Learning-based Control for Hybrid Battery Management Systems

Abstract

Battery packs of electric vehicles are prone to capacity, thermal, and aging imbalances in their cells, which limit power delivery to the vehicle. In this chapter, a hybrid battery management system (HBMS), capable of simultaneously equalizing battery capacity and temperature while enabling hybridization with supercapacitors, is investigated. We use model-free reinforcement learning to control the HBMS, where the control policy is obtained through direct interaction with the system’s model. Our approach exploits the soft actor-critic algorithm to handle continuous control actions and feedback states, and deep neural networks as function approximators. The validation of the proposed control method is performed through numerical simulations, making use of numerically efficient models of the energy storage and power converters developed in Modelica language.