Enhancing Active Power Scheduling with Multimodal Grid Data Based on Multi-Agent Reinforcement Learning

The rapid and widespread integration of renewable energy sources introduces significant challenges for power system dispatch. Owing to the inherent intermittency and variability of renewable outputs, conventional active power scheduling methods based on static models are often inadequate for capturing system dynamics and managing operational uncertainties. To address these issues, this paper proposes an optimization approach that integrates multimodal data sensing with multi-agent deep reinforcement learning. The proposed framework, named Heterogeneous Multi-Agent Transformer with Proximal Policy Optimization (HMAT-PPO), combines a Transformer-based architecture with PPO to jointly capture the spatial structures, temporal dynamics, and operational features of power grids. Through the incorporation of multimodal alignment and gated fusion mechanisms, the proposed framework enables the integration of heterogeneous information sources—such as grid topology, load fluctuations, and generator states—which significantly augments the agents’ environmental awareness and promotes collaborative, context-aware decision-making. Experimental results demonstrate that the proposed method consistently outperforms baseline models in minimizing generation cost, reducing transmission losses, and satisfying operational constraints, thereby offering both theoretical significance and practical value.