MS-CGDM: Multi-scale conditional graph diffusion model for extreme weather source-load scenario generation

The frequent occurrence of extreme scenarios of large-scale access to renewable energy and extreme weather leading to dramatic fluctuations in power generation and power demand has posed a serious challenge to the safe and stable operation of the power system. However, there is little methodology available for extreme weather scenario generation and a lack of identification of potential risks to the power system during extreme weather. Moreover, existing scenario generation methods are unable to fully characterize the long-tailed distribution of extreme weather, resulting in significant deviations between the generated scenarios and the actual extreme weather scenarios, which affect the scheduling decisions of the power system in extreme weather. To address this challenge, this paper proposes the multi-scale conditional graph diffusion model (MS-CGDM) for reliable and diverse extreme weather source-load scenario generation. Specifically, MS-CGDM employs a multi-scale feature extraction mechanism and a graph neural network (GNN) to learn the dynamic coupling relationships among wind power, photovoltaic power, and load by dynamically modeling them as graph nodes. Then, a targeted weighting strategy is implemented using an asymmetric loss function to allow the improved diffusion model to generate extreme weather scenarios that satisfy the long-tail characteristics of extreme weather. After experimental validation, the MS-CGDM proposed in this paper outperforms other scenario generation models in generating extreme weather source-load scenarios, demonstrating superior reliability and diversity.