• A multi-scale condition adaptive diffusion model (MS-CADM) is developed for controllable wind power scenario generation. • A multi-scale conditional embedding network is proposed to hierarchically extract complex NWP. • A time-step adaptive Transformer is introduced to decouple temporal conditions from external NWP inputs. • A learnable variance formulation and random conditional masking strategy are incorporated to improve uncertainty modeling and enhance training stability. • Extensive experiments on GEFCom2014 wind power dataset demonstrate that MS-CADM achieves improved uncertainty quantification, capture of peak variability and robustness under complex meteorological conditions. The large-scale integration of wind power poses significant challenges to power system operation, as the intermittency and stochasticity of wind power generation can adversely affect grid stability. While diffusion models have been applied to wind power scenario generation, their performance remains sensitive due to an inadequate capacity for processing conditional information. To address these issues, this manuscript proposes an multi-scale condition adaptive diffusion model(MS-CADM). Specifically, a multi-scale embedding network is proposed to extract complex conditions at multiple hierarchical levels, while a time step adaptation Transformer architecture is introduced to decouple the time-step condition unique to diffusion models from the external control conditions, which makes model to learn different controllable conditions accurately. During training, a learnable variance term is incorporated to enhance posterior expressiveness, and a random conditional masking strategy is applied for regularization to stabilize training. On Global Energy Forecasting Competition 2014 (GEFCom2014) wind power dataset, MS-CADM reduces MAE by 4.26% and RMSE by 1.91% compared to state-of-the-art methods and achieves competitive performance across six metrics, demonstrating its ability to effectively leverage conditional information and achieve more efficient convergence.
Zhang et al. (Tue,) studied this question.