Safe deep reinforcement learning hybrid electric vehicle energy management

Abstract

The optimality-based design of the energy management of a hybrid electric vehicle is a challenging task due to the extensive and complex nonlinear reciprocal effects in the system, as well as the unknown vehicle use in real traffic. The optimization has to consider multiple continuous values of sensor and control variables and has to handle uncertain knowledge. The resulting decision making agent directly influences the objectives like fuel consumption. This contribution presents a concept which solves the energy management using a Deep Reinforcement Learning algorithm which simultaneously permits inadmissible actions during the learning process. Additionally, this approach can include further state variables like the battery temperature, which is not considered in classic energy management approaches. The contribution focuses on the used environment and the interaction with the Deep Reinforcement Learning algorithm.