Energy Management for Ship Integrated Power Systems via Mode-Aware Safe Reinforcement Learning

Energy management in ship integrated power systems (IPSs) requires real-time dispatch of diesel generators, battery storage, and shore power under strict operational constraints. Existing deep reinforcement learning (DRL) approaches are economically competitive but cannot guarantee that device constraints are satisfied during training or deployment. This paper proposes MA-SRL, a safe reinforcement learning framework for ship IPSs that couples an execution-layer Safe Projection Layer (SPL) with a training-stage Lyapunov-based policy update. The SPL projects each raw action onto the feasible set of the current mode before execution, whenever that set is non-empty, and quantifies the departure as a constraint-cost signal that drives a Mode-Dependent Constrained Markov Decision Process (MD-CMDP), making feasibility observable to the learner. The Lyapunov update is designed to control the expected discounted constraint cost through a budget condition during training. Under the nominal scenario, this signal drives the raw policy close to the feasible set, lowering constraint violation by 2.2–6.0× and cost by 7.7–16.9% over DRL baselines, with the lowest constraint violation retained under storm conditions. The same architecture and settings are replicated on a second vessel, plant, route, and operating profile, attaining the lowest cost and constraint violation of all compared methods, and sustaining the lowest distance shortfall and unmet load under a single-generator-loss contingency.