A method for predicting the remaining capacity of a vehicle-mounted lithium battery

Abstract

This invention discloses a method for predicting the remaining capacity of vehicle-mounted lithium batteries. The steps include: collecting battery data using a battery management system; analyzing the correlation of multidimensional features of battery sample data from multiple electric vehicles using the Pearson correlation coefficient algorithm on a cloud server to identify strong features that significantly impact battery capacity; standardizing the data; constructing an LSTM neural network prediction model based on the dimension of the input data; setting the parameters of a chaotic particle swarm optimization algorithm according to the neural network structure; and using the optimal parameters obtained from the chaotic particle swarm optimization algorithm for the LSTM neural network; finally, training and testing the neural network model to obtain an LSTM neural network model suitable for prediction. This invention utilizes the Pearson correlation coefficient algorithm for battery capacity-related feature analysis to identify strong features that significantly impact battery capacity, improving the model's accuracy; and employs an LSTM neural network model optimized by the chaotic particle swarm optimization algorithm for prediction, thus improving prediction precision.

Description

Technical Field

[0001] This invention belongs to the field of vehicle-mounted lithium battery capacity prediction, specifically relating to a method for predicting the remaining capacity of vehicle-mounted lithium batteries based on LSTM neural networks. Background Technology

[0002] Lithium-ion batteries for vehicles are characterized by high voltage, stable discharge, and long storage life, making them suitable for applications in vehicles and other fields with high energy requirements. However, under high-temperature environments, the battery's capacity characteristics will degrade to some extent, affecting its continued use. Therefore, to understand the battery's storage condition, it is necessary to predict its capacity to assess its performance.

[0003] Currently, battery capacity prediction is divided into two main methods: model-based prediction and data-driven prediction. Model-based prediction primarily uses the equivalent circuit model method, which treats the power source as an equivalent circuit model and estimates the battery capacity based on the power source's operating history and conditions. The equivalent circuit model prediction method is based on data about the power source's operating state; it first establishes an equivalent circuit model of the power source and then calculates the remaining lifespan, thus predicting battery capacity more accurately. However, this method has some limitations, such as the model's accuracy being highly dependent on the modeling accuracy, and the need for additional models to be built for predicting new types of power sources.

[0004] Data-driven approaches have become mainstream, primarily relying on historical data for model training. This involves analyzing and modeling historical battery data to predict battery capacity. Early methods mainly used statistical methods for lifespan prediction, such as particle filtering, which updated empirical formulas for battery capacity degradation using early cycle data. However, statistical methods for battery capacity prediction depend heavily on the choice of empirical formulas. If the battery capacity degradation curve deviates significantly from the trend of the empirical formula, the accuracy of lifespan prediction will be greatly reduced. Furthermore, statistical methods for battery capacity prediction are not well-suited to situations where the battery experiences jumps in capacity degradation.

[0005] With the gradual development of electric vehicle technology, the functions and applications of Battery Management Systems (BMS) have also been expanded. The earliest BMS systems only collected primary battery parameters (voltage, current, temperature, etc.), but later evolved to include the measurement and prediction of secondary parameters (SOC, internal resistance). This technology provides sufficient data for data-driven approaches.

[0006] Currently, data-driven methods primarily rely on deep learning, using historical datasets of onboard lithium battery operating status data, including parameters such as voltage and current, for model training. Deep learning models can automatically learn and extract features from the data, improving the model's generalization ability. Furthermore, they can capture the underlying features of the data, which is of great significance for understanding the working mechanism of batteries. Summary of the Invention

[0007] To improve the accuracy of predicting the remaining capacity of vehicle lithium batteries, this invention proposes a method for predicting battery remaining capacity based on deep learning optimized by swarm intelligence algorithms, which effectively improves the accuracy of battery capacity prediction.

[0008] To achieve the above objectives, the battery remaining capacity prediction method provided by the present invention includes the following steps:

[0009] Step 1: Utilize the battery management system to collect battery data. Parameters include battery voltage, battery current, battery temperature, remaining capacity, discharge time, maximum discharge voltage, maximum charging voltage, constant current time, and charging time for each battery cycle (a battery cycle refers to a complete charging and discharging process). This data is then uploaded to a cloud server. The cloud server uses the Pearson correlation coefficient algorithm to analyze the correlation of multidimensional features in the collected battery sample data from multiple electric vehicles, filtering out strong features that significantly impact battery capacity. Let the total number of batteries be N, and the number of battery cycles for the j-th battery sample be denoted as n. j The Pearson correlation coefficient of a characteristic data of the j-th battery is as follows:

[0010]

[0011] Then calculate the average Pearson correlation coefficient for this characteristic.

[0012]

[0013] In the formula, x i,j Let y be the capacity data of the j-th battery in the i-th battery cycle number. i,j For each characteristic, r is calculated as the sampled data value of the i-th battery in the j-th battery, representing the number of cycles. xy ;

[0014] Based on the Pearson correlation coefficient, the strong features that have a significant impact on battery capacity include: discharge time, maximum discharge voltage, maximum charging voltage, constant current time, and charging time.

[0015] Step 2: Standardize the sample data selected based on 5 features to obtain normalized data, and divide the data of N batteries into training set and test set according to the ratio.

[0016] Step 3: Set the initial input dimension of the LSTM neural network to 5 and the output dimension of the fully connected layer to 1 to build a neural network model for predicting the capacity of automotive lithium batteries.

[0017] Step 4: Optimize the number of neurons, number of layers, learning rate, and number of iterations of the LSTM neural network using the chaotic particle swarm optimization algorithm, and assign the obtained optimal hyperparameters to the LSTM neural network model.

[0018] Step 5: Train the LSTM neural network model optimized in Step 4 using the training set data from Step 2. Then, use the test set data to predict the battery capacity of the vehicle lithium battery. Finally, calculate the evaluation index. LSTM neural network models that pass the evaluation index are used for battery capacity prediction.

[0019] The method for optimizing the number of neurons, layers, learning rate, and number of iterations using the chaotic particle swarm optimization algorithm in step four is as follows:

[0020] Step 4.1 Initialize Particle Swarm: First, initialize a group of particles, each particle representing a solution, i.e., a combination of the number of neurons, the number of layers, the learning rate, and the number of iterations; introduce Tent chaotic mapping into the particle swarm algorithm, and use the randomness of chaotic variables to generate a uniformly distributed particle population in the solution space.

[0021] Step 4.2: Calculate the fitness value for each particle: that is, use the number of neurons, layers, learning rate and number of iterations corresponding to the particle to train the LSTM model, and calculate the root mean square prediction error of the training set as the fitness value.

[0022] Step 4.3: Update individual optimal position and global optimal position: Based on the fitness value of each particle, update the individual optimal position and global optimal position of each particle;

[0023] Step 4.4: Update particle position and velocity: Based on the individual optimal position and the global optimal position, as well as the current position and velocity, update the particle's position and velocity;

[0024] Step 4.5: Determine if the termination condition is met (the maximum number of iterations has been reached or the fitness value has converged to the set value). If the termination condition is not met, continue with steps 4.2-4.5 until the termination condition is met, and output the optimal solution.

[0025] In step 4.1, the Tent chaotic mapping formula is:

[0026]

[0027] In the formula, x n Let x represent the result of the nth iteration of the chaotic mapping. n+1 This represents the (n+1)th iteration result of the chaotic mapping, where a is the scale parameter, β is the nonlinear exponent, and μ is the result of the (n+1)th iteration. n It is a random variable that follows a standard normal distribution, ε n It is a noise term, a uniform random variable with values ​​between [0,1].

[0028] The Tent chaotic map above is used to generate the initial particle population positions for the particle swarm optimization algorithm. Leveraging the chaotic properties of the Tent chaotic map, the initial positions of the diffused particles in the search space are determined using the following formula:

[0029] pos ij =a + (ba)·x n+1

[0030] In the formula, pos ij Let be the j-th position of the i-th particle, and a and b be the lower and upper bounds of the search dimension.

[0031] In step 4.4, the chaotic particle swarm optimization algorithm uses an adaptive function to determine the value of the inertia weight, which is given by the following formula:

[0032]

[0033] The particle position update formula and velocity update formula of the algorithm are as follows:

[0034] x i (t+1)=x i (t)+v i (t+1)

[0035] v i (t+1)=ω i (t)·v i (t)+c1·r1·(p i (t)-x i (t))+c2·r2·(p g (t)-x i (t))

[0036] In the formula, v i (t) represents the velocity of particle i at time t, x i (t) represents the position of particle i at time t, p i (t) represents the optimal position of particle i at time t, p g (t) represents the global optimal position, ω i(t) represents the inertial weight at time t, c1 and c2 are acceleration factors, r1 and r2 represent random number factors, and ω max ω is the maximum inertia coefficient. min It is the minimum inertia coefficient.

[0037] The beneficial effects of this invention are:

[0038] 1. Considering the multi-dimensional characteristics of batteries, this invention uses the Pearson correlation coefficient algorithm to analyze the battery capacity-related features, find strong features that have a significant impact on battery capacity, avoid redundant data input, and thus improve the accuracy of the model.

[0039] 2. Normalize the raw data to eliminate the differences in parameters between different batteries and avoid the influence of specific battery data input, thereby improving the stability and versatility of the prediction model.

[0040] 3. This invention uses a chaotic particle swarm optimization algorithm to optimize the number of neurons, layers, learning rate, and iterations of an LSTM neural network. It introduces a Tent chaotic mapping improved by a normally distributed random variable to diffuse the initial position of particles in the search space and uses an adaptive function to improve the inertia weight to determine the value of the inertia weight, thereby improving the search capability of the particle swarm algorithm and thus improving the prediction accuracy of the prediction model. Attached Figure Description

[0041] Figure 1 This is a flowchart of the vehicle battery capacity prediction method proposed in this invention.

[0042] Figure 2 This invention relates to the battery management system (BMS) data acquisition framework.

[0043] Figure 3 The diagram shows the LSTM neural network structure constructed in this invention. Detailed Implementation

[0044] like Figure 1 As shown, the present invention generally includes the following steps:

[0045] Step 1: First, determine the correlation between the battery's various parameters and its capacity.

[0046] Utilize Figure 2 The battery management system collects battery data and then uploads it to the cloud server. Figure 2 Solid arrows represent data flow, and dashed arrows represent power flow. Cloud servers collect sample data from multiple electric vehicle batteries and use the Pearson correlation coefficient algorithm to analyze the correlation of multidimensional features in the battery data, filtering out strong features that significantly impact battery capacity.

[0047] Step 2: Standardize the sample data of the features selected in Step 1 to obtain normalized data, and divide the battery sample data into training set and test set according to the proportion.

[0048] Step 3: Based on the characteristics of the input data, set the input and output dimensions of the LSTM neural network.

[0049] Step 4: Use the chaotic particle swarm optimization algorithm to optimize the number of neurons, layers, learning rate, and number of iterations of the LSTM neural network, and assign the obtained optimal hyperparameters to the LSTM model.

[0050] Step 5: Train the LSTM prediction model using the training set data, validate the model's performance using the test set, and finally calculate the model's evaluation metric to obtain an LSTM neural network model that can be used for battery capacity prediction. During real-time data acquisition, in addition to updating the model, the predicted battery capacity results are also sent to the electric vehicle.

[0051] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0052] In step one, the Pearson correlation coefficient algorithm is used to analyze the correlation of multidimensional features of battery data, as shown in the following formula:

[0053]

[0054] In the formula, r is the Pearson correlation coefficient for a characteristic of the j-th battery. xy It is the average Pearson correlation coefficient of this characteristic, x i,j Let y be the capacity data of the j-th battery in the i-th battery cycle number. i,j Let r be the sampled data value of this characteristic for the cycle number of the i-th battery in the j-th battery. Calculate its r for each characteristic. xy For example, consider relevant characteristics such as battery voltage, battery current, battery temperature, remaining capacity, discharge time, maximum discharge voltage, maximum charging voltage, constant current time, and charging time.

[0055] Table 1 shows the correlation strength criteria corresponding to the Pearson correlation coefficient:

[0056] Table 1. Criteria for the degree of association of Pearson correlation coefficients

[0057]

[0058] The correlation coefficients between battery capacity and its related factors were calculated using the Pearson correlation coefficient algorithm. The correlation coefficient values ​​between each influencing factor (feature) and battery capacity in the embodiment are shown in the table below:

[0059] Table 2 Correlation coefficients between influencing factors and battery capacity

[0060]

[0061] Table 2 lists only the correlation coefficients between strongly and extremely strongly correlated influencing factors and battery capacity. Based on the correlation coefficients, influencing factors with higher coefficients are selected as input features for the model. In this invention, influencing factors with correlation coefficients greater than 0.6 are selected, including five feature sequences: discharge time (seconds), maximum discharge voltage (volts), maximum charging voltage (volts), constant current time (seconds), and charging time (seconds).

[0062] The discharge time refers to the time required for the battery to be fully discharged from a fully charged state.

[0063] Maximum discharge voltage refers to the maximum voltage value that a battery is allowed to withstand during discharge.

[0064] Maximum charging voltage refers to the maximum voltage value that a battery is allowed to withstand during charging.

[0065] Constant current time refers to the time during which a battery operates at a constant current during charging or discharging.

[0066] Charging time refers to the time required for a battery to go from an unloaded state to a fully charged state.

[0067] After filtering out the input features, in step two, an example of the historical data matrix P collected from one of the lithium batteries is shown below.

[0068]

[0069] In matrix P, each row represents data with the same number of cycles, and each column represents an influencing factor.

[0070] Step two involves standardizing the collected data to eliminate the impact of dimensional differences between the characteristic data of different battery influencing factors. The normalization formula is as follows:

[0071]

[0072] In the formula: P i,j This represents the element in the i-th row and j-th column of the historical data matrix. This represents the element in the i-th row and j-th column of the standardized historical data matrix. This represents the maximum value of the j-th feature of matrix P. Let represent the minimum value of the j-th feature of matrix P.

[0073] The normalized historical data is as follows.

[0074]

[0075] Based on the screening results from step one, the model input is determined to consist of five feature sequences: discharge time (seconds), maximum discharge voltage (volts), maximum charging voltage (volts), constant current time (seconds), and charging time (seconds). Therefore, in step three, we set the input dimension of the LSTM neural network to 5 and the output dimension of the fully connected layer to 1.

[0076] Step Four: This invention further optimizes the number of neurons, layers, learning rate, and iteration count of the LSTM neural network, specifically using a chaotic particle swarm optimization algorithm. The chaotic particle swarm optimization algorithm is an improvement upon the PSO (Particle Swarm Optimization) algorithm. Addressing the weaknesses of the PSO algorithm, such as uneven initial population distribution and imbalance between global and local search capabilities, a chaotic mapping is introduced to improve the algorithm. The method used in this invention to improve the number of neurons, layers, learning rate, and iteration count using the chaotic particle swarm optimization algorithm is as follows:

[0077] Step 1: Initialize the particle swarm.

[0078] First, a group of particles is initialized, with each particle representing a solution, i.e., a combination of the number of hidden layer neurons, the learning rate, and the number of iterations. Initially, the position and velocity of each particle can be randomly generated. This invention introduces the Tent chaotic mapping, utilizing the randomness of chaotic variables to generate a uniformly distributed particle population in the solution space. The improved Tent chaotic mapping formula, incorporating normally distributed random variables, can be expressed as:

[0079]

[0080] Where: x n Let x represent the result of the nth iteration of the chaotic mapping. n+1 This represents the (n+1)th iteration result of the chaotic mapping, where a is the scale parameter, β is the nonlinear exponent, and μ is the result of the (n+1)th iteration. n It is a random variable that follows a standard normal distribution, ε n This is the noise term, typically a uniform random variable with values ​​between [0, 1]. By using μ... n This term replaces the ordinary random uniform distribution between 0 and 1. At the same time, a noise term ε is added. n It can be used to quickly eliminate high-frequency noise in mapping.

[0081] The Tent chaotic map, improved using the aforementioned normally distributed random variables, is used to generate the initial particle population positions for the particle swarm optimization algorithm. Leveraging the chaotic properties of the Tent chaotic map, the initial positions of the diffused particles in the search space are determined using the following formula:

[0082] pos ij=a + (ba)·x n+1 (5)

[0083] In the formula, pos ij x is the j-th position of the i-th particle, and a and b are the lower and upper bounds of the search dimension. n+1 It is a Tent chaotic mapping formula improved from a normally distributed random variable.

[0084] Step 2: Calculate the fitness value.

[0085] The fitness value for each particle is calculated by using the number of neurons, layers, learning rate, and iterations corresponding to that particle to train the LSTM model, and the root mean square prediction error of the training set is calculated as the fitness value.

[0086] After setting the relevant parameters of the LSTM neural network, the objective function of the algorithm is determined to be the root mean square prediction error (RMSE) of the LSTM training set, as shown in the following formula.

[0087]

[0088] In the formula, T is the number of samples in the training set. t The actual values ​​in the training set. Given the predicted values ​​of the training set, the optimal number of neurons, layers, learning rate, and number of iterations for the LSTM neural network are obtained by minimizing the fitness function.

[0089] Step 3: Update the individual optimal position and the global optimal position.

[0090] Based on the fitness value of each particle, update the individual optimal position and the global optimal position of each particle. The individual optimal position is the position with the best fitness value in the history of that particle, and the global optimal position is the position with the best fitness value in the entire particle swarm in history.

[0091] Step 4: Update the position and velocity of the particles.

[0092] Based on the individual optimal position and the global optimal position, as well as the current position and velocity, update the particle's position and velocity. This step can be performed using the update formula of the standard PSO algorithm, as shown in equations (8) and (9), which include the calculation of velocity and position updates.

[0093] For the inertia weight in the particle swarm optimization algorithm, an adaptive function is used to improve the inertia weight and determine its value. The formula for the inertia weight improved by the adaptive function is as follows:

[0094]

[0095] The improved particle swarm optimization algorithm's particle position update formula and velocity update formula are shown in equations (8) and (9):

[0096] x i (t+1)=x i (t)+v i (t+1) (8)

[0097] v i (t+1)=ω i (t)·v i (t)+c1·r1·(p i (t)-x i (t))+c2·r2·(p g (t)-x i (t)) (9)

[0098] In the formula, v i (t) represents the velocity of particle i at time t, v i (t+1) represents the velocity of particle i at time t+1, x i (t) represents the position of particle i at time t, x i (t+1) represents the position of particle i at time t+1, p i (t) represents the optimal position of particle i at time t, p g (t) represents the global optimal position, ω i (t) represents the inertial weight at time t, c1 and c2 are acceleration factors, r1 and r2 represent random number factors, and ω max ω is the maximum inertia coefficient. min It is the minimum inertia coefficient.

[0099] Step 5: Determine the termination condition.

[0100] Determine if the termination condition is met, such as reaching the maximum number of iterations or the fitness value converging to a certain level. If the termination condition is not yet met, continue with Steps 2-5 until the termination condition is met.

[0101] The parameters of the chaotic particle swarm optimization algorithm include population size and learning factor, and the selection of these parameters affects the model's prediction results. In the original PSO algorithm, the learning factors c1 and c2 are generally set to 2.0 or 1.49, depending on the algorithm's search requirements. In this embodiment, the selection of the learning factor and population size is based on simulation experiments, choosing the set of parameters with the smallest error. In this example, the learning factor is selected as 1.49, and the population size is set to 10.

[0102] The optimization results from step four are used as parameters for the LSTM neural network to construct an LSTM prediction model. In step two, the normalized data of N batteries are divided into training and test sets according to a ratio. In step five, the LSTM prediction model is trained based on the training set, and finally, the trained LSTM prediction model is tested using the test set to obtain the trained LSTM neural network model for predicting the capacity of automotive lithium batteries.

[0103] In this embodiment, the battery capacity of the vehicle-mounted lithium battery is predicted and analyzed. To verify the effectiveness of the present invention, the prediction model proposed in this invention is compared with the prediction model without the improved algorithm and the original model. Specifically, the prediction models include: the unimproved LSTM neural network (first row of Table 3), the LSTM neural network optimized by the standard PSO algorithm (second row of Table 3), and the LSTM neural network optimized by the chaotic particle swarm optimization algorithm of this invention (third row of Table 3). The evaluation absolute error (MAE) and mean absolute percentage error (MAPE) are used as evaluation indicators, and the calculation formulas are as follows:

[0104]

[0105] The battery capacity prediction error is shown in the table below:

[0106] Table 3 Prediction Error

[0107]

[0108] As can be seen, the prediction model of this invention performs random perturbation of particles before optimizing the model parameters, which improves the distribution of randomness and performs chaotic perturbation on its inertial weights, making them exhibit a change from large to small. The resulting solution has a significant advantage over the standard PSO, and the model prediction effect is better.

[0109] In summary, this invention improves and optimizes the particle swarm optimization algorithm by using chaotic back-learning and chaotic mapping, and uses it to adjust the number of neurons, layers, learning rate, and iterations of the LSTM neural network, thereby improving the solution capability, accelerating the convergence speed, and improving the prediction accuracy.

Claims

1. A method for predicting the remaining capacity of a vehicle-mounted lithium battery, characterized in that, Includes the following steps: Step 1: Utilize the battery management system to collect battery data, including battery voltage, battery current, battery temperature, remaining capacity, discharge time, maximum discharge voltage, maximum charging voltage, constant current time, and charging time for each battery cycle. Upload this data to a cloud server. The cloud server uses the Pearson correlation coefficient algorithm to analyze the correlation of multidimensional features of the collected sample data from multiple electric vehicle batteries, filtering out strong features that significantly impact battery capacity. The total number of batteries is N, and the number of battery cycles for the j-th battery sample data is denoted as n. j The Pearson correlation coefficient of a characteristic data of the j-th battery is as follows: Then calculate the average Pearson correlation coefficient for this characteristic. In the formula, x i,j Let y be the capacity data of the j-th battery in the i-th battery cycle number. i,j For each characteristic, r is calculated as the sampled data value of the i-th battery in the j-th battery, representing the number of cycles. xy ; Based on the Pearson correlation coefficient, the strong features that have a significant impact on battery capacity include: discharge time, maximum discharge voltage, maximum charging voltage, constant current time, and charging time. Step 2: Standardize the sample data selected based on 5 features to obtain normalized data, and divide the data of N batteries into training set and test set according to the ratio. Step 3: Set the initial input dimension of the LSTM neural network to 5 and the output dimension of the fully connected layer to 1 to construct a neural network model for predicting the remaining capacity of vehicle lithium batteries. Step 4: Optimize the number of neurons, number of layers, learning rate, and number of iterations of the LSTM neural network using the chaotic particle swarm optimization algorithm, and assign the obtained optimal hyperparameters to the LSTM neural network model. Step 5: Train the LSTM neural network model optimized in Step 4 using the training set data from Step 2. Then, use the test set data to predict the battery capacity of the vehicle lithium battery. Finally, calculate the evaluation index. LSTM neural network models that pass the evaluation index are used for battery capacity prediction. The method for optimizing the number of neurons, layers, learning rate, and number of iterations using the chaotic particle swarm optimization algorithm in step four is as follows: Step 4.1 Initialize Particle Swarm: First, initialize a group of particles, each particle representing a solution, i.e., a combination of the number of neurons, the number of layers, the learning rate, and the number of iterations; introduce Tent chaotic mapping into the particle swarm algorithm, and use the randomness of chaotic variables to generate a uniformly distributed particle population in the solution space. Step 4.2: Calculate the fitness value for each particle: that is, use the number of neurons, layers, learning rate and number of iterations corresponding to the particle to train the LSTM model, and calculate the root mean square prediction error of the training set as the fitness value. Step 4.3: Update individual optimal position and global optimal position: Based on the fitness value of each particle, update the individual optimal position and global optimal position of each particle; Step 4.4: Update particle position and velocity: Based on the individual optimal position and the global optimal position, as well as the current position and velocity, update the particle's position and velocity; Step 4.5: Determine if the termination condition is met. If the termination condition is not met, continue with steps 4.2-4.5 until the termination condition is met, and output the optimal solution. In step 4.1, the Tent chaotic mapping formula is: In the formula, x n Let x represent the result of the nth iteration of the chaotic mapping. n+1 This represents the (n+1)th iteration result of the chaotic mapping, where a is the scale parameter, β is the nonlinear exponent, and μ is the result of the (n+1)th iteration. n It is a random variable that follows a standard normal distribution, ε n It is a noise term, a uniform random variable with values ​​between [0,1]. The Tent chaotic map above is used to generate the initial particle population positions for the particle swarm optimization algorithm. Leveraging the chaotic properties of the Tent chaotic map, the initial positions of the diffused particles in the search space are determined using the following formula: pos ij =a+(b-a)·x n+1 In the formula, pos ij Let be the j-th position of the i-th particle, and a and b be the lower and upper bounds of the search dimension.

2. The method for predicting the remaining capacity of an on-board lithium battery as described in claim 1, characterized in that, In step 4.4, the chaotic particle swarm optimization algorithm uses an adaptive function to determine the value of the inertia weight, which is given by the following formula: The particle position update formula and velocity update formula of the algorithm are as follows: x i (t+1)=x i (t)+v i (t+1) v i (t+1)=ω i (t)·v i (t)+c1·r1·(p i (t)-x i (t))+c2·r2·(p g (t)-x i (t)) In the formula, v i (t) represents the velocity of particle i at time t, x i (t) represents the position of particle i at time t, p i (t) represents the optimal position of particle i at time t, p g (t) represents the global optimal position, ω i (t) represents the inertial weight at time t, c1 and c2 are acceleration factors, r1 and r2 represent random number factors, and ω max ω is the maximum inertia coefficient. min It is the minimum inertia coefficient.

3. The method for predicting the remaining capacity of an on-board lithium battery as described in claim 1, characterized in that, The termination condition described in step 4.5 is reaching the maximum number of iterations or the fitness value has converged to a set value.

Images