Research On Battery State Of Charge Estimation Based On Bi-directional Long Short-Term Memory Network

With the rapid development of electric vehicles,the lithium-ion battery and its corresponding battery management system have attracted more attention from researchers.Accurate estimation of battery state of charge(SOC)can provide the battery management system(BMS)with a clear understanding of the battery’s status,thus enabling the system to manage the battery reasonably and prevent reduced battery life due to overcharging and over-discharging,as well as safety issues caused by overheating.Therefore,this article focuses on the low accuracy of current battery SOC estimation,and aims to improve SOC estimation accuracy based on the Bi-directional Long Short-Term Memory network(Bi-LSTM),a deep learning model,from the perspectives of data optimization and model optimization.(1)Aiming at the low accuracy of battery state of charge estimation,this article uses the Bi-LSTM model,a variant of recurrent neural network,to estimate SOC.To ensure the practical significance of the research results,the dataset used in this study was obtained from charge-discharge experiments performed under various working conditions and temperatures.After constructing the Bi-LSTM model and dividing the processed battery data into training,validation,and test sets,the article compared the Bi-LSTM model with the RNN and LSTM models and verified that the Bi-LSTM model has a better SOC prediction effect.Subsequently,the article further investigated the influence of different hidden layer neuron numbers,hidden layer depths,learning rates,and input data sequence lengths on model performance.Finally,it was determined that the Bi-LSTM model has the best estimation effect with 128 hidden layer neurons,one hidden layer depth,a learning rate of 0.005,and a sequence length of 40.The model was also experimentally verified to be applicable to SOC prediction under other fixed and variable environmental temperatures.(2)Aiming at the problem of insufficient mining of raw data and small amount of raw data in Bi-LSTM model,two data optimization methods are proposed: feature selection and Gaussian data augmentation.First,the original data was subjected to feature engineering,a total of 14 features were produced,and then the Permutaion Importance(PI)method was used to select the 7 important features,and when the Bi-LSTM model was trained with these 7 features,the estimation accuracy of SOC was greatly improved,which proved the effectiveness of PI feature selection.Then,by adding Gaussian noise to the original data,more effective data are obtained to smooth the input space,which further improves the SOC estimation accuracy of the Bi-LSTM model,and proves the effectiveness of the Gaussian data augmentation method.(3)Aiming at the issue that Bi-LSTM is unable to utilize its advantages of using future information in battery state of charge(SOC)prediction,a Bi-LSTM-TDSE model based on delayed second-order prediction is proposed.This model can use past SOC estimation results to adjust the current SOC estimation results to obtain more accurate SOC estimates.This method also provides new ideas for time series estimation problems.In summary,this article addresses the issue of low battery state of charge prediction accuracy by utilizing feature selection,Gaussian data augmentation,and delayed second-order estimation methods in addition to the Bi-LSTM model.This provides an algorithmic foundation for precise battery management and has broad application prospects in the field of electric vehicles.