Proper determination of the state of health (SoH) of electric vehicle (EV) batteries is essential to guarantee reliable performance, operational safety, and durability of EV transportation systems. Nevertheless, current SoH estimation approaches do not scale to different driving patterns, environmental factors, and usage patterns. This results in decreased accuracy and generalisation to real-world scenarios. To address these issues, this paper proposes an adaptive SoH estimation framework utilising reinforcement learning (RL). It can learn and adapt to various driving behaviours like taxis, delivery vans, and family EVs. It initially models battery dynamics with three different driving styles, and subsequently creates time-series current, voltage, state of charge (SoC) and temperature data, including the effects of degradation. These signals are processed to extract engineered statistical features. Further, gradient boosting regressors are trained as style-specific machine learning (ML) experts to forecast subsequent SoH. Further, a deep Q-network (DQN) agent is trained to operate within a custom environment so as to choose the most trusted expert dynamically based on prediction error feedback and transitions between driving styles. During testing, it was found that the model has a prediction accuracy of over 97%. Overall, experimental findings and visualisation support the robustness, stability, and flexibility of the model in a variety of driving conditions.
Singh et al. (Mon,) studied this question.