Adaptive Kalman filter and self-designed early stopping strategy optimized convolutional neural network for state of energy estimation of lithium-ion battery in complex temperature environment

Abstract

To achieve accurate State of Energy (SOE) estimation of Battery Management System (BMS), the Adaptive Kalman Filter and self-designed Early Stopping Optimized Convolutional Neural Network (AKF-ESCNN) is innovatively introduced. It is based on a self-designed Early Stopping (ES) strategy to optimize the training of Convolutional Neural Network (CNN) models, addressing the issue of network overfitting. By integrating Adaptive Kalman Filtering (AKF) for smoothing and filtering the network outputs, it reduces erroneous abrupt variations in results, ultimately achieving precise estimation of SOE. After different experimental data verification (5 °C, 10 °C and 25 °C), compared the loss values of model training. AKF-ESCNN model training accuracy is 10 % higher than CNN. In the whole temperature range of this paper, AKF-ESCNN also has a better performance. At cold −5 °C the Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) of AKF-ESCNN in the HPPC working condition are 0.268 % and 0.449 %, while the MAE and RMSE of CNN before optimization are 1.411 % and 1.973 %, and the estimation accuracy has been significantly improved. AKF-ESCNN provides a new way to solve the problems faced by data-driven SOE estimation of lithium-ion batteries.