Lithium battery temperature prediction method and system

Abstract

The invention discloses a lithium battery temperature prediction method and system, and the method comprises the steps: obtaining a lithium battery data set, optimizing a TVF-EMD model through a PSA algorithm, obtaining an improved PSA-TVF-EMD, decomposing the data set through the PSA-TVF-EMD, and dividing the decomposed components into high-frequency components and low-frequency components through Hilbert transformation; secondly, replacing an Inception module in the Times Net by using ConvGRU to obtain an improved GTimes Net model, and establishing a GTimes Net prediction model to predict a high-frequency component and a low-frequency component; the method comprises the following steps: initializing a population of a gold scrubber algorithm by adopting Gaussian distribution initialization, and adding a chaos inertia weight and a sine function learning factor strategy in an exploration stage to obtain an improved gold scrubber optimizer algorithm IGRO to optimize model hyper-parameters; and finally, performing secondary weighting on the prediction results of the high-frequency component and the low-frequency component by utilizing IGRO to obtain a final prediction result. The precision and effectiveness of lithium battery temperature prediction can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium batteries, and in particular to a lithium battery temperature prediction method and system. Background Art

[0002] The lithium battery industry currently presents a multi-faceted development trend of policy support, steady growth in shipments, strong market demand, fierce corporate competition and a sound industrial chain. Lithium batteries are widely used in many fields such as electric vehicles, energy storage systems, 3C products, military and aerospace. Lithium batteries have the characteristics of high energy density, long cycle life and low self-discharge rate, and also show good performance in terms of safety. It is widely used in new energy vehicles, portable electronic devices and energy storage systems, promoting energy transformation and technological progress. The key performance indicators of lithium batteries include energy density, power density, cycle life, safety and fast charging capability. These indicators jointly determine the performance and scope of application of the battery. The impact of temperature on the performance and safety of lithium batteries is significant. In low temperature environments, the available capacity is reduced, the charge and discharge power is limited, and the electrochemical reaction rate at the interface of the electrode and electrolyte is slowed down, resulting in a decrease in the battery charge and discharge performance. The capacity loss caused by low temperature discharge is usually reversible, while low temperature charging may cause the formation of metal lithium dendrites, resulting in permanent capacity loss. In high temperature environments, high temperature will accelerate the structural destruction of the positive electrode material and the volatilization of the electrolyte, resulting in a decrease in battery capacity. At the same time, high temperature increases the risk of thermal runaway of the battery, which may cause leakage, gassing, smoking, or even explosion. Therefore, the lithium battery temperature prediction method is crucial to prevent battery overheating, maintain stable battery performance, and ensure safe use. By accurately monitoring and predicting battery temperature, safety accidents caused by abnormal temperature can be effectively prevented and the battery life can be extended.

[0003] Existing lithium battery temperature prediction methods have some shortcomings, including time-varying model parameters, sensitivity to aging effects, high cost of measurement equipment, difficult feature selection, limited model generalization ability, temperature measurement noise interference and high model complexity. These problems limit the accuracy and application scope of the prediction model. Further research and development of more efficient, accurate and adaptable prediction models are needed to overcome these challenges. Summary of the invention

[0004] Purpose of the invention: The present invention provides a lithium battery temperature prediction method and system, which effectively improves the accuracy of lithium battery temperature prediction.

[0005] Technical solution: A lithium battery temperature prediction method according to the present invention comprises the following steps:

[0006] (1) Obtaining lithium battery temperature related data;

[0007] (2) Conduct in-depth feature selection on the raw data of lithium batteries and use regression models to identify and select the feature variables most relevant to battery temperature changes;

[0008] (3) Using the PSA algorithm to optimize TVF-EMD to obtain the improved PSA-TVF-EMD;

[0009] (4) The lithium battery data set is decomposed using the improved PSA-TVF-EMD model, and the decomposed components are divided into high-frequency components and low-frequency components through Hilbert transform;

[0010] (5) Use ConvGRU to replace the Inception module in TimesNet to obtain an improved GTimesNet model, and establish a GTimesNet prediction model to predict high-frequency components and low-frequency components;

[0011] (6) Improve the gold panning optimization algorithm GRO: Use Gaussian distribution initialization to initialize the population of the gold panning algorithm, and add chaotic inertia weight CIW and sine function learning factor strategy in the exploration stage to obtain the improved gold panning optimizer algorithm IGRO;

[0012] (7) Use the improved gold mining optimizer algorithm IGRO to optimize the parameters of GTimesNet;

[0013] (8) Use IGRO to perform secondary weighting on the prediction results of the high-frequency component and the low-frequency component to obtain the final prediction result.

[0014] Furthermore, the implementation process of step (1) is as follows:

[0015] Lithium battery temperature-related data include current, voltage, charge state, temperature, internal resistance, ambient temperature and aging degree;

[0016] Create a data matrix with m sample data, each sample contains n feature data, forming the data matrix X of lithium battery:

[0017]

[0018] Where, X i is the parameter of the i-th lithium battery sample, X i (k) is the kth characteristic data parameter of the parameters of the i-th lithium battery sample.

[0019] Furthermore, the implementation process of step (2) is as follows:

[0020] (21) Normalize the data:

[0021]

[0022]

[0023] Where, X i '(k) is the kth indicator data in the temperature parameter of the i-th lithium battery sample after dimensionless conversion; i = 0, 1, ..., m; k = 0, 1, 2, 3; AM is the mean of the data, σ is the standard deviation of the data;

[0024] (22) Constructing the observation value vector where y1,y2...y m is the temperature of the lithium battery;

[0025] (23) Minimize the residual sum of squares:

[0026]

[0027] In the formula, is the model prediction value, X is the lithium battery data set;

[0028] (24) Taking the derivative of S(β) and setting it to zero, we obtain the normal equation:

[0029] X T Xβ=X T Y (6)

[0030] Solving the normal equations yields least squares estimates of the regression coefficients:

[0031]

[0032] In the formula, is the vector of regression coefficients;

[0033] (25) Construct a linear regression model:

[0034] Y=β0+β1X1+β2X2+…+β n X n + (8)

[0035] Among them, β0 is the intercept term, β1, β2...β n is the regression coefficient corresponding to the respective variable, ò is the error term, which represents the random variation that cannot be explained by the model;

[0036] (26) Calculate the coefficient of determination:

[0037]

[0038] In the formula, SS res is the residual sum of squares, that is SS tot is the total sum of squares, i.e. is the sample mean of the dependent variable Y; if the coefficient of determination R 2 The larger the value is, the greater the impact of this characteristic data on the temperature of the lithium battery.

[0039] Furthermore, the implementation process of step (3) is as follows:

[0040] (31) Parameter encoding, where the bandwidth threshold and B-spline order are encoded as decision variables, and each particle in PSA represents a potential solution, i.e., a pair of parameter values;

[0041] (32) Initialize the population, randomly initialize the particle swarm, and the position of each particle represents a parameter combination:

[0042]

[0043] In the formula, is the bandwidth threshold of the ith particle, is the B-spline order of the i-th particle;

[0044] (33) Define the fitness function and calculate the initial fitness, using the signal-to-noise ratio (SNR) as the objective function:

[0045] SNR = 10log 10 (P signal / P noise )(11)

[0046] Where P signal is the signal power, P noise is the noise power;

[0047] (34) Use PID control theory to update the particle position and system deviation:

[0048] Δu(t)=K p ·r2·[e k (t)-e k-1 (t)]+K i ·r3·e k (t)+K d ·r4·[e k (t)-2e k-1 (t)+e k-2 (t)](12)

[0049]

[0050] In the formula, K p , K i and K dare the proportional, integral and differential coefficients of the PID controller, r2, r3 and r4 are random number vectors in the interval [0,1]; The individual x is the minimum value in the overall history psa (t-1) is the particle position at the previous iteration;

[0051] (35) Update the particle position according to the PID control value:

[0052] x psa (t+1)=x psa (t)+η·Δu(t)+(1-η)·o(t) (14)

[0053] Where η is the adjustment factor composed of random numbers, and o(t) is the zero output adjustment factor;

[0054] (36) Check the termination condition and output the optimal solution. If the termination condition is met, stop the iteration; otherwise, continue the iteration; output the bandwidth threshold and B-spline order combination that maximizes the fitness function to obtain the improved PSA-TVF-EMD decomposition model.

[0055] Furthermore, the implementation process of step (4) is as follows:

[0056] (41) The original signal is preprocessed and the trend term and high-frequency noise are removed using TVF-EMD;

[0057] (42) A time-varying filter is constructed to filter the signal and extract the intrinsic mode function of the signal until the remaining signal is a monotonic function; the time-varying filter h(t) is:

[0058]

[0059] The signal x(t) is decomposed into several IMFs and a residual term r(t):

[0060]

[0061] (43) Determine the local cutoff frequency using Hilbert transform:

[0062] First, calculate the Hilbert transform of the signal x(t)

[0063]

[0064] Where H{x(t)} is the Hilbert transform of the signal x(t), and P represents the Cauchy principal value;

[0065] Secondly, the local maximum sequence and local minimum sequence of the instantaneous amplitude are determined as:

[0066] {A max (t)} and {A min (t)}

[0067] Then, the local maximum and minimum sequences are interpolated to calculate the instantaneous mean and instantaneous envelope to obtain β1(t) and β2(t):

[0068] β1(t)=Interpolation(A max (t))

[0069] β2(t)=Interpolation(A min (t))

[0070] Instantaneous mean α1(t):

[0071] Instantaneous mean α2(t):

[0072] Finally, calculate the instantaneous frequency component and calculate the local cutoff frequency:

[0073] Instantaneous frequency component φ′1(t):

[0074]

[0075] Instantaneous frequency component φ′2(t):

[0076]

[0077] Where η1(t) and η2(t) are obtained by max (t)} and {A min (t)} interpolated;

[0078] Local cutoff frequency ω′ bis (t):

[0079]

[0080] (44) Continuously iterate and extract IMF according to the stopping criterion:

[0081] Instantaneous frequency:

[0082]

[0083] Where ω(t) is the instantaneous phase of the signal, which is obtained from the phase of the signal after Hilbert transformation; Loughlin instantaneous bandwidth:

[0084]

[0085] Where sgn is the sign function, which is used to determine the direction of frequency change;

[0086] Weighted average instantaneous frequency:

[0087]

[0088] Where w(t) is the weight function;

[0089] Local mean of instantaneous envelope:

[0090]

[0091] Where A(t) is the instantaneous amplitude;

[0092] Stopping criterion θ(t):

[0093]

[0094] If θ(t) < ξ, where ξ is a preset threshold, it is considered that an IMF has been successfully extracted and the iteration stops;

[0095] The expression of the original signal X(t) decomposed by TVF-EMD is:

[0096]

[0097] Where u(t) is the health status of the lithium battery, IMF i (t) is the i-th intrinsic mode function, R(t) is the residual term after decomposition, and N is the number of IMFs obtained by decomposition;

[0098] (45) The decomposed signal is divided into high-frequency components and low-frequency components using Hilbert transform:

[0099] Obtain the analytical signal. For each IMFu(t), its analytical signal z(t) is expressed as:

[0100] z(t)=u(t)+j·H{u(t)}(30)

[0101] Where j is the imaginary unit, H{u(t)} is the Hilbert transform of u(t);

[0102] The first bandpass filter is applied to the real part of the analytical signal to obtain the low-frequency components; the second bandpass filter is applied to the imaginary part of the analytical signal to obtain the high-frequency components.

[0103] Furthermore, the implementation process of step (5) is as follows:

[0104] (51) The ConvGRU module will contain convolutional layers to process spatial data and GRU units to process time series data; in TimesNet, the Inception module is replaced by the ConvGRU module to ensure that the input and output dimensions of the ConvGRU module match those of the Inception module; according to the characteristics of the ConvGRU module, the parameters and activation functions of the convolutional layers in TimesNet are adjusted to obtain an improved GTimesNet model;

[0105] (52) Use the GTimesNet model to predict high-frequency components and low-frequency components:

[0106] Compute the improved update gate:

[0107]

[0108] In the formula, ug t is the output of the update gate is a low frequency signal (high frequency signal), W ug and U ug is the input of the convolution and acts on the current time step respectively and the hidden state H at the previous time step t-1 , b ug is the bias term, σ is the sigmoid activation function;

[0109] Calculate the improved reset gate:

[0110] rg t =σ(W rg *X t +U rg *H t-1 +b rg ) (32)

[0111] In the formula, rg t is the output of the reset gate, b rg is the bias term;

[0112] Compute candidate hidden states:

[0113]

[0114] In the formula, is the candidate hidden state after the activation function, b h is the bias term, tanh is the hyperbolic tangent activation function;

[0115] Output predictions, final hidden state:

[0116]

[0117] In the formula, Ht is the predicted low-frequency component or high-frequency component.

[0118] Furthermore, the implementation process of step (6) is as follows:

[0119] (61) Set the population size and maximum number of iterations of the IGRO algorithm, as well as the upper and lower limits of the search space;

[0120] (62) The position of the gold prospector is initialized using Gaussian distribution, which is defined as follows:

[0121] Q ij =μ·σN(0,1) (35)

[0122] In the formula, Q ij represents the position of the i-th individual in the j-th dimension; μ is the mean of the Gaussian distribution set to the current best gold mine position; σ is the standard deviation of the Gaussian distribution, a smaller σ means that the new position is closer to the current best position, while a larger σ allows the new position to explore a wider range; N(0,1) is a random number from a standard normal distribution;

[0123] The positions of the gold prospectors are randomly initialized, and the position of each gold prospector is regarded as the approximate position of the gold mine:

[0124]

[0125] In the formula, Represents the mean A random vector drawn from a Gaussian distribution with a standard deviation of σ, ep is the current iteration number;

[0126] (63) During the migration phase of the gold prospector, the chaotic inertia weight CIW and sine function learning factor strategy are used to update the position of the gold prospector:

[0127] Q new =Q+Ω(ep)·C(ep)·(Q * -Q) (37)

[0128] Where Q is the current position, Q * is the current optimal position, inertia weight ep max is the maximum number of iterations, k is a constant used to control the intensity of chaotic behavior; the sine function learning factor

[0129] (64) Gold prospector cooperation stage:

[0130] In the cooperation stage, prospectors will cooperate with each other to improve efficiency. The update formula combining CIW and sine function is:

[0131]

[0132] in, and are the positions of two randomly selected prospectors at the current iteration number ep;

[0133] (65) Gold prospector gold panning stage:

[0134]

[0135] In the formula, is the position of the randomly selected reference prospector at the current iteration number ep;

[0136] In order to decide whether the prospector should stay in the original position or move to a new position, the fitness value is compared; basic fitness function:

[0137]

[0138] In the formula, G represents a dim-dimensional vector, G i Represents the i-th element of vector G, dim is the dimension;

[0139] Dynamic fitness function:

[0140] F(G,ep)=F0(G)·Ω(ep)·C(ep) (41)

[0141] Using the dynamic fitness function F(G,ep), compare the prospector's new position Q new and the fitness of the current position Q are Fitness new =F(G new ,ep) and Fitness current =F(G,ep); According to the fitness comparison result, decide whether to update the prospector's position:

[0142]

[0143] If Fitness new >Fitness current , then accept the new position: Q(ep+1)=Q new ; If Fitness new <Fitness current , then accept the new position with a certain probability: Q(ep+1)=Q new .

[0144] Furthermore, the implementation process of step (7) is as follows:

[0145] (71) Find the objective function MSE of GTimesNet and randomly generate a set of hyperparameter combinations as the initial population:

[0146]

[0147] Where n GT is the sample size;

[0148] (72) The GTimesNet model was used to evaluate the performance of each set of hyperparameters on the validation set and calculate the fitness value. The objective function MSE was used as the fitness evaluation object. The smaller the MSE, the smaller the prediction error and the best performance.

[0149] (73) Update position: According to the update rule of the IGRO algorithm, combined with the chaotic inertia weight CIW and the sine function learning factor, the position of each individual in the population is updated:

[0150]

[0151] Where β is the inertia weight is the gradient of the objective function F at position Q;

[0152] (74) According to the fitness value, select the hyperparameter combination with better performance and replace the combination with poor performance;

[0153] (75) Repeat steps (73) and (74) until the termination condition is met, such as reaching the maximum number of iterations or the fitness value converges.

[0154] Furthermore, the implementation process of step (8) is as follows:

[0155]

[0156] In the formula, MSE High and MSE Low are the errors of high-frequency component and low-frequency component respectively, α1 is the weight of high-frequency component, and α2 is the weight of low-frequency component;

[0157] The calculated weight α is used to fuse the prediction results of high-frequency and low-frequency components:

[0158] Final Prediction=α1×Prediction High+α2×Prediction Low (47)

[0159] If α1 is close to 1, it means that the high-frequency component has a greater impact on the final prediction result; if α2 is close to 1, it means that the low-frequency component has a greater impact on the final prediction result.

[0160] A lithium battery temperature prediction system according to the present invention comprises the following modules:

[0161] Lithium battery temperature-related data collection module, which collects data including current, voltage, charging status, temperature, internal resistance, ambient temperature and aging degree;

[0162] The correlation detection module performs in-depth feature selection on the raw data of lithium batteries and uses regression models to identify and select the feature variables most relevant to battery temperature changes;

[0163] The signal decomposition module decomposes the lithium battery data set and divides the decomposed components into high-frequency components and low-frequency components through Hilbert transform;

[0164] The lithium battery temperature prediction module uses ConvGRU to replace the Inception module in TimesNet to obtain an improved GTimesNet model, and establishes a GTimesNet prediction model for lithium battery temperature prediction model;

[0165] The intelligent optimization module for lithium battery temperature prediction uses Gaussian distribution initialization to initialize the population of the gold panning algorithm. In the exploration phase, chaotic inertia weights and sinusoidal function learning factor strategies are added to obtain the improved gold panning optimizer algorithm IGRO. The hyperparameters of the GTimesNet model are optimized, and IGRO is used to perform secondary weighting on the prediction results of high-frequency components and low-frequency components to obtain the final lithium battery temperature prediction results.

[0166] Beneficial effects: Compared with the prior art, the beneficial effects of the present invention are as follows:

[0167] 1. The present invention obtains key data including lithium batteries to form a data set, establishes a linear regression equation to analyze the importance of lithium battery characteristics, and uses TVF-EMD to decompose the data set, and divides the decomposed components into high-frequency components and low-frequency components through Hilbert transformation; this method can more accurately capture the subtle characteristics of lithium battery temperature changes, improve the model's sensitivity to data and the accuracy of prediction;

[0168] 2. The present invention uses the PID search algorithm to improve TVF-EMD to obtain the PSA-TVF-EMD decomposition model, uses this model to decompose the data set, and divides the decomposed components into high-frequency components and low-frequency components through Hilbert transform; this advanced signal processing technology can more effectively capture the nonlinear and non-stationary characteristics in the data compared to traditional time series analysis methods, such as the moving average model (MA) and the autoregressive moving average model (ARMA). Through this classification, the model can more accurately identify and predict the trend and periodicity of lithium battery temperature changes, improving the accuracy of the prediction;

[0169] 3. The present invention replaces the Inception module in TimesNet with ConvGRU to obtain an improved GTimesNet model. Compared with traditional RNN models, such as LSTM and GRU, this improvement can better process time series data, especially when processing lithium battery temperature data with obvious time dependence, it can more accurately capture the long-term dependencies in the time series; compared with the Transformer model, the GTimesNet model can more effectively utilize local time information and improve the prediction ability of the model when processing time series prediction problems due to its unique time convolution structure;

[0170] 4. The present invention uses Gaussian Gaussian distribution initialization to initialize the population of the gold panning algorithm, and adds chaotic inertia weight (CIW) and sine function learning factor strategy in the exploration stage to obtain an improved gold panning optimizer algorithm IGRO; compared with the standard gold panning optimization algorithm, this improved algorithm can more effectively optimize the model hyperparameters and improve the generalization ability of the model. Compared with the commonly used grid search or random search method, the IGRO algorithm can converge to the optimal solution faster and can better adapt to the complex search space of the model hyperparameters, thereby improving the accuracy and effectiveness of the lithium battery temperature prediction model; by IGRO, the prediction results of the high-frequency component and the low-frequency component are weighted twice, and a more accurate final prediction result can be obtained. Compared with the simple average or single component prediction, this weighting method can more comprehensively consider the influence of different frequency components on temperature prediction, further improving the accuracy of the prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0171] Figure 1 This is a flow chart of the lithium battery temperature prediction method proposed by the present invention;

[0172] Figure 2 Flowchart for classifying the importance of lithium battery features;

[0173] Figure 3 Optimize TVF-EMD flow chart for PID search algorithm;

[0174] Figure 4 The improved IGRO algorithm flow chart proposed by the present invention;

[0175] Figure 5 Optimizing the GTimesNet model flow chart for the IGRO algorithm. DETAILED DESCRIPTION

[0176] The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0177] like Figure 1As shown, the present invention discloses a method for accurately predicting the temperature of a lithium battery, which specifically includes the following steps.

[0178] Step 1: Obtain a lithium battery data set (lithium battery temperature related data) and preprocess the data set.

[0179] In order to predict the temperature of lithium batteries, a series of data related to the temperature of lithium batteries need to be collected from the lithium batteries, including current, voltage, charging state, temperature, internal resistance, ambient temperature and aging degree.

[0180] Create a data matrix. Assume that there are m sample data and each sample contains n feature data to form the data matrix X of the lithium battery:

[0181]

[0182] Where, X i is the parameter of the i-th lithium battery sample, X i (k) is the kth characteristic data parameter of the parameters of the i-th lithium battery sample.

[0183] Step 2: If Figure 2 As shown, an in-depth feature selection is performed on the raw data of lithium batteries, and a regression model is used to identify and select the feature variables most relevant to battery temperature changes.

[0184] Step 2.1: The data needs to be normalized to improve the prediction fit of the model. The formula is as follows:

[0185]

[0186] Where, X i '(k) is the kth indicator data in the temperature parameter of the i-th lithium battery sample after dimensionless conversion; i=0,1,...,m; k=0,1,2,3; AM is the mean of the data, and σ is the standard deviation of the data.

[0187] Step 2.2: Construct the observation vector where y1,y2...y m is the temperature of the lithium battery.

[0188] Step 2.3: Minimize the residual sum of squares:

[0189]

[0190] In the formula, is the model prediction value, and X is the lithium battery dataset.

[0191] Step 2.4: To find the value of β that minimizes S(β), we take the derivative of S(β) and set it to zero, obtaining the normal equation:

[0192] X T Xβ=X T Y (6)

[0193] Solving the normal equations yields least squares estimates of the regression coefficients:

[0194]

[0195] In the formula, is the vector of regression coefficients.

[0196] Step 2.5: Build a linear regression model:

[0197] Y=β0+β1X1+β2X2+…+β n X n + (8)

[0198] Among them, β0 is the intercept term, β1, β2...β n is the regression coefficient corresponding to each variable (from Equation 7), and ò is the error term, which represents the random variation that cannot be explained by the model.

[0199] Step 2.6: Calculate the coefficient of determination:

[0200]

[0201] In the formula, SS res is the residual sum of squares, that is SS tot is the total sum of squares, i.e. is the sample mean of the dependent variable Y. If the coefficient of determination R 2 The larger the value is, the greater the impact of this characteristic data on the temperature of the lithium battery.

[0202] Step 3: If Figure 3 As shown, the improved PSA-TVF-EMD is obtained by optimizing the TVF-EMD using the PID search algorithm (PSA).

[0203] Step 3.1: Parameter encoding, the bandwidth threshold and B-spline order are encoded as decision variables. Each particle in PSA represents a potential solution, i.e., a pair of parameter values ​​(bandwidth threshold, B-spline order).

[0204] Step 3.2: Initialize the population, randomly initialize the particle swarm, and the position of each particle represents a parameter combination:

[0205]

[0206] In the formula, is the bandwidth threshold of the ith particle, is the B-spline order of the ith particle.

[0207] Step 3.3: Define the fitness function and calculate the initial fitness. Use the signal-to-noise ratio (SNR) as the objective function. The calculation formula can be expressed as:

[0208] SNR = 10log 10 (P signal / P noise )(11)

[0209] Where P signal is the signal power, P noise is the noise power.

[0210] Step 3.4: Update the particle position and system deviation using PID control theory:

[0211] Δu(t)=K p ·r2·[e k (t)-e k-1 (t)]+K i ·r3·e k (t)+K d ·r4·[e k (t)-2e k-1 (t)+e k-2 (t)](12)

[0212]

[0213] In the formula, K p , K i and K d are the proportional, integral and differential coefficients of the PID controller respectively, and r2, r3 and r4 are random number vectors in the interval [0,1]. The individual x is the minimum value in the overall history psa (t-1) is the particle position at the previous iteration.

[0214] Step 3.5: Update the particle position according to the PID control amount:

[0215] x psa (t+1)=x psa (t)+η·Δu(t)+(1-η)·o(t) (14)

[0216] Where η is a regulation factor composed of random numbers, and o(t) is a zero output regulation factor.

[0217] Step 3.6: Check the termination condition and output the optimal solution. If the termination condition is met (such as reaching the maximum number of iterations or fitness convergence), stop the iteration; otherwise, continue the iteration. Output the bandwidth threshold and B-spline order combination that maximizes the fitness function, and obtain the improved PSA-TVF-EMD decomposition model.

[0218] The TVFEMD improved by the PID search algorithm significantly improves the accuracy and stability of signal decomposition through adaptive filters and multi-scale analysis capabilities, effectively reduces modal aliasing, and enhances robustness to noise. This method is particularly suitable for the processing of nonlinear and non-stationary signals, and can provide more comprehensive signal understanding and analysis results. In addition, it can reduce modeling costs and computational workload while maintaining predictive performance.

[0219] Step 4: Use time-varying filtered empirical mode decomposition (TVF-EMD) to decompose the lithium battery data set, and divide the decomposed components into high-frequency components and low-frequency components through Hilbert Transform.

[0220] Step 4.1: Preprocess the original signal and use TVF-EMD to remove trend terms and high-frequency noise

[0221] The first IMF component usually contains the component with the smallest time scale (highest frequency) in the original signal. As the IMF order increases, the corresponding frequency component gradually decreases. The frequency component of the residual r_n is the lowest. When r_n becomes a monotonic function, it can be used as a trend term. The high-frequency component of the signal in the first IMF component obtained by TVF-EMD decomposition realizes the removal of trend terms and high-frequency noise.

[0222] Step 4.2: Construct a time-varying filter to filter the signal and extract the intrinsic mode function of the signal until the remaining signal is a monotonic function. The time-varying filter h(t) can be defined as:

[0223]

[0224] The signal x(t) can be decomposed into several IMFs and a residual term r(t):

[0225]

[0226] Step 4.3: Determine the local cutoff frequency using the Hilbert transform:

[0227] First, calculate the Hilbert transform of the signal x(t)

[0228]

[0229] Where H{x(t)} is the Hilbert transform of the signal x(t), and P represents the Cauchy principal value.

[0230] Secondly, the local maximum sequence and local minimum sequence of the instantaneous amplitude are determined as:

[0231] {A max (t)} and {A min (t)}

[0232] Then, the local maximum and minimum sequences are interpolated to calculate the instantaneous mean and instantaneous envelope to obtain β1(t) and β2(t):

[0233] β1(t)=Interpolation(A max (t))

[0234] β2(t)=Interpolation(A min (t))

[0235] Instantaneous mean α1(t):

[0236] Instantaneous mean α2(t):

[0237] Finally, calculate the instantaneous frequency component and calculate the local cutoff frequency:

[0238] Instantaneous frequency component φ′1(t):

[0239]

[0240] Instantaneous frequency component φ′2(t):

[0241]

[0242] Where η1(t) and η2(t) are obtained by max (t)} and {A min (t)} interpolated.

[0243] Local cutoff frequency ω′ bis (t):

[0244]

[0245] Step 4.4: Iterate and extract IMF according to the stopping criteria:

[0246] Instantaneous frequency:

[0247]

[0248] Where ω(t) is the instantaneous phase of the signal, which is obtained from the signal phase after Hilbert transformation.

[0249] Loughlin instantaneous bandwidth:

[0250]

[0251] Where sgn is the sign function, which is used to determine the direction of frequency change.

[0252] Weighted average instantaneous frequency:

[0253]

[0254] Where w(t) is the weight function.

[0255] Local mean of instantaneous envelope:

[0256]

[0257] Where A(t) is the instantaneous amplitude.

[0258] Stopping criterion θ(t):

[0259]

[0260] If θ(t)<ξ, where ξ is a preset threshold, it is considered that an IMF has been successfully extracted and the iteration can be stopped.

[0261] The expression of the original signal X(t) decomposed by TVF-EMD is:

[0262]

[0263] Where u(t) is the health status of the lithium battery, IMF i (t) is the i-th intrinsic mode function, R(t) is the residual term after decomposition, and N is the number of IMFs obtained by decomposition.

[0264] Step 4.5: Use Hilbert transform to separate the decomposed signal into high-frequency components and low-frequency components:

[0265] Obtain the analytical signal. For each IMFu(t), its analytical signal z(t) can be expressed as:

[0266] z(t)=u(t)+j·H{u(t)}(30)

[0267] Where j is the imaginary unit, H{u(t)} is the Hilbert transform of u(t), which is obtained by Equation 12.

[0268] Apply the first bandpass filter to the real part of the analytical signal to obtain the low-frequency components. Apply the second bandpass filter to the imaginary part of the analytical signal to obtain the high-frequency components.

[0269] Step 5: Decompose the lithium battery data set using the improved PSA-TVF-EMD model, and divide the decomposed components into high-frequency components and low-frequency components through Hilbert Transform;

[0270] Step 5.1: Design a ConvGRU module to replace the Inception module. The ConvGRU module will contain convolutional layers to process spatial data, and GRU units to process time series data. In the TimesNet code, find the location of the Inception module and replace it with the designed ConvGRU module. Make sure that the input and output dimensions of the ConvGRU module match those of the Inception module. According to the characteristics of the ConvGRU module, other parts in TimesNet may need to be adjusted, such as the parameters of the convolutional layer, the activation function, etc., to ensure the performance of the entire network and obtain an improved GTimesNet model. Using the ConvGRU module to replace the Inception module in lithium battery temperature prediction can bring many benefits such as improved prediction accuracy, enhanced model adaptability, optimized computational efficiency, and improved generalization ability.

[0271] Step 5.2: Use the GTimesNet model to predict high-frequency components and low-frequency components:

[0272] Compute the improved update gate:

[0273]

[0274] In the formula, ug t is the output of the update gate is a low frequency signal (high frequency signal), W ug and U ug is the input of the convolution and acts on the current time step respectively and the hidden state H at the previous time step t-1 , b ug is the bias term, and σ is the sigmoid activation function.

[0275] Calculate the improved reset gate:

[0276] rg t =σ(W rg *X t +U rg *H t-1 +b rg ) (32)

[0277] In the formula, rg t is the output of the reset gate, b rg is the bias term

[0278] Compute candidate hidden states:

[0279]

[0280] In the formula, is the candidate hidden state after the activation function, b h is the bias term, and tanh is the hyperbolic tangent activation function.

[0281] Output predictions, final hidden state:

[0282]

[0283] In the formula, H t is the predicted low-frequency component (high-frequency component).

[0284] Using the GTimesNet model to predict the high-frequency and low-frequency components of lithium battery data can significantly improve the prediction accuracy and adaptability of the model. This model combines the advantages of convolutional neural networks and gated recurrent units to effectively capture local features and long-term dependencies in time series, thereby more accurately identifying temperature change trends. At the same time, the gating mechanism of the GTimesNet model enables it to dynamically adjust its response to input data, enhancing the model's adaptability to changing conditions. In addition, the GTimesNet model exhibits higher computational efficiency when processing sequence data, reduces the computational burden of traditional complex feature extraction, and improves the model's generalization ability for unknown data by learning the inherent laws of the data. For nonlinear and non-stationary lithium battery temperature data, the GTimesNet model can also better capture its complex dynamic changes, thus playing an important role in lithium battery temperature prediction.

[0285] Step 6: Figure 4 As shown, the gold panning optimization algorithm (GRO) is improved: the population of the gold panning algorithm is initialized using Gaussian distribution initialization, and the chaotic inertia weight (CIW) and sine function learning factor strategy are added in the exploration stage to obtain the improved gold panning optimizer algorithm IGRO.

[0286] Step 6.1: Set the population size and maximum number of iterations of the IGRO algorithm, as well as the upper and lower limits of the search space;

[0287] Step 6.2: Use Gaussian distribution to initialize the position of the gold prospector, which is defined as follows:

[0288] Q ij =μ+σ·N(0,1) (35)

[0289] In the formula, Q ij represents the position of the i-th individual in the j-th dimension. μ is the mean of the Gaussian distribution set to the current best gold mine position. σ is the standard deviation of the Gaussian distribution, a smaller σ means that the new position is closer to the current best position, while a larger σ allows the new position to explore a wider range. N(0,1) is a random number from a standard normal distribution.

[0290] The positions of the gold prospectors are randomly initialized, and the position of each gold prospector is regarded as the approximate position of the gold mine:

[0291]

[0292] In the formula, Represents the mean A random vector drawn from a Gaussian distribution with standard deviation σ. ep is the current iteration number.

[0293] This method enhances the algorithm's search efficiency and diversity by providing a wide initial search space, helping to avoid local optimal solutions and improve global search capabilities. At the same time, it reduces dependence on specific initial conditions, enhances the robustness of the model, and accelerates the convergence of the algorithm. In addition, Gaussian distribution initialization also improves the model's generalization ability for different lithium battery operating conditions, ensuring accurate prediction results in all situations.

[0294] Step 6.3: In the gold prospector migration phase, the chaotic inertia weight (CIW) and sine function learning factor strategy are used to update the position of the gold prospector:

[0295] Q new =Q+Ω(ep)·C(ep)·(Q * -Q) (37)

[0296] Where Q is the current position, Q * is the current optimal position, inertia weight (ep max is the maximum number of iterations), k is a constant used to control the intensity of chaotic behavior. Sine function learning factor

[0297] In the migration phase of the gold panning algorithm, the chaotic inertia weight (CIW) and sine function learning factor strategy are used to update the position of the explorer, which can significantly improve the performance of lithium battery temperature prediction. This strategy enhances the global search capability of the algorithm by introducing chaos theory, effectively avoiding the trap of local optimal solutions. At the same time, the dynamic adjustment of the sine function learning factor accelerates the convergence speed of the algorithm. The combination of these methods not only improves the accuracy and robustness of the prediction, but also enhances the adaptability of the algorithm to nonlinear and dynamic changes, thereby achieving a better global optimization effect in lithium battery temperature prediction.

[0298] Step 6.4: Gold Prospector Cooperation Phase:

[0299] In the cooperation stage, prospectors will cooperate with each other to improve efficiency. The update formula combining CIW and sine function is:

[0300]

[0301] In the formula, and are the positions of two randomly selected prospectors at the current iteration number ep.

[0302] Step 6.5: Gold prospector gold panning stage:

[0303]

[0304] In the formula, is the position of the randomly selected reference prospector at the current iteration number ep.

[0305] Gold prospectors are constantly moving to get more gold. In order to decide whether the prospector should stay in the original position or move to a new position, the fitness value is compared:

[0306] Basic fitness function:

[0307]

[0308] In the formula, G represents a dim-dimensional vector, G i Represents the i-th element of vector G, dim is the dimension.

[0309] Dynamic fitness function:

[0310] F(G,ep)=F0(G)·Ω(ep)·C(ep) (41)

[0311] Using the dynamic fitness function F(G,ep), we can compare the prospector's performance at the new position Q new and the fitness of the current position Q are Fitness new =F(Gnew ,ep) and Fitness current =F(G,ep). Based on the fitness comparison results, we can decide whether to update the prospector's position:

[0312]

[0313] If Fitness new >Fitness current , then accept the new position: Q(ep+1)=Q new ; If Fitness new <Fitness current , then accept the new position with a certain probability: Q(ep+1)=Q new .

[0314] Step 7: Figure 5 As shown, the parameters of GTimesNet are optimized using the improved gold mining optimizer algorithm IGRO.

[0315] Step 7.1: Find an objective function MSE of GTimesNet and randomly generate a set of hyperparameter combinations as the initial population:

[0316]

[0317] Where n GT is the sample size.

[0318] Step 7.2: Use the GTimesNet model to evaluate the performance of each set of hyperparameters on the validation set and calculate the fitness value, using Formula 38 as the fitness evaluation object. The smaller the MSE, the smaller the prediction error and the best performance.

[0319] Step 7.3: Update position: According to the update rule of the IGRO algorithm, combined with the chaotic inertia weight (CIW) and the sine function learning factor, update the position of each individual (hyperparameter combination) in the population:

[0320]

[0321] Where β is the inertia weight is the gradient of the objective function F at position Q.

[0322] Step 7.4: Based on the fitness value, select the hyperparameter combination with better performance and replace the combination with poor performance.

[0323] Step 7.5: Repeat steps (73) and (74) until the termination condition is met, such as reaching the maximum number of iterations or the fitness value converges.

[0324] The IGRO algorithm speeds up the convergence of the model, improves the generalization ability, and reduces the risk of overfitting, making the GTimesNet model more efficient and reliable in lithium battery temperature prediction. In addition, the IGRO algorithm helps the model converge to the optimal solution faster by searching the parameter space more efficiently, thereby achieving faster response time and higher prediction efficiency in lithium battery temperature prediction.

[0325] Step 8: Use IGRO to perform secondary weighting on the prediction results of the high-frequency component and the low-frequency component to obtain the final prediction result.

[0326] Step 8.1: Obtain the prediction errors of high-frequency and low-frequency components according to the mean square error formula:

[0327] According to formula 38, the error MSE of high-frequency component and low-frequency component is obtained High and MSE Low .

[0328] Step 8.2: Determine the weights:

[0329]

[0330] Where α1 is the weight of the high-frequency component and α2 is the weight of the low-frequency component.

[0331] Step 8.3: The calculated weight α is used to fuse the prediction results of high-frequency and low-frequency components:

[0332] Final Prediction=α1×Prediction High+α2×Prediction Low (47)

[0333] If α1 is close to 1, it means that the high-frequency component has a greater impact on the final prediction result; if α2 is close to 1, it means that the low-frequency component has a greater impact on the final prediction result.

[0334] The present invention also provides a lithium battery temperature prediction system, comprising the following modules: a lithium battery temperature related data collection module, which collects data including current, voltage, charging state, temperature, internal resistance, ambient temperature and aging degree; a correlation detection module, which performs in-depth feature selection on the original data of the lithium battery, and uses a regression model to identify and select the feature variables most relevant to the battery temperature change; a signal decomposition module, which decomposes the lithium battery data set, and divides the decomposed components into high-frequency components and low-frequency components through Hilbert transform; a lithium battery temperature prediction module, which replaces the Inception module in TimesNet with ConvGRU to obtain an improved GTimesNet model, and establishes a GTimesNet prediction model for lithium battery temperature prediction model; a lithium battery temperature prediction intelligent optimization module, which uses Gaussian Gaussian distribution initialization to initialize the population of the gold panning algorithm, adds chaotic inertia weights and sine function learning factor strategies in the exploration stage to obtain an improved gold panning optimizer algorithm IGRO, optimizes the hyperparameters of the GTimesNet model, and uses IGRO to perform secondary weighting on the prediction results of the high-frequency components and the low-frequency components to obtain the final lithium battery temperature prediction result.

[0335] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A lithium battery temperature prediction method, characterized in that: The following steps are involved: (1) Obtaining lithium battery temperature related data; (2) Conduct in-depth feature selection on the raw data of lithium batteries and use regression models to identify and select the feature variables most relevant to battery temperature changes; (3) Using the PSA algorithm to optimize TVF-EMD to obtain the improved PSA-TVF-EMD; (4) The lithium battery data set is decomposed using the improved PSA-TVF-EMD model, and the decomposed components are divided into high-frequency components and low-frequency components through Hilbert transform; (5) Use ConvGRU to replace the Inception module in TimesNet to obtain an improved GTimesNet model, and establish a GTimesNet prediction model to predict high-frequency components and low-frequency components; (6) Improve the gold panning optimization algorithm GRO: Use Gaussian distribution initialization to initialize the population of the gold panning algorithm, and add chaotic inertia weight CIW and sine function learning factor strategy in the exploration stage to obtain the improved gold panning optimizer algorithm IGRO; (7) Use the improved gold mining optimizer algorithm IGRO to optimize the parameters of GTimesNet; (8) Use IGRO to perform secondary weighting on the prediction results of the high-frequency component and the low-frequency component to obtain the final prediction result.

2. A lithium battery temperature prediction method according to claim 1, characterized in that: The implementation process of step (1) is as follows: Lithium battery temperature-related data include current, voltage, charge state, temperature, internal resistance, ambient temperature and aging degree; Create a data matrix with m sample data, each sample contains n feature data, forming the data matrix X of lithium battery: Where, X i is the parameter of the i-th lithium battery sample, X i (k) is the kth characteristic data parameter of the parameters of the i-th lithium battery sample.

3. A lithium battery temperature prediction method according to claim 1, characterized in that: The implementation process of step (2) is as follows: (21) Normalize the data: Where, X i '(k) is the kth indicator data in the temperature parameter of the i-th lithium battery sample after dimensionless conversion; i=0,1,...,m; k=0,1,2,3; AM is the mean of the data, σ is the standard deviation of the data; (22) Constructing the observation value vector where y1,y2…y m is the temperature of the lithium battery; (23) Minimize the residual sum of squares: In the formula, is the model prediction value, X is the lithium battery data set; (24) Taking the derivative of S(β) and setting it to zero, we obtain the normal equation: X T Xβ=X T Y (6) Solving the normal equations yields least squares estimates of the regression coefficients: In the formula, is the vector of regression coefficients; (25)Construct a linear regression model: Y=β0+β1X1+β2X2+…+β n X n +ò (8) Among them, β0 is the intercept term, β1, β2…β n is the regression coefficient corresponding to the respective variable, ò is the error term, which represents the random variation that cannot be explained by the model; (26) Calculate the coefficient of determination: In the formula, SS res is the residual sum of squares, that is SS tot is the total sum of squares, i.e. is the sample mean of the dependent variable Y; if the coefficient of determination R 2 The larger the value is, the greater the impact of this characteristic data on the temperature of the lithium battery.

4. A lithium battery temperature prediction method according to claim 1, characterized in that: The implementation process of step (3) is as follows: (31) Parameter encoding, where the bandwidth threshold and B-spline order are encoded as decision variables, and each particle in PSA represents a potential solution, i.e., a pair of parameter values; (32) Initialize the population, randomly initialize the particle swarm, and the position of each particle represents a parameter combination: In the formula, is the bandwidth threshold of the ith particle, is the B-spline order of the i-th particle; (33) Define the fitness function and calculate the initial fitness, using the signal-to-noise ratio (SNR) as the objective function: SNR=10log 10 (P signal / P noise ) (11) Where P signal is the signal power, P noise is the noise power; (34) Use PID control theory to update the particle position and system deviation: Δu(t)=K p ·r2·[e k (t)-e k-1 (t)]+K i ·r3·e k (t)+K d ·r4·[e k (t)-2e k-1 (t)+e k-2 (t)] (12) In the formula, K p , K i and K d are the proportional, integral and differential coefficients of the PID controller, r2, r3 and r4 are random number vectors in the interval [0,1]; The individual x is the minimum value in the overall history psa (t-1) is the particle position at the previous iteration; (35) Update the particle position according to the PID control value: x psa (t+1)=x psa (t)+η·Δu(t)+(1-η)·o(t) (14) Where η is the adjustment factor composed of random numbers, and o(t) is the zero output adjustment factor; (36) Check the termination condition and output the optimal solution. If the termination condition is met, stop the iteration; otherwise, continue the iteration; The bandwidth threshold and B-spline order combination that maximizes the fitness function is output, and the improved PSA-TVF-EMD decomposition model is obtained.

5. A lithium battery temperature prediction method according to claim 1, characterized in that: The implementation process of step (4) is as follows: (41) The original signal is preprocessed and the trend term and high-frequency noise are removed using TVF-EMD; (42) A time-varying filter is constructed to filter the signal and extract the intrinsic mode function of the signal until the remaining signal is a monotonic function; the time-varying filter h(t) is: The signal x(t) is decomposed into several IMFs and a residual term r(t): (43) Determine the local cutoff frequency using Hilbert transform: First, calculate the Hilbert transform of the signal x(t) Where H{x(t)} is the Hilbert transform of the signal x(t), and P represents the Cauchy principal value; Secondly, the local maximum sequence and local minimum sequence of the instantaneous amplitude are determined as: {A max (t)} and {A min (t)} Then, the local maximum and minimum sequences are interpolated to calculate the instantaneous mean and instantaneous envelope to obtain β1(t) and β2(t): β1(t)=Interpolation(A max (t)) β2(t)=Interpolation(A min (t)) Instantaneous mean α1(t): Instantaneous mean α2(t): Finally, calculate the instantaneous frequency component and calculate the local cutoff frequency: Instantaneous frequency component φ′1(t): Instantaneous frequency component φ′2(t): Where η1(t) and η2(t) are obtained by max (t)} and {A min (t)} interpolated; Local cutoff frequency ω′ bis (t): (44) Continuously iterate and extract IMF according to the stopping criterion: Instantaneous frequency: Where ω(t) is the instantaneous phase of the signal, which is obtained from the signal phase after Hilbert transformation; Loughlin instantaneous bandwidth: Where sgn is the sign function, which is used to determine the direction of frequency change; Weighted average instantaneous frequency: Where w(t) is the weight function; Local mean of instantaneous envelope: Where A(t) is the instantaneous amplitude; Stopping criterion θ(t): If θ(t) < ξ, where ξ is a preset threshold, it is considered that an IMF has been successfully extracted and the iteration stops; The expression of the original signal X(t) decomposed by TVF-EMD is: Where u(t) is the health status of the lithium battery, IMF i (t) is the i-th intrinsic mode function, R(t) is the residual term after decomposition, and N is the number of IMFs obtained by decomposition; (45) The decomposed signal is divided into high-frequency components and low-frequency components using Hilbert transform: Obtain the analytical signal. For each IMFu(t), its analytical signal z(t) is expressed as: z(t)=u(t)+j·H{u(t)}(30)where j is the imaginary unit and H{u(t)} is the Hilbert transform of u(t); The first bandpass filter is applied to the real part of the analytical signal to obtain the low-frequency components; the second bandpass filter is applied to the imaginary part of the analytical signal to obtain the high-frequency components.

6. A lithium battery temperature prediction method according to claim 1, characterized in that: The implementation process of step (5) is as follows: (51) The ConvGRU module will contain convolutional layers to process spatial data and GRU units to process time series data; in TimesNet, the Inception module is replaced by the ConvGRU module to ensure that the input and output dimensions of the ConvGRU module match those of the Inception module; according to the characteristics of the ConvGRU module, the parameters and activation functions of the convolutional layers in TimesNet are adjusted to obtain an improved GTimesNet model; (52) Use the GTimesNet model to predict high-frequency components and low-frequency components: Compute the improved update gate: In the formula, ug t is the output of the update gate is a low frequency signal (high frequency signal), W ug and U ug is the input of the convolution and acts on the current time step respectively and the hidden state H at the previous time step t-1 , b ug is the bias term, σ is the sigmoid activation function; Calculate the improved reset gate: rg t =σ(W rg *X t +U rg *H t-1 +b rg ) (32) In the formula, rg t is the output of the reset gate, b rg is the bias term; Compute candidate hidden states: In the formula, is the candidate hidden state after the activation function, b h is the bias term, tanh is the hyperbolic tangent activation function; Output predictions, final hidden state: In the formula, H t is the predicted low-frequency component or high-frequency component.

7. A lithium battery temperature prediction method according to claim 1, characterized in that: The implementation process of step (6) is as follows: (61) Set the population size and maximum number of iterations of the IGRO algorithm, as well as the upper and lower limits of the search space; (62) The position of the gold prospector is initialized using Gaussian distribution, which is defined as follows: Q ij =μ+σ·N(0,1) (35) In the formula, Q ij represents the position of the i-th individual in the j-th dimension; μ is the mean of the Gaussian distribution set to the current best gold mine position; σ is the standard deviation of the Gaussian distribution, a smaller σ means that the new position is closer to the current best position, while a larger σ allows the new position to explore a wider range; N(0,1) is a random number from a standard normal distribution; The positions of the gold prospectors are randomly initialized, and the position of each gold prospector is regarded as the approximate position of the gold mine: In the formula, Represents the mean A random vector drawn from a Gaussian distribution with a standard deviation of σ, ep is the current iteration number; (63) During the gold prospector migration phase, the chaotic inertia weight CIW and sine function learning factor strategy are used to update the position of the gold prospector: Q new =Q+Ω(ep)·C(ep)·(Q * -Q) (37) Where Q is the current position, Q * is the current optimal position, inertia weight ep max is the maximum number of iterations, k is a constant used to control the intensity of chaotic behavior; the sine function learning factor (64) Gold prospector cooperation stage: In the cooperation stage, prospectors will cooperate with each other to improve efficiency. The update formula combining CIW and sine function is: in, and are the positions of two randomly selected prospectors at the current iteration number ep; (65) Gold prospector gold panning stage: In the formula, is the position of the randomly selected reference prospector at the current iteration number ep; In order to decide whether the prospector should stay in the original position or move to a new position, the fitness value is compared; basic fitness function: In the formula, G represents a dim-dimensional vector, G i Represents the i-th element of vector G, dim is the dimension; Dynamic fitness function: F(G,ep)=F0(G)·Ω(ep)·C(ep) (41) Using the dynamic fitness function F(G,ep), compare the prospector's new position Q new and the fitness of the current position Q are Fitness new =F(G new ,ep) and Fitness current =F(G,ep); According to the fitness comparison result, decide whether to update the prospector's position: If Fitness new >Fitness current , then accept the new position: Q(ep+1)=Q new ; If Fitness new <Fitness current , then accept the new position with a certain probability: Q(ep+1)=Q new .

8. A lithium battery temperature prediction method according to claim 1, characterized in that: The implementation process of step (7) is as follows: (71) Find the objective function MSE of GTimesNet and randomly generate a set of hyperparameter combinations as the initial population: Where n GT is the sample size; (72) The GTimesNet model was used to evaluate the performance of each set of hyperparameters on the validation set and calculate the fitness value. The objective function MSE was used as the fitness evaluation object. The smaller the MSE, the smaller the prediction error and the best performance. (73) Update position: According to the update rule of the IGRO algorithm, combined with the chaotic inertia weight CIW and the sine function learning factor, the position of each individual in the population is updated: Where β is the inertia weight is the gradient of the objective function F at position Q; (74) According to the fitness value, select the hyperparameter combination with better performance and replace the combination with poor performance; (75) Repeat steps (73) and (74) until the termination condition is met, such as reaching the maximum number of iterations or the fitness value converges.

9. A lithium battery temperature prediction method according to claim 1, characterized in that: The implementation process of step (8) is as follows: In the formula, MSE High and MSE Low are the errors of high-frequency component and low-frequency component respectively, α1 is the weight of high-frequency component, and α2 is the weight of low-frequency component; The calculated weight α is used to fuse the prediction results of high-frequency and low-frequency components: Final Prediction=α1×Prediction High+α2×Prediction Low (47) If α1 is close to 1, it means that the high-frequency component has a greater impact on the final prediction result; if α2 is close to 1, it means that the low-frequency component has a greater impact on the final prediction result.

10. A lithium battery temperature prediction system using the method according to any one of claims 1 to 9, characterized in that: The system includes the following modules: Lithium battery temperature-related data collection module, which collects data including current, voltage, charging status, temperature, internal resistance, ambient temperature and aging degree; The correlation detection module performs in-depth feature selection on the raw data of lithium batteries and uses regression models to identify and select the feature variables most relevant to battery temperature changes; The signal decomposition module decomposes the lithium battery data set and divides the decomposed components into high-frequency components and low-frequency components through Hilbert transform; The lithium battery temperature prediction module uses ConvGRU to replace the Inception module in TimesNet to obtain an improved GTimesNet model, and establishes a GTimesNet prediction model for lithium battery temperature prediction model; The intelligent optimization module for lithium battery temperature prediction uses Gaussian distribution initialization to initialize the population of the gold panning algorithm. In the exploration phase, chaotic inertia weights and sinusoidal function learning factor strategies are added to obtain the improved gold panning optimizer algorithm IGRO. The hyperparameters of the GTimesNet model are optimized, and IGRO is used to perform secondary weighting on the prediction results of high-frequency components and low-frequency components to obtain the final lithium battery temperature prediction results.

Images