr , b lr , Q 2，input )=Sigmoid(w lr Q 2，input +b lr ) (10);

[0101] Where Sigmoid(·) represents the activation function;

[0102] The result(w lr , b lr , Q 2，input ) function as a whole H, H is regarded as a hidden layer output coefficient matrix, assuming that the current real temperature is T true , and the output layer has v neurons, then equation (9) is simplified to:

[0103]

[0104] The H matrix is ​​represented as:

[0105]

[0106] By using the matrix inversion method, we can solve the problem of c in equation (11) lr Solving for:

[0107] c lr =SC·H + (13);

[0108] Among them, H + is the pseudo-inverse form of the H matrix; respectively, the obtained c lr Substitute it into the SSA-ELM model to find the predicted value of the battery module temperature Tem at time t t ;

[0109] S03, such as Figure 3 As shown in the figure, a physical information neural network PINN is constructed. The physical information neural network PINN uses the residual of the thermal balance equation of the battery module as the loss function. There are two unknown variables in the loss function, namely, the temperature convergence value of the battery module and the variable of the temperature change of the battery module over time; the parameter value in the loss function is measured, and the predicted value of the battery module temperature Tem is used. t As the initial value of the battery module temperature convergence value, the initial value of the battery module temperature convergence value and the measured parameter value are input into the loss function to obtain the variable of the battery module temperature changing with time, and then the variable of the battery module temperature changing with time is input into the loss function. In this process, the result of the battery module temperature changing with time is generated, and the battery module temperature convergence value at that time is obtained; by updating the battery module temperature convergence value and the variable of the battery module temperature changing with time, the value of the loss function is continuously reduced until the loss function converges, and finally the optimal solution of the battery module temperature convergence value at that time is obtained;

[0110] The physical information neural network PINN physically constrains the battery module temperature through the battery module thermal balance equation, so that the battery module temperature follows the laws of physics. The introduction of the thermal balance equation injects physical prior knowledge of the battery module into the physical information neural network PINN. The thermal balance equation of the battery module is:

[0111]

[0112] Where mass is the mass of the battery module, cap is the heat capacity of the battery module, T n Indicates the temperature of the outer metal shell of the battery module, T e Indicates the temperature of the metal shell inside the battery module, R c，n Represents the resistance value of the outer metal shell of the battery module, R c，erepresents the internal metal shell resistance value, t represents time, ω represents the convection heat transfer coefficient of the battery module, area represents the cross-sectional area of ​​the battery module, T r Indicates the ambient temperature of the battery module, and heat indicates the heat generated by the battery module;

[0113] The Loss loss function is introduced, and the residual of the heat balance equation is regarded as the loss term, which is expressed as:

[0114]

[0115] Formula (15) is the loss function composed of the residual of the thermal balance equation of the battery module. The unknown variables in the loss function include the variables of the battery module temperature changing with time: Battery module temperature convergence value T, is a function of T with respect to time; measurement parameters are mass, cap, heat, T n , R c，n 、T e , R c，e ,ω,area,T r The data obtained in step S02 is t As the initial value of T, the initial value of T and the parameters mass, cap, heat, T n , R c，n 、T e , R c，e ,ω,area,T r The data is input into the loss function to obtain Then the obtained Input the loss function and find T at time t; in the training process of the physical information neural network PINN, as With the continuous updating of and T, the loss function gradually decreases, and the optimal solution of T at time t is solved.

[0116] The SSA-ELM model uses small sample data for training and uses the inverse matrix method to solve the hidden layer neuron weights in the SSA-ELM model, which has an advantage in prediction speed. The role of the physical information neural network PINN is to make the battery module temperature convergence value T gradually approach the actual temperature value through physical drive.

[0117] The convergence value T of the battery module temperature obtained by the physical information neural network PINN is detected, and the threshold δ is set to judge the predicted value Tem of the battery module temperature output by the SSA-ELM model. t The difference between the battery module temperature convergence value T output by the physical information neural network PINN, the threshold δ is selected as a random number in [0,0.3], and the battery module temperature prediction value Tem tThe judgment between the convergence value T of the battery module temperature is expressed as:

[0118] |T-Tem t |<δ (16);

[0119] When the battery module temperature prediction value Tem t If the battery module temperature convergence value T satisfies the condition of formula (16), the battery module temperature convergence value T is used as the result of the battery module temperature to be predicted at time t; if the condition of formula (16) is not satisfied, steps S01-S03 are repeated.

[0120] The foregoing description of specific exemplary embodiments of the present invention is for the purpose of illustration and demonstration. These descriptions are not intended to limit the present invention to the precise form disclosed, and it is clear that many changes and variations can be made based on the above teachings. The purpose of selecting and describing the exemplary embodiments is to explain the specific principles of the present invention and its practical application, so that those skilled in the art can realize and utilize various different exemplary embodiments of the present invention and various different selections and changes. The scope of the present invention is intended to be limited by the claims and their equivalents.

Claims

1. A battery module temperature prediction method based on physical information neural network and SSA-ELM, characterized by: The steps are: S01. Select a battery module, and calculate the mutual information estimation value of each characteristic factor and the battery module temperature respectively by using the mutual information method according to the characteristic factor historical data. The magnitude of the mutual information estimation value reflects the correlation between the characteristic factor and the battery module temperature. By setting a threshold for the mutual information estimation value, the types of characteristic factors whose mutual information estimation value exceeds the threshold are put into a preset set; the characteristic factors include chamber temperature, voltage, current, and state of charge; S02, the characteristic factor historical data and the corresponding battery module temperature historical data are combined into a pre-training data set, The characteristic factor historical data and the corresponding battery module temperature historical data are all data before time t. The SSA algorithm is used to optimize the ELM model to obtain the SSA-ELM model. The pre-trained data set is brought into the SSA-ELM model for training. The training content is to obtain the neuron weights and thresholds in the SSA-ELM model by the matrix inversion method. After the training is completed, the characteristic factors in the preset set are input into the SSA-ELM model, and the SSA-ELM model outputs the battery module temperature prediction value Tem t ; S03. Construct a physical information neural network PINN. The physical information neural network PINN uses the residual of the thermal balance equation of the battery module as a loss function. There are two unknown variables in the loss function, namely, the temperature convergence value of the battery module and the temperature variation of the battery module over time. The parameter values ​​in the loss function are measured and the predicted value of the battery module temperature Tem is used. t As the initial value of the battery module temperature convergence value, the initial value of the battery module temperature convergence value and the measured parameter value are input into the loss function to obtain the variable of the battery module temperature changing with time, and then the variable of the battery module temperature changing with time is input into the loss function. In this process, the result of the battery module temperature changing with time is generated, and the convergence value of the battery module temperature at that time is obtained; by updating the battery module temperature convergence value and the variable of the battery module temperature changing with time, the value of the loss function is continuously reduced until the loss function converges, and finally the optimal solution of the battery module temperature convergence value at that time is obtained.

2. The battery module temperature prediction method based on physical information neural network and SSA-ELM according to claim 1 is characterized in that: In step S01, the mutual information method represents the correlation between the battery module temperature and the characteristic factors; The characteristic factors are expressed as in represents i (i=1, 2, 3...N) characteristic factors at time t, where N represents the number of characteristic factors; Since the multiple historical data of each characteristic factor are in different quantity ranges, each characteristic factor historical data is normalized, and the data normalization is expressed as: in represents the maximum value among the i characteristic factors at time t, represents the minimum value of the i-th characteristic factor at time t, represents the average value of i characteristic factors at time t; represents the i characteristic factors after normalization; At time tL (L>0), calculate the battery module temperature history data Ta t-L Estimated mutual information with each normalized feature factor The expression of mutual information is expressed as: Where x represents the variable distribution of the battery module temperature, y represents the variable distribution of the characteristic factor, p(x) and p(y) are the marginal probability distribution functions of x and y respectively, and p(x, y) is the joint probability distribution function of x and y. The mutual information estimation value in formula (2) is placed in a set Q1, A threshold β is set, and the mutual information estimation value in Q1 is compared with the threshold β. The mutual information estimation value in Q1 that is greater than the threshold β is placed in the preset set Q2. Since the mutual information estimation value affects the prediction of the battery module temperature, and the mutual information estimation value is positively correlated with the influence of the characteristic factor on the battery module temperature, the characteristic factors with the mutual information estimation value above the threshold are selected. The preset set Q2 is expressed as: 。 3. The battery module temperature prediction method based on physical information neural network and SSA-ELM according to claim 2 is characterized in that: In step S02, the ELM model is a single hidden layer neural network, and the network structure consists of an input layer of d neurons, a hidden layer of m neurons, and an output layer of v neurons, wherein the length d of the input preset set Q2 corresponds to d neurons; the weights and thresholds of the input layer and the hidden layer are randomly generated during the training process, and the weight matrix w of the input layer j =[w1 w2……w d ] T , the threshold matrix b of the input layer k =[b1 b2……b d ] T , the weight matrix c of the hidden layer g =[c1 c2……c m ] T , the vector matrix out output by the output layer index =[out1 out2……out v ] T , where T represents the transpose operation of the matrix, w d represents the weight of the jth (j=1,2,3…d) neuron in the input layer, b d represents the threshold of the kth (k=1,2,3…d)th neuron in the input layer, c m represents the weight of the gth (g=1,2,3...m)th neuron in the hidden layer, out v Represents the output of the index(index=1,2,3……v)th neuron in the output layer.

4. The battery module temperature prediction method based on physical information neural network and SSA-ELM according to claim 3 is characterized by: In step S02, during the optimization of the ELM model, the SSA algorithm first transforms the weight matrix w of the input layer into j and the threshold matrix b of the input layer k =[b1 b2……b d ] T Concatenate into a weight threshold matrix Link d*2 , In the weight threshold matrix of formula (4), each row represents the weight and threshold of a neuron. The fitness value of each neuron is calculated respectively. The fitness value represents the quality of the data. The higher the fitness value, the more conducive the data generated by the neuron is to training. The calculation process of the fitness value is expressed as: ADP good Represents a matrix composed of fitness values, adp(w d , b d ) indicates that each row of the matrix represents the fitness value of different neurons, and adp(·) indicates the mapping function for calculating the fitness value; The SSA algorithm includes explorers, trackers, and scouts, where the explorer iteration process is expressed as: The Link matrix is ​​a weight threshold matrix with d rows and 2 columns. represents the data of the pth dimension in the oth row in the cntth iteration, p=1 represents the weight, and p=2 represents the threshold; oter max represents the preset maximum number of iterations, α is a random number in the interval [0,1]; R2 represents the warning value, ST represents the data safety value; ram is a random number that obeys the normal distribution, and L is a 1*d matrix of all 1s; Based on the explorer's optimization, the iterative process of the tracker is expressed as: in represents the worst data in the cnt-th iteration, represents the best position of the tracker at the cnt+1th time, A represents a 1*d matrix, and the values ​​in the matrix are randomly selected from -1 or 1, and re represents the number of trackers; Based on the optimization of the explorer and the tracker, the weight threshold matrix is ​​further optimized, and the iterative process of the scout is expressed as: in represents the global optimal position at the cnt-th iteration, ρ represents a random number that follows a normal distribution with a mean of 0 and a variance of 1, and f now represents the fitness value of the current individual, f minf Represents the minimum fitness value among all individuals, f maxf represents the maximum fitness value among all individuals, ε represents a non-zero minimum, and μ represents a random number in the range of [-1,1]; Through multiple iterations of explorers, trackers, and scouts, the weights and thresholds of each neuron are continuously updated in a direction suitable for training, and the updated results are returned to Link d*2 In the matrix.

5. The battery module temperature prediction method based on physical information neural network and SSA-ELM according to claim 4 is characterized in that: After the weights and thresholds of the input layer and hidden layer of the SSA-ELM model are optimized, when the weights and thresholds of each neuron in the input layer and hidden layer are known, the preset set Q2 is used as the input of the SSA-ELM model. The output of the output layer of the SSA-ELM model is expressed as: Where SC represents the output value of the output layer, c lr represents the weight of the hidden layer output, w lr represents the weight of the input layer, b lr represents the threshold of the input layer, lr = 1, 2, 3 ... m, input = 1, 2, 3 ... d, Q 2，input Represents the input feature factor in the preset set Q2; Result represents a function, expressed as: result(w lr ,b lr ,Q 2，input )=Sigmoid(w lr ·Q 2，input +b lr ) (10); Where Sigmoid(·) represents the activation function; The result(w lr , b lr , Q 2，input ) function as a whole H, H is regarded as a hidden layer output coefficient matrix, then formula (9) can be simplified as: The H matrix is ​​represented as: By using the matrix inversion method, we can solve the problem of c in equation (11) lr Solving for: c lr =SC·H + (13); Among them, H + is the pseudo-inverse form of the H matrix; The obtained c lr Substitute it into the SSA-ELM model to find the predicted value of the battery module temperature Tem at time t t .

6. The battery module temperature prediction method based on physical information neural network and SSA-ELM according to claim 1 is characterized in that: In step S03, the physical information neural network PINN physically constrains the battery module temperature through the battery module thermal balance equation, so that the battery module temperature follows the laws of physics. The introduction of the thermal balance equation injects physical prior knowledge of the battery module into the physical information neural network PINN. The thermal balance equation of the battery module is: Where mass is the mass of the battery module, cap is the heat capacity of the battery module, T n Indicates the temperature of the outer metal shell of the battery module, T e Indicates the temperature of the metal shell inside the battery module, R c，n Represents the resistance value of the outer metal shell of the battery module, R c，e represents the internal metal shell resistance value, t represents time, ω represents the convection heat transfer coefficient of the battery module, area represents the cross-sectional area of ​​the battery module, T r Indicates the ambient temperature of the battery module, and heat indicates the heat generated by the battery module; The Loss loss function is introduced, and the residual of the heat balance equation is regarded as the loss term, which is expressed as: Formula (15) is the loss function composed of the residual of the thermal balance equation of the battery module. The unknown variables in the loss function include the variables of the battery module temperature changing with time: Battery module temperature convergence value T, is a function of T with respect to time; measurement parameters are mass, cap, heat, T n , R c，n , T e , R c，e ,ω,area,T r The data obtained in step S02 is t As the initial value of T, the initial value of T and the parameters mass, cap, heat, T n , R c，n , T e , R c，e ,ω,area,T r The data is input into the loss function to obtain Then the obtained Input the loss function and find T at time t; in the training process of the physical information neural network PINN, as With the continuous updating of and T, the loss function gradually decreases, and the optimal solution of T at time t is solved.

7. The battery module temperature prediction method based on physical information neural network and SSA-ELM according to claim 6 is characterized by: In step S03, the battery module temperature convergence value T obtained by the physical information neural network PINN is detected, and a threshold δ is set to determine the battery module temperature prediction value Tem output by the SSA-ELM model. t The difference between the battery module temperature convergence value T output by the physical information neural network PINN, the threshold δ is selected as a random number in [0,0.3], and the battery module temperature prediction value Tem t The judgment between the convergence value T of the battery module temperature is expressed as: |T-Time t |<δ(16); When the battery module temperature prediction value Tem t If the battery module temperature convergence value T satisfies the condition of formula (16), the battery module temperature convergence value T is used as the result of the battery module temperature to be predicted at time t; if the condition of formula (16) is not satisfied, steps S01-S03 are repeated.

Images