Lithium battery energy state estimation method and system based on multi-strategy weighted fusion

Abstract

The invention discloses a lithium battery energy state estimation method and system based on multi-strategy weighted fusion. The method comprises the steps that voltage, current, temperature and energy state data of a lithium ion battery in the charging and discharging process are collected; decomposing the voltage data by using a robust empirical mode decomposition technology, and screening out features of voltage, current, temperature and temperature with high correlation with an energy state by using a recursive feature elimination method RFE; dividing the training set into three training subsets; a time sequence dense encoder TiDE is used as a prediction model; the hyper-parameters of the TiDE model are optimized through an aurora optimization algorithm PLO, and the energy state of the lithium battery is predicted; a final energy state estimation result is obtained by combining a self-adaptive weighted fusion method with the generated prediction result; according to the method, multiple estimation is performed on the battery energy state from different working conditions and temperature conditions, and the reliability and accuracy of a prediction result are remarkably improved.

Description

Technical Field

[0001] The present invention relates to a method and system for estimating the state of energy of a lithium battery, and more particularly to a method and system for estimating the state of energy of a lithium battery based on multi-strategy weighted fusion. Background Art

[0002] As a new battery technology with high energy density, long life and environmental friendliness, lithium batteries have gradually replaced traditional batteries in the fields of electric vehicles, portable electronic devices and energy storage systems, and have been widely used. However, the state of energy of a lithium battery is of great significance for its safe operation, extending the battery life and optimizing the energy management system.

[0003] Traditional methods for estimating the state of energy, such as the ampere-hour integration method, although simple to implement, have many deficiencies in practical applications. First, the battery will age during actual use, resulting in a decline in its performance, thereby affecting the accuracy of the state of energy. Second, changes in temperature and charge-discharge conditions will cause dynamic changes in the internal parameters of the battery, and traditional estimation methods are difficult to accurately adapt to these changes, further reducing the estimation accuracy. In addition, the estimation of the state of energy under complex operating conditions also faces higher challenges, and traditional static analysis methods cannot cope with these changes, thus limiting their performance and effectiveness in practical applications. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a method for estimating the state of energy of a lithium battery based on multi-strategy weighted fusion to improve the accuracy and reliability of predicting the state of energy of a lithium battery. On the other hand, a system for estimating the state of energy of a lithium battery based on multi-strategy weighted fusion is provided.

[0005] Technical Solution: A method for estimating the state of energy of a lithium battery based on multi-strategy weighted fusion according to the present invention includes the following steps:

[0006] S1. Collect voltage, current, temperature and state of energy data of a lithium-ion battery during charge and discharge processes to construct a data set;

[0007] S2. Decompose the voltage data by using robust empirical mode decomposition (REMD), and screen out the features with high correlation with the state of energy in the voltage, current, temperature and temperature by using the recursive feature elimination (RFE) method;

[0008] S3. Divide the data set described in S1 into a training set and a test set, and further divide the training set into three training subsets according to the operating conditions and temperature conditions. The training subsets include a training subset under the same operating conditions and the same temperature conditions, a training subset under different operating conditions and the same temperature conditions, and a training subset under different operating conditions and different temperature conditions;

[0009] S4. Use the time series dense encoder TiDE as the prediction model, take the features described in S2 as the model input features, and use the state of energy of the lithium-ion battery as the output target for training;

[0010] S5. Use the aurora optimization algorithm PLO to optimize the hyperparameters of the TiDE model, and combine the three training subsets and the optimized TiDE model to predict the state of energy of the lithium battery;

[0011] S6. Obtain the final state of energy estimation result by combining the prediction results generated by the three training subsets through the adaptive weighted fusion method.

[0012] Preferably, in S1, data during the charge and discharge processes of the lithium-ion battery are collected under four temperature conditions of 0°C, 10°C, 25°C, and 40°C; each temperature includes data during the charge and discharge processes under three working conditions of the dynamic stress test condition DST, the highway driving condition US06, and the federal urban driving condition FUDS.

[0013] Preferably, the implementation process of the robust empirical mode decomposition REMD described in S2 is as follows:

[0014] Set the initial signal as the original voltage signal and initialize the original data;

[0015] Obtain the upper envelope and the lower envelope of the signal after screening and iteration, and calculate the average value of the upper and lower envelopes;

[0016] Calculate the difference between the signal and the average value of the envelope to obtain the signal obtained in the first iteration;

[0017] Judge whether the extracted one meets the conditions. If it does not meet the conditions, use the signal obtained in the first iteration as the new initial signal and repeat the above steps.

[0018] Preferably, the conditions are:

[0019] (1) The number of zero points is equal to the number of extreme points, or the difference between the number of zero points and the number of extreme points is less than 1;

[0020] (2) f(k - 2) < f(k - 1) and f(k - 1) < f(k), the screening process stops and returns the decomposition result of the (k - 2)th time;

[0021] If the conditions are met, stop the screening process; if the conditions are not met, continue to iterate until the maximum number of iterations is reached to obtain a new initial signal. Judge whether the initial signal is monotonic or constant. If not, repeat the above steps until the signal cannot be decomposed.

[0022] Preferably, the implementation process of the recursive feature elimination method RFE in S2 is as follows: Train a support vector machine model using all features and calculate the performance of the model through cross-validation; evaluate the importance of each feature for the model prediction result through the feature importance index of the support vector machine, remove the feature with the smallest importance, and update the feature subset; repeat the process of training-evaluation-elimination until the number of remaining features meets the predetermined conditions; select the feature subset that minimizes the root mean square error RMSE as the final feature selection result.

[0023] Preferably, the implementation process of the time series dense encoder TiDE in S4 is as follows:

[0024] Map the past and covariates of the time series into a dense representation of features through an encoder, and use a residual block to map the input signal at each time step into a low dimension;

[0025] As the input of the dense encoder, stack and flatten all past and future prediction covariates, concatenate them with the static attributes and the past of the time series, and map them into an embedding using an encoder containing multiple residual blocks;

[0026] The first decoding unit, like the encoder, is a stack of several residual blocks with the same hidden layer size, takes the output of the encoder as input, and maps it into a vector of size H×p, and reshapes the vector into a matrix;

[0027] Use a time decoder to generate the final prediction result. The time decoder is a residual block with an output size of 1, and maps the decoded vector and the projected covariates at the t-th horizontal time step;

[0028] Add a global residual connection to linearly map the final prediction result into a vector with the same size as the horizontal line.

[0029] Preferably, the optimization process of the aurora optimization algorithm PLO in S5 is as follows:

[0030] Determine the hyperparameters to be optimized. The hyperparameters include the number of layers of the encoder, the number of layers of the decoder, the size of the hidden layer, the learning rate, and the dropout rate, and initialize the particle swarm;

[0031] Calculate the fitness value of each particle under the current hyperparameter combination. The fitness value is the root mean square error of the TiDE model on the training set;

[0032] Update the positions of the particles, simulate the precession motion of charged particles in a magnetic field, and find the local optimal solution;

[0033] The chaotic collision between particles causes the PLO to leave the local optimal solution. In each iteration, compare the fitness values of all particles, update the global optimal solution, and find the optimal hyperparameter combination in the current iteration;

[0034] When the number of iterations reaches the maximum value or the fitness value converges, stop the iteration, output the hyperparameter combination corresponding to the global optimal solution, and complete the hyperparameter optimization of the TiDE model.

[0035] Preferably, the formula for updating the position of the particle is:

[0036] X new (i,j) = X(i,j) + r2 × (W1 × v(t) + W2 × A O );

[0037] A O = Levy(d) × (X avg (j) - X(i,j)) + L B + r1 × (U B × L B );

[0038]

[0039] Among them, X new (i,j) represents the position of the i-th particle in the j-th dimension after position update, X(i,j) represents the position of the i-th particle in the j-th dimension, r2 represents the interference of uncontrollable factors on the particle, taking [0,1], v(t) represents the rotary motion, A O represents the aurora elliptical trail, W1 represents the weight of the rotary motion, W2 represents the weight of the aurora elliptical walk; Levy(d) represents the Levy flight, X avg (j) represents the central position in the j-th dimension, r1 takes [0,1], C represents the integral constant, q represents the charge carried by the particle, B represents the geomagnetic field strength, m represents the mass, t represents the current number of iterations, and T represents the maximum number of iterations;

[0040] The mathematical model for PLO to leave the local optimal solution is as follows:

[0041] X new (i,j) = X(i,j) + sin(r3 × π) × (X(i,j) - X(a,j)), r4 < K and r5 < 0.05;

[0042]

[0043] Among them, X(a,j) represents any particle in the particle, r3, r4, and r5 represent random values, taking [0,1], and K represents the collision probability.

[0044] Preferably, the implementation process of the adaptive weighted fusion method in S6 is:

[0045] There are three results predicted by the model. A weight is assigned to each result, and the final predicted result is obtained by comprehensively considering the three predicted results. The formula is as follows:

[0046]

[0047] Among them, represents the predicted values of the three models, k1, k2, and k3 represent the weight coefficients, and y p represents the combined predicted value;

[0048] The absolute error of the prediction is: represents the fitting predicted value of the p-th sample of the i-th prediction model, and Y p represents the actual value of the p-th sample;

[0049] Let e p = y p - Y p ,

[0050] The weight coefficients k1, k2, and k3 are obtained by minimizing the sum of squared errors. Their expression is:

[0051]

[0052] The solution of the weight coefficients uses the following model:

[0053]

[0054] Among them, R T = [1,1], k i ≥0 (i = 1, 2);

[0055] Solving the above equation by the Lagrange multiplier method, we get:

[0056]

[0057] Substitute the weight coefficients k1, k2, and k3 into y p to obtain the predicted result of weighted fusion.

[0058] A lithium battery state of energy estimation system based on multi-strategy weighted fusion according to the present invention includes:

[0059] A data acquisition and processing module, which is used to collect voltage, current, temperature, and state of energy data of a lithium-ion battery during charge and discharge, construct a data set and divide it into a training set and a test set, and further divide the training set into three training subsets;

[0060] A feature extraction and selection module, which is used to perform robust empirical mode decomposition on the voltage signal, and combine the recursive feature elimination method to screen out features highly correlated with the state of energy;

[0061] A model construction and optimization module, which is used to construct a TiDE model according to the selected features and optimize the hyperparameters of the TiDE model by using the aurora optimization algorithm PLO;

[0062] A prediction and fusion module, which uses the optimized TiDE model to predict the energy state of lithium batteries in three training subsets, and uses an adaptive weighted fusion method to comprehensively combine the prediction results of the three training subsets to obtain the final energy state estimation result.

[0063] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: 1. By using the robust empirical mode decomposition REMD to decompose the voltage data and combining the recursive feature elimination method RFE to select the features highly correlated with the energy state, it can more accurately capture the energy state characteristics of lithium batteries under different working conditions and temperatures, thus improving the accuracy of predicting the energy state of lithium batteries; 2. When training the model, the data under different working conditions and temperature conditions are considered, which can better adapt to the complex working conditions and temperature changes in various practical applications. By adaptively weighting and fusing the prediction results of three different training subsets, the reliability of the prediction results can be improved; 3. The hyperparameters of the time series dense encoder TiDE model are optimized by using the aurora optimization algorithm PLO, so that the optimal hyperparameter combination can be found, further improving the prediction performance of the model. Description of the Drawings

[0064] Figure 1 It is a schematic flow chart of the lithium battery energy state estimation system of the present invention;

[0065] Figure 2 It is a schematic flow chart of the robust empirical mode decomposition of the present invention;

[0066] Figure 3 It is a schematic diagram of the structure of the time series dense encoder TiDE of the present invention. Detailed Embodiments

[0067] Next, in combination with the drawings, the technical solutions of the present invention will be described in detail.

[0068] S1. Collect the voltage, current, temperature and energy state data of the lithium-ion battery during the charging and discharging process to construct a data set.

[0069] S2. Use the robust empirical mode decomposition REMD to decompose the voltage data, and screen out the features highly correlated with the energy state among the voltage, current, temperature and temperature by using the recursive feature elimination method RFE.

[0070] Set the initial signal as the original voltage signal R j (n)=x(n), and set the maximum number of iterations as I max, initialize the original data:

[0071] h k (n)=R j (n); (1)

[0072] where h k (n) represents the signal after screening and iteration, and R j (n) represents the initial signal. k = 0, 1, 2,..., j = 1, 2,..., n represents the signal length;

[0073] Obtain the upper envelope g k (n) and the lower envelope g max (n) from h min (n);

[0074] Calculate the average value of the upper and lower envelopes:

[0075]

[0076] where m k (n) represents the average value of the upper and lower envelopes. Calculate the difference between the signal and the average value of the envelope to obtain h1(n):

[0077] h1(n)=h0(n)-m0(n); (3)

[0078] where h1(n) represents the signal obtained after the first iteration, h0(n) represents the initial signal, and m0(n) represents the average value of the upper and lower envelopes when k = 0;

[0079] Judge whether the extracted component R IMF meets the conditions. If it does not meet the conditions, use h1(n) as the new initial signal and repeat the above steps.

[0080] The definition of the f(k) condition is as follows:

[0081] f(k)=RMS k +|EK k |; (4)

[0082]

[0083] where g k (n) represents the component obtained after k iterations of the signal, represents the average value of g k (n), RMS k represents the root mean square of g k (n), EK k represents the excess peak of g k (n), and N s represents the signal length.

[0084] The screening process stops if the following conditions are met:

[0085] (1) The number of zeros (N zp ) and the number of extreme points (N ep ) are equal, or the difference between them is less than 1;

[0086] (2) f(k - 2) < f(k - 1) and f(k - 1) < f(k), the screening process stops and returns the decomposition result of the (k - 2)th time;

[0087] If not, continue to iterate until the maximum number of iterations is reached to obtain a new initial signal R j (n), and determine whether R j (n) is monotonic or constant. If not, repeat the above steps until the signal cannot be decomposed;

[0088]

[0089] wherein, represents the jth component.

[0090] The recursive feature elimination method evaluates the importance of features by repeatedly training a model, evaluating the importance of each feature, and gradually removing unimportant features, in combination with a support vector machine model:

[0091] First, train a support vector machine model using all features and calculate the performance of the model through cross - validation;

[0092] Evaluate the contribution of each feature to the model prediction result through the feature importance index of the support vector machine, remove the feature with the smallest contribution, and update the feature subset;

[0093] Repeat the process of training - evaluation - elimination until the number of remaining features meets the predetermined conditions;

[0094] Select the feature subset that minimizes the RMSE as the final feature selection result.

[0095] S3. Divide the data set into a training set and a test set. Further divide the training set into three training subsets according to three working conditions, namely the dynamic pressure test condition DST, the highway driving condition US06, and the federal urban driving condition FUDS, and four temperature conditions of 0°C, 10°C, 25°C, and 40°C. The three training subsets include the training subset with the same working condition and the same temperature condition, the training subset with different working conditions and the same temperature condition, and the training subset with different working conditions and different temperature conditions.

[0096] Taking the charge - discharge data at 25°C and under the dynamic pressure test condition DST as the test set, the three groups of training sets constructed are as follows:

[0097] (1) Under DST conditions, the charge and discharge data at temperatures of 0°C, 10°C, and 40°C are used as the training set;

[0098] (2) Under the condition of 25°C, the charge and discharge data of US06 conditions and FUDS conditions are used as the training set;

[0099] (3) The charge and discharge data of lithium batteries of the same model under DST conditions and 25°C are used as the training set.

[0100] S4. Use the time series dense encoder TiDE as the prediction model, use the features described in S2 as the model input features, and use the energy state of the lithium-ion battery as the output target for training.

[0101] The time series dense encoder TiDE is implemented as follows:

[0102] First, the encoder maps the past and covariates of the time series into a dense representation of features, and uses a residual block to map each time step of the into a low dimension;

[0103]

[0104] Among them, ResidualBlock represents the residual block, represents the input signal, represents the feature variable after mapping.

[0105] As the input of the dense encoder, stack and flatten all past and future prediction covariates, and concatenate them with the static attributes and the past of the time series. Use an encoder containing multiple residual blocks to map them into the embedding;

[0106]

[0107] Among them, Encoder represents the encoding process, y (i) represents the historical sequence; a (i) represents the static attribute, represents the output of the prediction covariate encoder;

[0108] The first decoding unit, like the encoder, is a stack of several residual blocks with the same hidden layer size. It takes the encoded e (i) as the input and maps it to a vector g of size H×p (i) , and then reshapes this vector into a matrix D (i) ∈R d×H , and the t-th column, that is, can be regarded as the decoding vector of the t-th time period in all horizontal lines of t∈[H].

[0109] The entire operation can be described as follows:

[0110] g (i) = Decoder(e (i) ); (10)

[0111] D (i) = Reshape(g (i) ); (11)

[0112] where Decoder represents the decoding process, Reshape represents reshaping; g (i) is the output of the decoder, and D (i) is the matrix after reshaping g (i) ;

[0113] The temporal decoder is used to generate the final prediction result. The temporal decoder is a residual block with an output size of 1, which maps the decoded vector and the projected covariates at the t-th horizontal time step, i.e.:

[0114]

[0115] where TemporalDecoder represents the temporal decoder, represents the decoded vector, represents the future covariates at the L + t time step, represents the final prediction result.

[0116] Finally, a global residual connection is added to linearly map to a vector with the same size as the horizontal line, and this vector is added to the prediction.

[0117] S5. Use the aurora optimization algorithm PLO to optimize the hyperparameters of the TiDE model, and combine the three training subsets and the optimized TiDE model to predict the state of charge of the lithium battery;

[0118] The process of using the aurora optimization algorithm PLO to optimize the hyperparameters of the TiDE model is as follows:

[0119] Determine the hyperparameters to be optimized, including the number of encoder layers, the number of decoder layers, the size of the hidden layer, the learning rate, and the dropout rate.

[0120] Initialize the particle swarm. Each particle of the algorithm represents a potential combination solution of hyperparameters, and the upper and lower limits of the particle swarm are determined according to the characteristics of the hyperparameters.

[0121]

[0122] Among them, U B represents the upper limit of the solution, and L B represents the lower limit of the solution; rand represents a random number between [0, 1].

[0123] Then, calculate the fitness value of each particle under the current hyperparameter combination. The fitness value is defined as the root mean square error of the TiDE model on the training set, and the goal is to minimize the root mean square error.

[0124] Update the position of the particle, simulating the precession motion of a charged particle in a magnetic field. Update the position of the particle according to the current position and velocity of the particle, as well as the simulated magnetic field direction and intensity, to find the local optimal solution;

[0125] The particle moves along an elliptical path to explore a new region of the solution space to enhance the global search ability of the algorithm. Combine the two strategies into PLO, and the proposed position update formula is as follows:

[0126] X new (i,j) = X(i,j) + r2 × (W1 × v(t) + W2 × A O ); (14)

[0127] Among them, X new (i,j) represents the position of the i-th particle in the j-th dimension after position update, X(i,j) represents the position of the i-th particle in the j-th dimension, r2 represents the interference of uncontrollable factors on the particle, taking [0, 1], v(t) represents the gyroscopic motion, and A O represents the auroral ellipse trail, W1 represents the weight of the gyroscopic motion, and W2 represents the weight of the auroral ellipse walk;

[0128] A O = Levy(d) × (X avg (j) - X(i,j)) + L B + r1 × (U B × L B ) / 2; (15)

[0129]

[0130] Levy(d) represents the Levy flight, X avg (j) represents the central position in the j-th dimension, r1 takes [0, 1], C represents the integral constant, q represents the charge carried by the particle, B represents the geomagnetic field intensity, m represents the mass, t represents the current iteration number, and T represents the maximum iteration number;

[0131] The chaotic collision between particles enables PLO to leave the local optimum. When these particles enter the atmosphere and the aurora converges within the ellipse, the collision occurs more frequently, resulting in the continuous change of the aurora shape.

[0132] The mathematical model is as follows:

[0133] X new (i,j) = X(i,j) + sin(r3×π)×(X(i,j) - X(a,j)), where r4 < K and r5 < 0.05; (19)

[0134]

[0135] Among them, X(a,j) represents any particle in the particle, r3, r4, and r5 represent random values, taking [0,1], and K represents the collision probability.

[0136] In each iteration, compare the fitness values of all particles and update the global optimal solution, that is, find the optimal hyperparameter combination in the current iteration; when the number of iterations reaches the maximum value or the fitness value converges, stop the iteration and output the hyperparameter combination corresponding to the global optimal solution, which is the optimal hyperparameter setting of the TiDE model.

[0137] S6. Combine the prediction results generated from the three training subsets in S3 through the adaptive weighted fusion method to obtain the final energy state estimation result;

[0138] There are three results predicted by the model. Assign a weight to each result and comprehensively consider the three prediction results to obtain the final prediction result. The formula is as follows:

[0139]

[0140] Among them, represents the predicted values of the three models, k1, k2, and k3 represent the weight coefficients, and y p represents the combined predicted value;

[0141] Suppose there are m prediction samples, represents the fitting predicted value of the pth sample of the ith prediction model, and Y p represents the actual value of the pth sample;

[0142] The absolute error of the prediction is: i = 1, 2, 3, let e p = y p - Y p ,

[0143] The weight coefficients k1, k2, and k3 are obtained by minimizing the sum of squared errors, and their expression is:

[0144]

[0145] The solution of the weight coefficients uses the following model:

[0146]

[0147] where, R T = [1, 1], k i ≥ 0 (i = 1, 2);

[0148] Solving the above equation using the Lagrange multiplier method, we get:

[0149]

[0150] Substitute the weight coefficients k1, k2, k3 obtained by solving Equation (23) into Equation (20) to obtain the predicted result of weighted fusion.

Claims

1. A method for estimating the state of energy of a lithium battery based on multi-strategy weighted fusion, characterized in that It includes the following steps: S1. Collect the voltage, current, temperature, and energy state data of the lithium-ion battery during charge and discharge processes to construct a data set; S2. Use the Robust Empirical Mode Decomposition (REMD) to decompose the voltage data, and screen out the features with high correlation with the energy state among voltage, current, temperature, and temperature through the Recursive Feature Elimination (RFE) method; S3. Divide the data set in S1 into a training set and a test set, and further divide the training set into three training subsets according to working conditions and temperature conditions. The training subsets include the training subset with the same working condition and the same temperature condition, the training subset with different working conditions and the same temperature condition, and the training subset with different working conditions and different temperature conditions; S4. Use the Time Series Dense Encoder (TiDE) as the prediction model, take the features in S2 as the model input features, and take the energy state of the lithium-ion battery as the output target for training; S5. Use the Polar Light Optimization (PLO) algorithm to optimize the hyperparameters of the TiDE model, and combine the three training subsets and the optimized TiDE model to predict the energy state of the lithium battery; S6. Combine the prediction results generated by the three training subsets through the adaptive weighted fusion method to obtain the final energy state estimation result.

2. The management method of the lithium battery material factory according to claim 1, wherein, S1 collects the data of the lithium-ion battery during charge and discharge processes under four temperature conditions of 0°C, 10°C, 25°C, and 40°C; each temperature includes the data of charge and discharge processes under three working condition conditions of the Dynamic Stress Test (DST), the US06 highway driving condition, and the Federal Urban Driving Schedule (FUDS).

3. The method for estimating the energy state of a lithium battery according to claim 1, wherein The implementation process of the Robust Empirical Mode Decomposition (REMD) in S2 is as follows: Set the initial signal as the original voltage signal and initialize the original data; Obtain the upper envelope and the lower envelope of the signal after screening and iteration, and calculate the average value of the upper and lower envelopes; Calculate the difference between the signal and the average value of the envelope to obtain the signal obtained in the first iteration; Judge whether the extracted component meets the conditions. If it does not meet the conditions, use the signal obtained in the first iteration as the new initial signal and repeat the above steps.

4. The method for estimating the energy state of a lithium battery according to claim 3, characterized in that, The conditions are as follows: (1) The number of zero points is equal to the number of extreme points, or the difference between the number of zero points and the number of extreme points is less than 1; (2) f(k - 2) < f(k - 1) and f(k - 1) < f(k), the screening process stops and returns the decomposition result of the (k - 2)th time; If the conditions are met, stop the screening process; If the conditions are not met, continue to iterate until the maximum number of iterations is reached, obtain a new initial signal, judge whether the initial signal is monotonic or constant. If not, repeat the above steps until the signal cannot be decomposed.

5. The method for estimating the energy state of a lithium battery according to claim 1, wherein The implementation process of the Recursive Feature Elimination (RFE) method in S2 is as follows: Use all features to train a support vector machine model, and calculate the performance of the model through cross-validation; evaluate the importance of each feature to the model prediction result through the feature importance index of the support vector machine, remove the feature with the smallest importance, and update the feature subset; Repeat the process of training - evaluation - elimination until the number of remaining features meets the predetermined conditions; select the feature subset that minimizes the Root Mean Square Error (RMSE) as the final feature selection result.

6. The method for estimating the energy state of a lithium battery according to claim 1, wherein The implementation process of the time series dense encoder TiDE described in S4 is as follows: The past and covariates of the time series are mapped to a dense representation of features through the encoder, and a residual block is used to map the input signal at each time step to a low dimension; As the input of the dense encoder, all past and future predictive covariates are stacked and flattened, concatenated with the static attributes and the past of the time series, and mapped into the embedding using an encoder containing multiple residual blocks; The first decoding unit, like the encoder, is a stack of several residual blocks with the same hidden layer size, takes the output of the encoder as the input, and maps it to a vector of size H×p, and reshapes the vector into a matrix; A time decoder is used to generate the final prediction result. The time decoder is a residual block with an output size of 1, and maps the decoded vector and the projected covariates at the t-th horizontal time step; A global residual connection is added to linearly map the final prediction result to a vector with the same size as the horizontal line.

7. The method for estimating the energy state of a lithium battery according to claim 1, wherein The optimization process of the aurora optimization algorithm PLO described in S5 is as follows: Determine the hyperparameters to be optimized, including the number of layers of the encoder, the number of layers of the decoder, the size of the hidden layer, the learning rate, and the dropout rate, and initialize the particle swarm; Calculate the fitness value of each particle under the current hyperparameter combination. The fitness value is the root mean square error of the TiDE model on the training set; Update the position of the particle, simulate the precession motion of charged particles in the magnetic field, and find the local optimal solution; The chaotic collision between particles causes PLO to leave the local optimal solution. In each iteration, compare the fitness values of all particles, update the global optimal solution, and find the optimal hyperparameter combination in the current iteration; When the number of iterations reaches the maximum value or the fitness value converges, stop the iteration, output the hyperparameter combination corresponding to the global optimal solution, and complete the hyperparameter optimization of the TiDE model.

8. The method for estimating the energy state of a lithium battery according to claim 7, wherein The formula for updating the position of the particle is as follows: X new (i,j) = X(i,j) + r2 × (W1 × v(t) + W2 × A O ); A O = Levy(d) × (X avg (j) - X(i, j)) + L B + r1 × (U B × L B ) / 2; Among them, X new (i,j) represents the position of the i-th particle in the j-th dimension after position update, X(i,j) represents the position of the i-th particle in the j-th dimension, r2 represents the interference of uncontrollable factors on the particle, taking [0,1], v(t) represents the rotary motion, A O represents the aurora ellipse path, W1 represents the weight of the rotary motion, W2 represents the weight of the aurora ellipse walk; Levy(d) represents the Levy flight, X avg (j) represents the central position in the j-th dimension, r1 takes [0,1], C represents the integral constant, q represents the charge carried by the particle, B represents the geomagnetic field strength, m represents the mass, t represents the number of the current iteration, T represents the maximum number of iterations; The mathematical model for PLO to leave the local optimal solution is as follows: X new (i,j) = X(i,j) + sin(r3 × π) × (X(i,j) - X(a,j)), r4 < K and r5 < 0.05; Among them, X(a,j) represents any particle in the particle, r3, r4, and r5 represent random values, taking [0,1], and K represents the collision probability.

9. The method for estimating the energy state of a lithium battery according to claim 1, wherein The implementation process of the adaptive weighted fusion method described in S6 is as follows: There are three results predicted by the model. A weight is assigned to each result, and the final prediction result is obtained by comprehensively considering the three prediction results. The formula is as follows: Among them, represents the predicted values of three models, k1, k2, and k3 represent weight coefficients, and y p represents the combined predicted value; The absolute prediction error is: i = 1, 2, 3, represents the fitted predicted value of the p-th sample of the i-th prediction model, Y p represents the actual value of the p-th sample; Let e p = y p -Y p , The weight coefficients k1, k2, and k3 are obtained by minimizing the sum of squared errors, and their expressions are: The solution of the weight coefficients uses the following model: wherein, R T = [1, 1], k i ≥ 0 (i = 1, 2); The above formula is solved by the Lagrange multiplier method to obtain: Substitute the weight coefficients k1, k2, and k3 into y p to obtain the prediction result of weighted fusion.

10. A lithium battery state of energy estimation system based on multi-strategy weighted fusion, characterized in that, Including: The data acquisition and processing module is used to collect the voltage, current, temperature, and energy state data of the lithium-ion battery during charging and discharging, construct a data set and divide it into a training set and a test set, and further divide the training set into three training subsets; The feature extraction and selection module is used to perform robust empirical mode decomposition on the voltage signal, and combine the recursive feature elimination method to screen out the features highly related to the energy state; The model construction and optimization module is used to construct a TiDE model according to the selected features, and optimize the hyperparameters of the TiDE model using the aurora optimization algorithm PLO; The prediction and fusion module uses the optimized TiDE model to predict the state of charge of lithium batteries in three training subsets, and uses the adaptive weighted fusion method to synthesize the prediction results of the three training subsets to obtain the final state of charge estimation result.

Images