Online Estimation of Model Parameters and State of Charge for Lithium-Ion Battery Using Multitimescale Recurrent Neural Networks
Zepei Zhang, Yuan Fan, Jiaqiang Tian, Huyong Kuang, Mince Li, Tianhong Pan, Yujie Wang, Xinghua Liu
Abstract
Accurate estimation of battery parameters and state is significant for new energy electric vehicles. In this work, a novel online estimation method is proposed for a second-order equivalent circuit model (ECM). The proposed online estimation method develops a dual recurrent neural networks (RNNs), one RNN updates model parameters on a long time scale, and an adaptive RNN (ARNN) estimates the state of charge (SOC) on a short time scale. The RNN-based method is proved to be convergent for battery state estimation, when the learning rate is suitable. To obtain the learning rate, the adaptive gradient (Adagrad) algorithm is employed to optimize the RNN-based parameter identification method. The ARNN algorithm utilizes the fuzzy controller to adaptively adjust the learning rate. The RNN–ARNN online estimation method is verified with different types of batteries under three different conditions, including the federal urban driving schedule (FUDS), dynamic stress test (DST), and US06 highway driving schedule. The filter-based and deep learning-based SOC estimation methods are utilized to make a comparison, including the recursive least square (RLS)–unscented Kalman filter (UKF), extended Kalman filter (EKF)–EKF, and gated RNN-based method. Under the conditions of FDUS, DST, and US06, RNN–ARNN can improve the voltage estimation accuracy of the model by 8.8%, 1.8%, and 24.9%. The convergence of RNN–ARNN method can be evidenced by the relative error of SOC less than 4% after 14 s. Compared with the gated RNN-based method, the proposed method can save at least 66% of calculation time and improve the accuracy by 30%.