Abstract: The state-of-charge(SOC) estimation of lithium-ion battery modules is recognized as a critical internal state that significantly impacts battery performance and serves as the foundation for other state estimations in battery packs. However, the accuracy of SOC estimation is easily influenced by external factors such as temperature, and the inconsistency among batteries poses additional challenges for SOC estimation of cells within the pack. A method combining the back propagation neural network(BPNN) and the extended Kalman filter(EKF) algorithm is proposed for SOC estimation of battery packs. In this method, the terminal voltage of the "leader" battery at different temperatures is first estimated using BPNN based on the prior SOC. The estimated voltage is then compared with the measured voltage, and the posterior SOC estimation is completed using the EKF algorithm. Simultaneously, the output layer weights of the BPNN are updated using the gradient descent(GD) algorithm based on the voltage difference to accelerate algorithm convergence. Furthermore, a correction strategy is designed, and the random forest(RF) algorithm is utilized to adjust the SOC estimation of the "follower" batteries. Experimental results demonstrate that the proposed BPNN-EKF-GD-RF algorithm achieves accurate SOC estimation for battery packs under varying temperatures, with the SOC estimation error remaining within 2.5% at room temperature and the maximum SOC estimation error for cells in the pack not exceeding 3.2% under temperature variations. The research results provide a high-precision and low-complexity solution for SOC estimation of lithium-ion battery packs in complex environments.

Key words: SOC estimation, BP neural network, extended Kalman filter, gradient descent algorithm, random forest algorithm, lithium-ion battery pack