Battery SOC closed-loop prediction method and system based on cross-model collaborative optimization

Abstract

The invention discloses a battery SOC closed-loop prediction method and system based on cross-model collaborative optimization, the system adopts a data generation module, a dual-channel prediction module, a dynamic correction module and an optimization control module, the data generation module is used for generating expanded battery data by using a WGAN-GP generator based on a WGAN-GP network; the dual-channel prediction module is used for constructing an SOC prediction model according to the Transform-LSTM neural network; the dynamic correction module is used for activating an LSSVR compensation mechanism when a prediction error in the SOC prediction value is greater than a preset error threshold value; and the optimization control module is used for designing an enhanced whale optimization algorithm, establishing a three-level optimization space linkage mechanism through a parameter incidence matrix, and forming a generation-prediction-feedback closed-loop system. The three-level collaborative optimization mechanism provided by the invention plays a key role in improving the SOC prediction precision in a small sample scene.

Description

Technical Field

[0001] The present invention relates to the field of battery detection technology, and in particular discloses a battery SOC closed-loop prediction method and system based on cross-model collaborative optimization. Background Art

[0002] With the widespread application of lithium-ion batteries in electric vehicles, energy storage systems and other fields, accurate prediction of battery state of charge (SOC) is crucial to ensure battery safety and extend battery life. Existing SOC prediction methods are mainly divided into two categories: (1) Traditional model-driven methods: such as the extended Kalman filter (EKF) and equivalent circuit models (such as the second-order RC model), rely on accurate battery parameter calibration. However, in actual operating conditions, battery parameters are highly time-varying, which can easily lead to error accumulation.

[0003] (2) Data-driven methods: such as support vector regression (SVR) and long short-term memory (LSTM) networks, rely on a large amount of battery charging and discharging data. However, in actual scenarios, the cost of obtaining battery life cycle data is high and samples are scarce, resulting in insufficient model generalization capabilities.

[0004] The current technology has the following problems: (1) High data dependence: Machine learning models require a large amount of measured data support and are prone to overfitting in small sample scenarios.

[0005] (2) Difficulty in parameter tuning: The hyperparameters of models such as generative adversarial networks (GANs) and Transformer-LSTM are coupled with structural parameters, making manual tuning inefficient.

[0006] (3) Lack of closed-loop correction: Traditional open-loop prediction systems are difficult to dynamically correct model deviations, and long-term prediction accuracy is significantly reduced.

[0007] Therefore, the above-mentioned defects in the existing SOC prediction methods are technical problems that need to be solved urgently. Summary of the invention

[0008] The present invention provides a battery SOC closed-loop prediction method and system based on cross-model collaborative optimization, aiming to solve at least one defect existing in the above-mentioned prior art.

[0009] One aspect of the present invention relates to a battery SOC closed-loop prediction system based on cross-model collaborative optimization, comprising: A data generation module, used for generating extended battery data based on the WGAN-GP network using the WGAN-GP generator; The dual-channel prediction module is used to build an SOC prediction model based on the Transformer-LSTM neural network, combine the expanded battery data, and the voltage and current identification parameters output by the EKF algorithm to output the SOC prediction value; The dynamic correction module is used to activate the LSSVR compensation mechanism when the prediction error in the SOC prediction value is greater than the preset error threshold, and use the Gaussian kernel function to weight the SOC results of the EKF and Transformer-LSTM neural network; The optimization control module is used to design an enhanced whale optimization algorithm. It establishes a three-level optimization space linkage mechanism through the parameter association matrix, and performs cross-model collaborative optimization on the parameters of the WGAN-GP network, the structure of the Transformer-LSTM neural network, and the LSSVR kernel parameters to form a generation-prediction-feedback closed-loop system.

[0010] Furthermore, in the optimization control module, the optimization of the parameters of the WGAN-GP network includes the generator learning rate, the discriminator gradient penalty coefficient, and the number of network hidden layer nodes.

[0011] Furthermore, in the optimization control module, the optimization of the structure of the Transformer-LSTM neural network includes the dynamic matching rule of the number of attention heads and the optimization strategy of the number of LSTM layers, where: The dynamic matching rule of the number of attention heads is used to select the number of heads in proportion to the length of the input sequence. If the sequence length is less than the preset first number threshold, the first number is selected; if the sequence length is in the preset number threshold interval in the pre-examination, the second number is selected; if the sequence length is greater than the preset second number threshold, the third number is selected; the first number threshold and the second number threshold are within the number threshold interval and the first number threshold is less than the second number threshold; The LSTM layer number optimization strategy adaptively selects the number of layers based on data complexity. LSTM includes single-layer LSTM and multi-layer LSTM. The LSTM layer number optimization strategy is used to select the preset single-layer LSTM if the data complexity is identified as a simple working condition; if the data complexity is identified as a complex working condition, the preset multi-layer LSTM is selected.

[0012] Furthermore, in the optimization control module, the dynamic weighted fusion formula is used to perform cross-model collaborative optimization on the parameters of the WGAN-GP network, the structure of the Transformer-LSTM neural network, and the LSSVR kernel parameters. The dynamic weighted fusion formula is:

[0013] in, For the end SOC Fusion results, α is the adjustment factor, For LSSVR SOC value, For EKF SOC value.

[0014] Furthermore, in the optimization control module, a three-level optimization space is defined in the enhanced whale optimization algorithm, which includes the first optimization space, the second optimization space and the third optimization space. The first optimization space adopts the hyperparameter set of WGAN-GP, including the generator learning rate, the discriminator λ value and the network depth; the second optimization space adopts the structural parameter set of Transformer-LSTM, including the number of attention heads, the number of LSTM layers and the dropout rate; the third optimization space adopts the LSSVR kernel parameter set, including the Gaussian kernel bandwidth σ and the regularization coefficient C.

[0015] Furthermore, in the optimization control module, the fitness function is used to construct a comprehensive evaluation index, which is:

[0016] in, Fitness As a comprehensive evaluation indicator, MAE is the SOC prediction error, FID To generate data quality scores, Delay It takes time to reason about the model.

[0017] Furthermore, in the optimization control module, cross-model optimization is performed, with WGAN-GP parameters being optimized first until FID < 30; Transformer-LSTM and LSSVR parameters being optimized jointly until MAE < 2%; all parameters are fine-tuned globally, and the iteration is terminated when the fitness function is improved by < 1%.

[0018] Furthermore, in the optimization control module, the enhanced whale optimization algorithm is used to calibrate the initial SOC, and the error correction formula is:

[0019] in, for SOC Correction value, is the SOC prediction value, K is the correction factor, To actually measure the voltage, Estimate the voltage for the model.

[0020] Another aspect of the present invention relates to a battery SOC closed-loop prediction method based on cross-model collaborative optimization, which is applied to the above-mentioned battery SOC closed-loop prediction system based on cross-model collaborative optimization. The battery SOC closed-loop prediction method based on cross-model collaborative optimization includes: Data generation and preprocessing: Collect battery charge and discharge data and build a small sample data set. The battery charge and discharge data includes voltage, current, temperature and SOC labels; Time series prediction model training: Build an SOC prediction model based on the Transformer-LSTM neural network. The input of the SOC prediction model is the voltage / current time series data output by the RC equivalent circuit model. Use EWOA to optimize the number of attention heads, the number of LSTM layers, and the hidden layer dimension, and combine the EKF algorithm to calibrate the initial SOC value. Train the SOC prediction model until convergence and save the optimal weight parameters. Closed-loop correction and dynamic fusion: Real-time monitoring of the prediction error in the SOC prediction value in the Transformer-LSTM neural network. If the MAE is found to be greater than the preset error threshold, the LSSVR auxiliary predictor is triggered. The LSSVR model is trained based on historical data, and the final SOC fusion result is calculated using a dynamic weighted fusion formula. The final SOC fusion result is fed back to the WGAN-GP generation module to update the generated data distribution. Parameter collaborative optimization: define the EWOA optimization parameter range and constraints; iteratively optimize the three-stage parameter space, update the parameter association matrix weights, and output the global optimal solution; Model verification and deployment: Test the performance of the SOC prediction model in public data sets, calculate the MAE, RMSE and FID indicators; and deploy the SOC prediction model to embedded devices to output the SOC prediction results in real time.

[0021] Furthermore, the steps of parameter collaborative optimization include: Define three levels of optimization space; Establish parameter correlation matrix; Design a fitness function and use it to construct a comprehensive evaluation index. The comprehensive evaluation index is:

[0022] in, Fitness As a comprehensive evaluation indicator, MAE is the SOC prediction error, FID To generate data quality scores, Delay It takes time to reason about the model.

[0023] The beneficial effects achieved by the present invention are: The present invention provides a battery SOC closed-loop prediction method and system based on cross-model collaborative optimization. The system adopts a data generation module, a dual-channel prediction module, a dynamic correction module and an optimization control module. The data generation module is used for generating extended battery data based on a WGAN-GP network and using a WGAN-GP generator; the dual-channel prediction module is used for constructing an SOC prediction model according to a Transformer-LSTM neural network, combining the extended battery data and voltage and current identification parameters output by an EKF algorithm, and outputting an SOC prediction value; the dynamic correction module is used for activating an LSSVR compensation mechanism when a prediction error in the SOC prediction value is greater than a preset error threshold, and using a Gaussian kernel function to perform weight fusion on the SOC results of an EKF and a Transformer-LSTM neural network; the optimization control module is used for designing an enhanced whale optimization algorithm, establishing a three-level optimization space linkage mechanism through a parameter association matrix, performing cross-model collaborative optimization on the parameters of the WGAN-GP network, the structure of the Transformer-LSTM neural network and the LSSVR kernel parameters, and forming a generation-prediction-feedback closed-loop system. The battery SOC closed-loop prediction system based on cross-model collaborative optimization provided by the present invention solves the problems of low SOC prediction accuracy and poor model robustness in small sample scenarios by generating adversarial network data expansion, time series model fusion prediction and parameter dynamic optimization; after canceling the EWOA cross-model optimization, the SOC prediction accuracy (MAE) dropped by 135%, and the generated data quality (FID) deteriorated by 502%, which fully proves that the three-level collaborative optimization mechanism proposed by the present invention plays a key role in improving the SOC prediction accuracy in small sample scenarios. The dynamic optimization priority adjustment realized by the parameter association matrix is ​​the core innovation point to ensure the efficient operation of the "generation-prediction-feedback" closed-loop system. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The figure is a flow chart of an embodiment of a battery SOC closed-loop prediction method based on cross-model collaborative optimization according to the present invention. DETAILED DESCRIPTION

[0025] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0026] like Figure 1As shown, the first embodiment of the present invention proposes a battery SOC closed-loop prediction system based on cross-model collaborative optimization, and establishes a dynamic mapping relationship between the generator learning rate of WGAN-GP (Wasserstein GAN with Gradient Penalty, an improved model of generative adversarial network based on Wasserstein distance) and the number of Transformer attention heads through a parameter association matrix. Distance, an important indicator for evaluating the quality of the generated model)>50, triggers the exponential decay mechanism of the gradient penalty coefficient λ. The battery SOC closed-loop prediction system based on cross-model collaborative optimization includes a data generation module, a dual-channel prediction module, a dynamic correction module and an optimization control module. Among them, the data generation module is used to generate extended battery data based on the WGAN-GP network using the WGAN-GP generator; the dual-channel prediction module is used to build an SOC prediction model based on the Transformer-LSTM (a hybrid model that combines the advantages of Transformer and LSTM (long short-term memory network) to improve the performance when processing sequence data) neural network, and output the SOC prediction value by combining the extended battery data and the voltage and current identification parameters output by the EKF algorithm; the dynamic correction module is used to activate the LSSVR compensation mechanism when the prediction error in the SOC prediction value is greater than the preset error threshold, and use the Gaussian kernel function to correct the EKF (Extended Kalman Filter) error. The optimization control module is used to design an enhanced whale optimization algorithm, establish a three-level optimization space linkage mechanism through the parameter association matrix, and perform cross-model collaborative optimization on the parameters of the WGAN-GP network, the structure of the Transformer-LSTM neural network, and the kernel parameters of the LSSVR (least squares support vector regression machine) to form a generation-prediction-feedback closed-loop system.

[0027] This example provides a battery SOC closed-loop prediction system based on cross-model collaborative optimization, including: (1) Data enhancement and parameter optimization: Based on the WGAN-GP network, a small amount of measured battery data is distributedly learned to generate high-realism extended data. EWOA is used to optimize the learning rate (0.0001-0.01), gradient penalty coefficient λ (1-10), network depth (3-5 layers) and other parameters of the generator / discriminator. The quality of the generated data is evaluated by the Frechet distance (FID).

[0028] (2) Construction of time series prediction model: A Transformer-LSTM hybrid network is designed to process multi-scale time series features. The voltage / current data output by the second-order RC equivalent circuit model is used as input. The number of Transformer attention heads (2 / 4 / 8 heads), the number of LSTM layers (1-3 layers) and the hidden layer dimension (32-128 units) are jointly optimized through EWOA, and the initial SOC value is calibrated in combination with the EKF algorithm.

[0029] (3) Closed-loop correction and parameter collaborative optimization: When the Transformer-LSTM prediction error exceeds the threshold (such as MAE>1.5%), the LSSVR auxiliary predictor is started, the kernel function parameters (Gaussian kernel bandwidth σ: 0.1-1.0) are trained based on historical data, and the dynamic weight fusion strategy (weight coefficient α=0.4-0.6) is adopted to integrate the SOC prediction values ​​of LSSVR and EKF, and the fusion results are fed back to the WGAN-GP generation module to form a closed-loop optimization.

[0030] Through EWOA, a cross-model parameter association matrix is ​​established (priority: WGAN-GP parameter weight 50%, Transformer-LSTM parameter weight 35%, LSSVR parameter weight 15%), and the three-stage parameter space is optimized simultaneously.

[0031] Furthermore, in the battery SOC closed-loop prediction system based on cross-model collaborative optimization proposed in this embodiment, in the optimization control module, the optimization of the parameters of the WGAN-GP network includes the generator learning rate (1e-5~1e-3), the discriminator gradient penalty coefficient λ (0.1~10) and the number of network hidden layer nodes (64~512), and the generated data needs to be verified by the FID score to ensure distribution consistency with the real data.

[0032] Preferably, in the battery SOC closed-loop prediction system based on cross-model collaborative optimization proposed in this embodiment, in the optimization control module, the optimization of the structure of the Transformer-LSTM neural network includes a dynamic matching rule for the number of attention heads and an optimization strategy for the number of LSTM layers, wherein the dynamic matching rule for the number of attention heads is used to select the number of heads in proportion to the length of the input sequence, and if the sequence length is less than a preset first quantity threshold, the first number is selected; if the sequence length is in a preset quantity threshold interval in a preliminary examination, the second number is selected; if the sequence length is greater than a preset second quantity threshold, the third number is selected; the first quantity threshold and the second quantity threshold are within the quantity threshold interval and the first quantity threshold is less than the second quantity threshold.

[0033] For example, in the dynamic matching rule of the number of attention heads, the number of heads is selected in proportion to the length of the input sequence. If the sequence length is ≤50, 2 heads are selected; if the sequence length is 50-100, 4 heads are selected; if the sequence length is >100, 8 heads are selected.

[0034] The LSTM layer number optimization strategy adaptively selects the number of layers based on data complexity. LSTM includes single-layer LSTM and multi-layer LSTM. The LSTM layer number optimization strategy is used to select the preset single-layer LSTM if the data complexity is identified as a simple working condition; if the data complexity is identified as a complex working condition, the preset multi-layer LSTM is selected.

[0035] For example, in the LSTM layer optimization strategy, the number of layers is adaptively selected based on data complexity. If the working condition is simple, 1 layer is selected; if the working condition is complex, 2-3 layers are selected.

[0036] Furthermore, in the battery SOC closed-loop prediction system based on cross-model collaborative optimization proposed in this embodiment, in the optimization control module, a dynamic weighted fusion formula is used to perform cross-model collaborative optimization on the parameters of the WGAN-GP network, the structure of the Transformer-LSTM neural network, and the LSSVR kernel parameters. The dynamic weighted fusion formula is: (1) In formula (1), For the end SOC Fusion results, α is the adjustment factor, For LSSVR SOC value, For EKF SOC value.

[0037] Furthermore, in the battery SOC closed-loop prediction system based on cross-model collaborative optimization proposed in this embodiment, in the optimization control module, a three-level optimization space is defined in the enhanced whale optimization algorithm, and the three-level optimization space includes a first optimization space, a second optimization space and a third optimization space. The first optimization space adopts the hyperparameter set of WGAN-GP, including generator learning rate, discriminator λ value and network depth; the second optimization space adopts the structural parameter set of Transformer-LSTM, including the number of attention heads, the number of LSTM layers and the dropout rate; the third optimization space adopts the LSSVR kernel parameter set, including Gaussian kernel bandwidth σ and regularization coefficient C.

[0038] Furthermore, in the battery SOC closed-loop prediction system based on cross-model collaborative optimization proposed in this embodiment, in the optimization control module, a fitness function is used to construct a comprehensive evaluation index, and the comprehensive evaluation index is:

[0039] In formula (2), Fitness As a comprehensive evaluation indicator, MAE is the SOC prediction error, FID To generate data quality scores,Delay It takes time to reason about the model.

[0040] Furthermore, in the battery SOC closed-loop prediction system based on cross-model collaborative optimization proposed in this embodiment, cross-model optimization is performed in the optimization control module, and WGAN-GP parameters are optimized first until FID < 30; Transformer-LSTM and LSSVR parameters are jointly optimized until MAE < 2%; all parameters are globally fine-tuned, and the iteration is terminated when the fitness function is improved by < 1%.

[0041] Preferably, in the battery SOC closed-loop prediction system based on cross-model collaborative optimization proposed in this embodiment, in the optimization control module, an enhanced whale optimization algorithm is used to calibrate the initial SOC, and the error correction formula is: (3) In formula (3), for SOC Correction value, is the SOC prediction value, K is the correction factor, To actually measure the voltage, Estimate the voltage for the model.

[0042] The EWOA algorithm implementation steps include: Step 1: Define the three-level optimization space (1) The first optimization space: the hyperparameter set of WGAN-GP, including the generator learning rate, the discriminator lambda value, and the network depth; (2) Second optimization space: Transformer-LSTM structural parameter set, including the number of attention heads, the number of LSTM layers, and the dropout rate (0.1-0.5); (3) The third optimization space: LSSVR kernel parameter set, including Gaussian kernel bandwidth σ (0.1-5) and regularization coefficient C (1-100).

[0043] Step 2: Create a parameter correlation matrix

[0044] Step 3: Design fitness function

[0045] Among them, MAE is the SOC prediction error, FID is the quality score of the generated data, Delay It takes time to reason about the model.

[0046] Step 4: Perform cross-model optimization (1) Prioritize optimizing WGAN-GP parameters until FID < 30.

[0047] (2) Jointly optimize the Transformer-LSTM and LSSVR parameters until MAE < 2%.

[0048] (3) Globally fine-tune all parameters and terminate the iteration when the fitness function improves by less than 1%.

[0049] Parameter association matrix, the optimization priority rule is: (1) When the FID of the data generated by WGAN-GP is greater than 50, the parameters of the second and third optimization spaces are frozen.

[0050] (2) When the MAE of Transformer-LSTM is greater than 5%, LSSVR forced correction (α=0.5) is triggered.

[0051] (3) Perform full parameter space scanning optimization once every 24 hours.

[0052] A method for implementing a battery SOC closed-loop prediction system based on cross-model collaborative optimization includes the following steps: Step 1: Data preparation and generation (1) Collect voltage, current, and temperature data during the battery charging and discharging process to construct the original data set.

[0053] (2) WGAN-GP is used to generate augmented data that is 10 times the amount of the original data and merged into the training set after FID verification.

[0054] Step 2: Model initialization (1) A second-order RC equivalent circuit model is established and the model parameters are identified using the FFRLS algorithm.

[0055] (2) Initialize the Transformer-LSTM network: embedding dimension = 64, feed-forward layer dimension = 256.

[0056] (3) Configure the initial parameters of LSSVR: σ = 1, C = 10.

[0057] Step 3: Closed-loop prediction and optimization (1) Input battery data into the EKF algorithm in real time to obtain the SOC baseline value; (2) Run Transformer-LSTM synchronously to predict SOC, and activate LSSVR correction when |SOC_{NN}-SOC_{EKF}|>3%.

[0058] (3) At the end of each cycle, the system parameters are updated through EWOA, and the adjustment range of the generator learning rate does not exceed ±20%.

[0059] Step 4: Exception handling mechanism (1) When the MAE of predictions is greater than 5% for five consecutive times, clear the generated data and restart WGAN-GP training.

[0060] (2) When the number of LSSVR corrections accounts for more than 30%, the Transformer-LSTM network structure reconstruction is triggered.

[0061] The present invention relates to a battery SOC closed-loop prediction method based on cross-model collaborative optimization, which is applied to the above-mentioned battery SOC closed-loop prediction system based on cross-model collaborative optimization. The battery SOC closed-loop prediction method based on cross-model collaborative optimization includes: Step S100, data generation and preprocessing: collect battery charge and discharge data and construct a small sample data set. The battery charge and discharge data includes voltage, current, temperature and SOC label.

[0062] Collect battery charging and discharging data, including voltage, current, temperature and SOC tags, and build a small sample data set.

[0063] Initialize the WGAN-GP network, optimize the generator / discriminator parameters through EWOA, and generate an extended dataset.

[0064] The generated data is denoised by Kalman filtering and divided into training set, validation set and test set.

[0065] Step S200, timing prediction model training: construct an SOC prediction model based on the Transformer-LSTM neural network, and the input of the SOC prediction model is the voltage / current timing data output by the RC equivalent circuit model; use EWOA to optimize the number of attention heads, the number of LSTM layers and the hidden layer dimension, and combine the EKF algorithm to calibrate the initial SOC value; train the SOC prediction model until convergence, and save the optimal weight parameters.

[0066] The SOC prediction model is built based on the Transformer-LSTM network, and the input is the voltage / current time series data output by the RC equivalent circuit model.

[0067] EWOA is used to optimize the number of attention heads, the number of LSTM layers, and the hidden layer dimension, and the EKF algorithm is used to calibrate the initial SOC value.

[0068] Train the model until convergence and save the optimal weight parameters.

[0069] Step S300, closed-loop correction and dynamic fusion: monitor the prediction error in the SOC prediction value in the Transformer-LSTM neural network in real time. If the MAE is found to be greater than the preset error threshold, the LSSVR auxiliary predictor is triggered; the LSSVR model is trained based on historical data, and the final SOC fusion result is calculated using the dynamic weighted fusion formula. The final SOC fusion result is fed back to the WGAN-GP generation module to update the generated data distribution.

[0070] The Transformer-LSTM prediction error is monitored in real time, and the LSSVR auxiliary predictor is triggered when MAE>1.5%.

[0071] The LSSVR model is trained based on historical data, using a dynamic weighted fusion formula:

[0072] Among them, α is dynamically adjusted according to the prediction error.

[0073] The fusion results are fed back to the WGAN-GP generation module to update the generated data distribution.

[0074] Step S400, parameter collaborative optimization: define the EWOA optimization parameter range and constraints; iteratively optimize the three-stage parameter space, update the parameter association matrix weights, and output the global optimal solution.

[0075] (1) Define the EWOA optimization parameter range and constraints (e.g., number of network layers ≤ 5).

[0076] (2) Iteratively optimize the three-stage parameter space, update the parameter association matrix weights, and output the global optimal solution.

[0077] Step S500, model verification and deployment: test the performance of the SOC prediction model in a public data set, calculate the MAE, RMSE and FID indicators; and deploy the SOC prediction model to an embedded device to output the SOC prediction results in real time.

[0078] (1) Test the model performance in public datasets (such as the NASA battery dataset) and calculate the MAE, RMSE, and FID indicators.

[0079] (2) Deploy to embedded devices (such as BMS controllers) and output SOC prediction results in real time.

[0080] Furthermore, the battery SOC closed-loop prediction method based on cross-model collaborative optimization involved in this embodiment, step S400 includes: Step S410: define a three-level optimization space.

[0081] Step S420: Establish a parameter association matrix.

[0082] Step S430: fitness function, using the fitness function to construct a comprehensive evaluation index, the comprehensive evaluation index is:

[0083] in, Fitness As a comprehensive evaluation indicator, MAE is the SOC prediction error, FID To generate data quality scores, Delay It takes time to reason about the model.

[0084] The battery SOC closed-loop prediction system and method based on cross-model collaborative optimization provided by the present application are described below with specific embodiments: Embodiment 1: refer to Figure 1 ,This example provides a SOC closed-loop prediction system based on public battery data, ,which achieves high-precision SOC estimation through WGAN-GP data enhancement, Transformer-LSTM time series modeling and EWOA cross-model optimization.

[0085] Implementation steps: 1. Public data collection and preprocessing: Obtain voltage, current, temperature and SOC label data from NASA battery public data sets (such as B0005 and B0006 battery charging and discharging data) to build an initial sample library (sample size ≤ 300 groups).

[0086] The original data is normalized, and Kalman filtering is used to eliminate abnormal data (such as invalid samples where the current suddenly changes to 0).

[0087] 2. WGAN-GP data generation and parameter optimization: Initialize the WGAN-GP network: the generator is a 3-layer fully connected network (128 hidden layer nodes), and the discriminator is a 2-layer convolutional network (convolution kernel 5×5).

[0088] EWOA is used to optimize the generator learning rate (0.0005-0.005) and gradient penalty coefficient λ (5-15), and an extended data set (sample size ≥ 3000 groups) is generated based on public data.

[0089] Generated data verification: Calculate the Frechet distance (FID≤10) and KL divergence (≤0.05) between the generated data and the NASA data distribution, and use them for model training after verification.

[0090] 3. Transformer-LSTM prediction model training: Build a hybrid timing model: Transformer module: Embedding dimension 32, the number of attention heads is dynamically selected by EWOA (2 / 4 heads).

[0091] LSTM module: The number of stacking layers is optimized to 2, and the hidden layer dimension is 64.

[0092] The input data is the voltage and current sequence simulated by the second-order RC equivalent circuit model (calibrated based on public data parameters), and the output is the SOC prediction value.

[0093] The initial SOC is calibrated by the EKF algorithm, and the error correction formula is:

[0094] Among them, the Kalman gain K is dynamically updated according to the covariance matrix.

[0095] 4. Closed-loop correction and LSSVR fusion: The prediction error (MAE) of Transformer-LSTM is monitored in real time, and the LSSVR auxiliary predictor is triggered when MAE>1.2%.

[0096] Based on the CALCE battery data set, the LSSVR model is trained and the Gaussian kernel bandwidth σ is optimized (range 0.2-1.5). The dynamic fusion formula is: =0.5× +0.5×

[0097] The fusion results are fed back to the WGAN-GP generation module to update the generated data distribution (iterate once every 24 hours).

[0098] 5. EWOA cross-model parameter optimization: Define the parameter optimization priority matrix:

[0099] Fitness function:

[0100] Constraints: Model single inference delay ≤ 30ms.

[0101] Brief description of the process: This embodiment is based on NASA public data, and generates extended data through WGAN-GP to solve the problem of small sample training; the Transformer-LSTM model captures the voltage / current time series characteristics, and the EKF calibration improves the initial accuracy; when the prediction error exceeds the limit, the LSSVR is dynamically integrated and fed back to the generation module. Experiments show that the SOC prediction MAE is ≤0.8%, which is 50% higher than that of a single model, and the data generation efficiency is increased by 40%.

[0102] Embodiment 2: This embodiment provides a SOC prediction comparison experiment based on the traditional model-driven method. The second-order RC equivalent circuit model is implemented using the Matlab platform in combination with the FFRLS parameter identification and the EKF algorithm. The same NASA battery data set (B0005, B0006) as in Example 1 is used for SOC prediction to verify the performance advantages of the method proposed in the present invention.

[0103] Implementation steps: 1. Data preparation and preprocessing: Using the same NASA battery data set as in Example 1, the complete charge and discharge cycle data of the B0005 battery was selected.

[0104] Input parameters: voltage, current, temperature (keep the same data dimensions as in Example 1).

[0105] Data normalization: Z-score standardization is used to eliminate the impact of dimension.

[0106] 2. Model establishment and parameter identification: (1) Construct a second-order RC equivalent circuit model: Model structure: including ohmic internal resistance R0, two RC parallel branches (R1 / / C1, R2 / / C2) Parameter definition: : The terminal voltage of two RC parallel branches (unit: V).

[0107] I: Battery operating current (unit: A, discharge is positive, charge is negative).

[0108] : Ohm internal resistance (unit: Ω).

[0109] : Polarization resistance (unit: Ω).

[0110] : Polarization capacitance (unit: F).

[0111] OCV(SOC): open circuit voltage function (unit: V), SOC is the state of charge (0-100%).

[0112] Equation of state:

[0113]

[0114]

[0115] (2) FFRLS parameter online identification: Sampling period: 1 second.

[0116] The forgetting factor λ=0.98 (determined by grid search).

[0117] Parameter definition: θ: parameter vector to be identified .

[0118] λ: Forgetting factor (range 0.95~1.0), controls the decay rate of historical data K(k): Kalman gain matrix for the kth iteration.

[0119] φ(k): regression vector, consisting of voltage and current observations.

[0120] P(k): Covariance matrix, representing the uncertainty of parameter estimation.

[0121] V_t(k): The terminal voltage measurement value of the kth sampling point (unit: V).

[0122] Recursive formula:

[0123]

[0124]

[0125] 3.EKF algorithm implementation: Parameter definition: State variables: .

[0126] (1) SOC: State of charge (0-100%).

[0127] (2) : RC branch voltage (unit: V).

[0128] Observation equation: .

[0129] (1) h(X): observation function, corresponding to the second-order RC model equation.

[0130] (2) v: observation noise, which follows N(0,R) distribution.

[0131] Process noise covariance: .

[0132] (1) Corresponding to SOC, The state estimation noise The observation noise covariance is R=1e-4.

[0133] SOC-OCV relationship: polynomial fitting (same coefficients as in Example 1) was used.

[0134] Sixth-order polynomial fitting

[0135] Among them, ai is the fitting coefficient.

[0136] 4.SOC prediction process: Step 1: Update the RC model parameters every 10 seconds via FFRLS.

[0137] Step 2: Predict and correct the EKF state based on the current parameters.

[0138] Step 3: Output the SOC estimation value without setting a closed-loop correction mechanism.

[0139] Experimental results analysis: Test set MAE: 2.3% (187.5% higher than Example 1).

[0140] RMSE: 3.1% (210.5% higher than Example 1).

[0141] Parameter identification time: 4.2ms on average for a single iteration.

[0142] Maximum instantaneous error: 5.8% (occurs during high current discharge stage).

[0143] Error cause analysis: (1) Model mismatch problem: The fixed second-order RC structure is difficult to adapt to the nonlinear changes in parameters caused by battery aging.

[0144] (2) Data dependence: Data enhancement is not used, and small samples lead to the accumulation of FFRLS parameter identification errors.

[0145] (3) Lack of dynamic correction: The open-loop prediction system cannot compensate for model deviations.

[0146] Comparison conclusion: In Example 1 of the present invention, the amount of training samples is increased by 10 times through WGAN-GP data enhancement. Transformer-LSTM captures the dynamic characteristics that the RC model fails to characterize. Combined with the LSSVR closed-loop correction mechanism, MAE ≤ 0.8% is finally achieved, which is 187.5% higher than the traditional method, verifying the effectiveness of the cross-model collaborative optimization architecture.

[0147] Embodiment 3: This example uses the same NASA battery data set (B0005, B0006) as Example 1, but removes the EWOA optimization algorithm to verify the necessity of cross-model collaborative optimization.Figure 1 The data generation, prediction model and dynamic correction modules are added, but the intelligent parameter adjustment function of the optimization control module is cancelled.

[0148] Implementation steps: 1. Data generation module: (1) A fixed parameter WGAN-GP network (generator learning rate = 0.001, discriminator λ = 10, hidden layer nodes = 128) was used.

[0149] (2) A total of 3,000 data sets were generated, and no FID score verification was performed.

[0150] (3) Directly merge the original data and generated data as the training set.

[0151] 2. Prediction model construction: (1) Transformer-LSTM adopts a fixed structure: the number of attention heads = 2 (does not adjust with the sequence length), the number of LSTM layers = 1.

[0152] (2) Manually set network parameters: embedding dimension = 32, feedforward layer dimension = 128.

[0153] (3) The EKF algorithm maintains the same initial parameters as in Example 1 (not dynamically calibrated).

[0154] 3. Dynamic correction module: (1) The LSSVR kernel parameters are fixed as σ = 0.5 and C = 10 (not optimized).

[0155] (2) The weight fusion coefficient α is fixed to 0.5 (not adjusted dynamically with the error).

[0156] 4. Training and validation: (1) Use the same training set (3000 groups) and test set (NASA original test data).

[0157] (2) The training rounds are consistent with Example 1 (200 epochs).

[0158] (3) The closed-loop feedback mechanism is turned off, and the generation module does not receive feedback on the prediction results.

[0159] Comparison of experimental results:

[0160] Key Difference Analysis: Data generation quality deteriorates: Fixed λ value leads to insufficient gradient penalty and mode collapse in generated data (FID>50).

[0161] The maximum deviation of the generated voltage curve is up to 12 mV (only 4 mV in Example 1).

[0162] Suboptimal model structure: The fixed 2-head attention mechanism fails to capture long sequence dependencies (MAE increases by 47% for >50 steps).

[0163] Single-layer LSTM is insufficient in extracting features of complex working conditions (peak error reaches 3.5%).

[0164] Parameter solidification defects: Fixed α = 0.5 results in the inability to enhance the LSSVR correction weight when the error is large.

[0165] Unoptimized LSSVR kernel parameters cause systematic deviations (about 0.8%) in the fusion results.

[0166] Conclusion verification: After canceling the EWOA cross-model optimization, the SOC prediction accuracy (MAE) of this embodiment decreased by 135%, and the quality of generated data (FID) deteriorated by 502%, which fully proved that the three-level collaborative optimization mechanism proposed in this invention plays a key role in improving the SOC prediction accuracy in small sample scenarios. The dynamic optimization priority adjustment realized by the parameter association matrix is ​​the core innovation point to ensure the efficient operation of the "generation-prediction-feedback" closed-loop system.

[0167] The present invention provides a battery SOC closed-loop prediction method and system based on cross-model collaborative optimization, and the beneficial effects achieved are: 1. Data enhancement capability: WGAN-GP generates highly realistic data, reducing dependence on measured data and improving data generation efficiency by 40%.

[0168] 2. Multi-model fusion accuracy: Transformer-LSTM combined with EKF calibration, the SOC prediction mean absolute error (MAE) is less than 1.0%, which is 45% higher than that of a single model.

[0169] 3. Closed-loop adaptability: Dynamically integrate LSSVR and EKF prediction results to improve long-term prediction stability by 30%.

[0170] 4. Parameter optimization efficiency: EWOA's cross-model collaborative optimization strategy reduces parameter tuning time by 50%, making it suitable for embedded device deployment.

[0171] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A battery SOC closed-loop prediction system based on cross-model collaborative optimization, characterized in that: include: A data generation module, used for generating extended battery data based on the WGAN-GP network using the WGAN-GP generator; A dual-channel prediction module is used to construct an SOC prediction model based on a Transformer-LSTM neural network, and output a SOC prediction value by combining the expanded battery data and the voltage and current identification parameters output by the EKF algorithm; A dynamic correction module, used for activating the LSSVR compensation mechanism when the prediction error in the SOC prediction value is greater than a preset error threshold, and using a Gaussian kernel function to perform weight fusion on the SOC results of the EKF and the Transformer-LSTM neural network; The optimization control module is used to design an enhanced whale optimization algorithm, establish a three-level optimization space linkage mechanism through a parameter association matrix, perform cross-model collaborative optimization on the parameters of the WGAN-GP network, the structure of the Transformer-LSTM neural network, and the LSSVR kernel parameters, and form a generation-prediction-feedback closed-loop system.

2. The battery SOC closed-loop prediction system based on cross-model collaborative optimization according to claim 1, characterized in that: In the optimization control module, the optimization of the parameters of the WGAN-GP network includes the generator learning rate, the discriminator gradient penalty coefficient and the number of network hidden layer nodes.

3. The battery SOC closed-loop prediction system based on cross-model collaborative optimization according to claim 1, characterized in that: In the optimization control module, the optimization of the structure of the Transformer-LSTM neural network includes a dynamic matching rule of the number of attention heads and an optimization strategy of the number of LSTM layers, wherein: The dynamic matching rule of the number of attention heads is used to select the number of heads in proportion to the length of the input sequence. If the length of the sequence is less than a preset first number threshold, the first number is selected; if the length of the sequence is within a preset number threshold interval in a preliminary examination, the second number is selected; if the length of the sequence is greater than a preset second number threshold, the third number is selected; the first number threshold and the second number threshold are within the number threshold interval and the first number threshold is less than the second number threshold; The LSTM layer number optimization strategy adaptively selects the number of layers based on data complexity. LSTM includes single-layer LSTM and multi-layer LSTM. The LSTM layer number optimization strategy is used to select a preset single-layer LSTM if the data complexity is identified as a simple working condition; if the data complexity is identified as a complex working condition, then select a preset multi-layer LSTM.

4. The battery SOC closed-loop prediction system based on cross-model collaborative optimization according to claim 1, characterized in that: In the optimization control module, a dynamic weighted fusion formula is used to perform cross-model collaborative optimization on the parameters of the WGAN-GP network, the structure of the Transformer-LSTM neural network, and the LSSVR kernel parameters. The dynamic weighted fusion formula is: in, For the end SOC Fusion results, α is the adjustment factor, For LSSVR SOC value, For EKF SOC value.

5. The battery SOC closed-loop prediction system based on cross-model collaborative optimization according to claim 1, characterized in that: In the optimization control module, a three-level optimization space is defined in the enhanced whale optimization algorithm, and the three-level optimization space includes a first optimization space, a second optimization space and a third optimization space. The first optimization space adopts a hyperparameter set of WGAN-GP, including a generator learning rate, a discriminator λ value and a network depth; the second optimization space adopts a structural parameter set of Transformer-LSTM, including the number of attention heads, the number of LSTM layers and the dropout rate; the third optimization space adopts a LSSVR kernel parameter set, including a Gaussian kernel bandwidth σ and a regularization coefficient C.

6. The battery SOC closed-loop prediction system based on cross-model collaborative optimization according to claim 5, characterized in that: In the optimization control module, a fitness function is used to construct a comprehensive evaluation index, which is: in, Fitness As a comprehensive evaluation indicator, MAE is the SOC prediction error, FID To generate data quality scores, Delay It takes time to reason about the model.

7. The battery SOC closed-loop prediction system based on cross-model collaborative optimization according to claim 6, characterized in that: In the optimization control module, cross-model optimization is performed, and WGAN-GP parameters are optimized first until FID < 30; Transformer-LSTM and LSSVR parameters are jointly optimized until MAE < 2%; all parameters are fine-tuned globally, and the iteration is terminated when the fitness function is improved by < 1%.

8. The battery SOC closed-loop prediction system based on cross-model collaborative optimization according to claim 1, characterized in that: In the optimization control module, the enhanced whale optimization algorithm is used to calibrate the initial SOC, and the error correction formula is: in, for SOC Correction value, is the SOC prediction value, K is the correction factor, To actually measure the voltage, Estimate the voltage for the model.

9. A battery SOC closed-loop prediction method based on cross-model collaborative optimization, applied to the battery SOC closed-loop prediction system based on cross-model collaborative optimization as claimed in any one of claims 1 to 8, characterized in that: The battery SOC closed-loop prediction method based on cross-model collaborative optimization includes: Data generation and preprocessing: Collect battery charge and discharge data to build a small sample data set. The battery charge and discharge data includes voltage, current, temperature and SOC labels; Time series prediction model training: Build a SOC prediction model based on the Transformer-LSTM neural network, where the input of the SOC prediction model is the voltage / current time series data output by the RC equivalent circuit model; use EWOA to optimize the number of attention heads, the number of LSTM layers, and the hidden layer dimension, and combine the EKF algorithm to calibrate the initial SOC value; train the SOC prediction model until convergence, and save the optimal weight parameters; Closed-loop correction and dynamic fusion: Real-time monitoring of the prediction error in the SOC prediction value in the Transformer-LSTM neural network. If the MAE is found to be greater than the preset error threshold, the LSSVR auxiliary predictor is triggered. The LSSVR model is trained based on historical data, and the final SOC fusion result is calculated using a dynamic weighted fusion formula. The final SOC fusion result is fed back to the WGAN-GP generation module to update the generated data distribution. Parameter collaborative optimization: define the EWOA optimization parameter range and constraints; iteratively optimize the three-stage parameter space, update the parameter association matrix weights, and output the global optimal solution; Model verification and deployment: Test the performance of the SOC prediction model in a public dataset, calculate the MAE, RMSE and FID indicators; and deploy the SOC prediction model to embedded devices to output SOC prediction results in real time.

10. The battery SOC closed-loop prediction method based on cross-model collaborative optimization according to claim 9, characterized in that: The step of collaborative parameter optimization includes: Define three levels of optimization space; Establish parameter correlation matrix; Design a fitness function and use the fitness function to construct a comprehensive evaluation index, which is: in, Fitness As a comprehensive evaluation indicator, MAE is the SOC prediction error, FID To generate data quality scores, Delay It takes time to reason about the model.

Images