Safe Reinforcement Learning-Based Robust Scheduling for Energy Systems: Field Experiments and Energy Management Insights

Safety-critical energy system scheduling under uncertainty poses a fundamental challenge: operators must balance economic efficiency with strict physical constraints such as battery lifetime protection and electric vehicle (EV) quality-of-service (QoS). Conventional reinforcement learning approaches often fail to guarantee constraint satisfaction, leading to unsafe battery operation or unmet EV demand. To address this issue, we propose a Decision Transformer-based framework designed for robust and safety-aware energy scheduling. Which introduces Cost-to-Go in the training process, enabling the model to jointly learn long-term rewards and cumulative safety costs, thereby proactively preventing equipment degradation and service failures. In addition, an \textbf{out-of-distribution (OOD) detection} mechanism is incorporated to determine whether inputs deviate from the training distribution, allowing the framework to distinguish between soft and hard constraints. This dual-layer design achieves a balanced trade-off between safety and flexibility during policy learning. Extensive experimental results across diverse scenarios demonstrate that algorithm reduces violation rates by more than 70% compared to baseline methods, while maintaining millisecond-level inference latency suitable for real-time energy system scheduling. These findings highlight our algorithm as a practical solution that simultaneously ensures equipment safety and user-centric reliability, offering a pathway toward robust and efficient operation in next-generation energy systems.