Safe Reinforcement Learning for Energy Management of Electrified Vehicle With Novel Physics-Informed Exploration Strategy

This article introduces a novel physics-informed exploration (PIE) strategy for a deep reinforcement learning (DRL)-based energy management system (EMS), specifically targeting the challenge of dealing with constrained action sets. RL-based controllers for electrified vehicle EMS have faced obstacles stemming from the selection of infeasible actions, obstructing their practical deployment. The absence of a mechanism for assessing control action feasibility prior to application has compounded this issue, primarily due to the model-free nature of RL-based controllers. Adding a safety layer to the RL-based controller addresses the abovementioned issue, but this often results in suboptimal policies and necessitates an in-depth understanding of the powertrain. Alternatively, theoretical remedies incorporate penalty terms into the immediate reward function to manage infeasible conditions. However, this approach can slow down the training process as the agent learns to avoid infeasible actions. To surmount these challenges, this article introduces a novel PIE strategy, coupled with prioritized experience replay (PER), enabling the agent to swiftly learn to avoid selecting infeasible control actions without the need for a separate safety layer. Real-time simulation results highlight the superior performance of the proposed DRL-based controller over the baseline DRL-based controller with a safety layer, particularly in terms of overall fuel consumption (FC).