Conditional denoising diffusion probabilistic model based ante-hoc explainable scenario generation for power systems dispatch

The data-driven methods represented by deep reinforcement learning (DRL) face critical challenges in solving power systems dispatch problems due to limited samples of extreme scenarios. Deep generative models are difficult to applied to generate various scenarios due to their poor explainability and stability. To address these challenges, this paper proposes a DRL framework based on conditional denoising diffusion probabilistic model (CDDPM), which integrates a CDDPM-based ante-hoc explainable generative model and a DRL-based systems dispatch model. It achieves the explainability of scenario generation by establishing explainable and explicitly-quantifiable forward diffusion and reverse denoising processes, and enhances the stability of scenario generation by constructing time-series denoising network (TSDN). The verification shows that the proposed framework can explain and stably generate various scenarios including extreme scenarios, and improve the performance of the DRL-based power systems dispatch.