Fast solution method for mid- to long-term SCUC scheduling using graph neural networks and reinforcement learning

Abstract Mid- to long-term Security-Constrained Unit Commitment (SCUC) faces severe dimensional scaling and complex spatiotemporal coupling, exacerbated by extreme renewable variability. Consequently, traditional mathematical programming struggles to balance solution efficiency with dynamic security under practical operational time limits. This study proposes an integrated computational framework combining Generative Adversarial Networks (GANs), Temporal Graph Neural Networks (TGNNs), and Proximal Policy Optimization (PPO). The framework employs physically constrained GANs to synthesize extreme scenario distributions, mitigating data scarcity. Subsequently, TGNNs utilizing temporal attention and graph convolutions extract underlying spatiotemporal network topologies, feeding a PPO agent designed to optimize economic dispatch and unit commitment. By establishing an asymmetric architecture that decouples neural policy generation from an optimization-based Linear Programming (LP) projection layer, the framework enforces linearized physical feasibility strictly under the Direct Current (DC) power flow approximation. Experimental evaluations demonstrate that under prolonged scheduling horizons facing out-of-distribution disturbances, the proposed method provides notable improvements in managing system operational costs while maintaining robust security constraint satisfaction rates (averaging 99.3%). Furthermore, the dual-mode online inference architecture reduces neural inference to the millisecond scale, achieving sub-second end-to-end execution latencies that significantly outperform traditional mathematical programming benchmarks, thereby demonstrating significant operational potential for both day-ahead base scheduling and rapid contingency screening.