Abstract:

In the context of high-penetration renewable energy integration， accurately constructing wind-solar-load joint scenarios that can capture the dynamic characteristics and complex correlations of various variables has emerged as a core requirement for power system scheduling and control. To this end， a generation method for wind-solar-load scenarios based on an adaptive multi-task diffusion model was proposed. A multi-task diffusion model learning architecture based on a joint denoising network was established. By jointly processing multivariate state vectors and integrating temporal information， realistic joint scenarios that integrated physical coupling relationships and temporal dependency patterns were generated. On this basis， an adaptive diffusion strategy module guided by the dynamic features of heterogeneous data was proposed. Dynamic statistical features of the generated data were extracted， and the noise scheduling of the diffusion process was dynamically adjusted accordingly， thereby effectively characterizing the non-stationary and time-varying dynamic characteristics of the data. Meanwhile， a training criterion guided by structured consistency was introduced， where the marginal distribution and joint dependency characteristics of the data structure were constrained within the training objectives. It effectively guided the model generation process and improved the quality of wind-solar-load scenario generation. Validation based on the IES-134 standard dataset demonstrated that the proposed model could effectively generate wind-solar-load joint scenarios with realistic physical characteristics and reasonable statistical patterns， providing a practical tool for optimal scheduling and risk assessment in power systems.

Key words: renewable energy, wind-solar-load joint scenarios, adaptive multi-task diffusion model, denoising diffusion probabilistic model, multi-task learning, dynamic feature adaptation