Reinforcement learning for the control of battery electrothermal behaviours in electric vehicles
Résumé
Battery lifetime is related, among other things, to the battery temperature and RMS battery current. This paper presents an improved energy management of battery / supercapacitors (SCs) hybrid energy storage system (HESS) in an electric vehicle (EV) aiming at reducing the RMS battery current and battery temperature. A reinforcement learning (RL) based real-time energy management framework is designed to ensure an optimal power flow distribution between battery and supercapacitors starting from historical observation of the RMS battery current. First, the battery and SCs storage devices are modeled. An electrical model is used for the SC and an electrothermal representation is adopted to follow the evolution of the battery temperature and its electrical parameters (current, voltage). Then the RL energy management problem is formulated satisfying the electrical HESS constraints. The proposed methodology generates in real time an optimal power sharing between battery and SCs without any prior knowledge of the load variations of the EV. In our work, we propose a novel approach combining the rule based controller Frequency sharing with RL learning to achieve the best solution optimality. This approach is effective to adapt the rule-based strategy to work in their efficiency region and introduce additional intelligence. Simulation results have confirmed the convergence of the RMS battery current to the minimum values and appreciable reductions of the battery temperature are obtained.