Estimation of battery SOC using a combined approach of temporal convolutional networks and Unscented Kalman Filter
by Huipin Lin, Xiaoying Chen, Lei Zhang, Zhengyi Bao, Sai Tang In recent years, power batteries have been widely used in electric vehicles, and the evaluation of state of charge (SOC) is an important parameter in battery management systems. Therefore, in this paper, we propose a time-series convolutional network that employs extended convolution and residual concatenation to efficiently process time-series data with parallelism and flexibility, and combines it with Unscented Kalman Filter (UKF) to further improve the accuracy and reduce the output fluctuation, so as to estimate the state of charge of lithium-ion batteries. We conducted experiments using the University of Maryland’s Dynamic Stress Test (DST), US06 test, and Federal Urban Driving Scheme (FUDS) datasets, and compared the proposed method with Convolutional Neural Networks (CNNs), Long and Short-Term Memory Networks (LSTMs), and Gated Recurrent Units (GRUs). Experimental results demonstrate that the proposed framework achieves superior estimation accuracy and robustness. Specifically, the proposed method achieves mean absolute error (MAE) values of 1.305%, 1.470%, and 1.015% under the DST, US06, and FUDS conditions, respectively, with an average Root Mean Square Error (RMSE) of 1.566% and a MAE below 1.263%. Compared with existing deep learning models, the proposed method reduces the SOC estimation error by approximately 7.6%–39.6% under different driving conditions. These results verify the effectiveness and robustness of the proposed hybrid SOC estimation framework.