Hybrid data-driven deep learning model for state of charge estimation of Li-ion battery in an electric vehicle

This study develops a knowledge-assisted data-driven method for state of charge (SoC) estimation of Li-ion batteries to address the limits of conventional model-based and data-driven approaches which require precise parameters and large data, respectively. The proposed method minimizes data dependence by combining parameter identification of electrochemical-thermal models through genetic algorithms with SoC estimation using deep learning models trained on simulation data. In addition, the unscented Kalman filter was then employed to remove noise from the estimates. The performance of the proposed model was analyzed using the estimation accuracy in various environments, such as different operating temperatures, driving patterns, and data scarcity levels. To illustrate the capability of the proposed method, the accuracy of the SoC estimation was compared to conventional data-driven models. As a result, the proposed method showed a high estimation accuracy (average absolute error less than 5 %) over a wide driving temperature range (0-25 degrees C) with only limited driving data. Especially, the proposed method showed reliable accuracy even in vehicle driving temperature ranges where data-driven models cannot be trained due to data scarcity.