关键词: 场景生成, 去噪概率扩散模型, 多任务学习, 动态特性自适应, 风光荷

Abstract: With the high-penetration integration of renewable energy sources such as wind and solar, the generation of joint scenarios that can accurately capture the dynamic characteristics and complex interdependencies of each variable is crucial for power system dispatch and control. To model the power outputs for wind, solar, and load scenarios, this paper proposes a scenario generation method based on an adaptive multi-task diffusion model. First, a multi-task learning framework is introduced, built upon a joint denoising network. This framework processes multivariate state vectors jointly while fusing temporal information to generate scenarios faithful to the intrinsic physical couplings and temporal dependency patterns. Secondly, an adaptive diffusion strategy module guided by the dynamic features of heterogeneous data is developed. This module extracts dynamic statistical features from the generated data and dynamically adjusts the noise schedule of the diffusion process accordingly, enabling the accurate modeling of the data's non-stationary and time-varying dynamics. Building on this, a structurally-consistent guided training criterion is proposed. By imposing constraints on two key data structure properties—marginal distributions and joint dependencies—within the training objective, this criterion effectively guides the model's generation process and enhances the quality of the generated wind-solar-load scenarios. Finally, case studies conducted on a power dataset from the IES-134 standard test system validate the effectiveness and superiority of the proposed method in generating joint wind-solar- load scenarios.

Key words: scenario generation, diffusion models, multi-task learning, wind-solar- load