Accurate state of charge (SOC) estimation is critical to ensure the reliable performance of lithium-ion batteries in electric vehicles (EV) and energy storage systems. However, challenges arise due to battery nonlinearity, parameter drift, and external disturbances. A novel SOC estimation framework is proposed which combines a dual-time-scale zonotopic Kalman filter (DZKF) with an optimizable convolutional long short term memory (LSTM) correction network. The DZKF performs real-time SOC estimation via an equivalent circuit model with adaptive resistance and capacity modeling under unknown-but-bounded (UBB) noise. The optimizable convolutional LSTM further refines the estimates by learning residual patterns. This approach integrates the robustness of physics-based models with the adaptability of deep learning. Validation on batteries under various temperatures and multiple dynamic operating conditions shows that the proposed method consistently achieves RMSEs below 1%, and experimental comparisons with various SOC estimation methods all show better superiority. • Dual-time-scale zonotopic Kalman filter for SOC is proposed. • Couples dual-time-scale zonotopic Kalman filter with convolutional LSTM to reduce uncertainty. • NMC and LFP validation: RMSE under 1% using only three driving cycles to train. • The proposed unknown-but-bounded estimation ensures robustness and tight SOC bounds.
Zhou et al. (Fri,) studied this question.