A Framework for Safe Reinforcement Learning in Battery Storage Scheduling Under Partial Observability

The integration of Battery Storage Systems (BSS) into microgrids can enhance renewable energy utilization, reduce operating costs, and improve stability. However, the inherent uncertainty of photovoltaic (PV) generation and electricity demand presents significant challenges for safe and efficient control. This paper proposes a unified framework for safe BSS scheduling, using a dual safety mechanism integrated into a deep reinforcement learning (RL) agent operating under realistic, uncertain conditions. This approach exploits a novel agent architecture based on the Proximal Policy Optimization (PPO) algorithm with a custom feature extractor that captures both present operational states and long-term dependencies. This feature extractor leverages a Long Short-Term Memory (LSTM) encoder-decoder for PV generation and demand forecasting. Multiple safety metrics are evaluated, including violation rates during active operation, and the results are benchmarked against state-of-the-art model predictive control and RL methods. The proposed approach yields the best trade-off between economic performance and operational safety, improving safety compliance by over 74% relative to a model predictive control baseline while maintaining comparable economic return. The novelty of this work lies in jointly addressing partial observability, heterogeneous temporal dependencies, and multi-metric safety evaluation in a single framework for BSS scheduling.