Multi-objective coordinated scheduling for flexible distribution network based on reward-decoupling normalization reinforcement learning

The interconnection of a high proportion of distributed generation, energy storage systems, and advanced regulation equipment in flexible distribution networks (FDNs) offers a powerful means of supporting global energy transitions by providing enhanced flexibility, efficiency, and reliability. However, despite recent advances in reinforcement learning (RL), the optimized scheduling of highly complex FDN operations remains extremely challenging owing to the synergies offered by heterogeneous flexible resources, the complex game-theoretical coupling of multiple objectives, dimensional discrepancies, and the inherent uncertainties of sources and loads. The present work addresses this issue by establishing a collaborative operation model for FDNs that incorporates multiple objectives, including operational economy, system security, and renewable energy consumption rates. We propose an RL solver based on a policy optimization strategy that substantially enhances training stability and convergence by eliminating strong coupling among the multiple objectives through a decoupled reward mechanism and resolving dimensional discrepancy issues via normalization. The results of numerical computations demonstrate the effectiveness of the proposed method for improving the multi-objective collaborative optimization of the system and facilitating flexible interactions among heterogeneous resources, while offering improved multi-objective compromise performance and convergence behavior.