Safe Deep Reinforcement Learning for Energy Management of Electrified Vehicles: Optimal Action Filtering and Battery-in-the-Loop Validation

Despite the promising performance of deep reinforcement learning (DRL)-based energy management systems (EMS) for electrified vehicles, persistent safety concerns regarding control actions hinder their real-world deployment. This paper proposes a safe DRL-based EMS for hybrid electric vehicles (HEVs) that guarantees zero safety violations during both training and deployment. A model-based safety layer is developed with prior knowledge to filter unsafe actions with minimal disruption to agent exploration. This safety metric evaluation is embedded into the reward function via a Lagrangian relaxation method, enabling adaptive penalization of constraint violations and prompting efficient policy learning. A safe soft actor-critic (SSAC) EMS is then trained under stochastic conditions, including random initial battery state-of-charge (SOC) and multi-modal driving cycles. The approach is validated through simulation and battery-in-the-loop (BIL) experiments. The proposed reward formulation accelerates safety-aware policy learning by 38.5% compared to conventional methods. In BIL tests, SSAC-EMS achieves over 95% fuel efficiency relative to the offline dynamic programming (DP) benchmark under unseen driving cycles and initial SOCs, with actions consistently respecting safety constraints. Compared to the adaptive equivalent consumption minimization strategy (AECMS), SSAC-EMS improves fuel economy by 6–10% while delivering more stable SOC regulation and smoother engine operation.