Ultra-short-term Output Scenario Generation for Renewable Energy Plants Based on Improved Denoising Diffusion Probabilistic Models and Model-based Transfer Learning

The output of renewable energy plants exhibits strong uncertainty, posing significant challenges to the dispatching and controlling of new-type power systems. To achieve accurate modeling of renewable energy output scenarios, first, addressing the uncertainty of ultra-short-term output for renewable energy plants, this paper proposes an improved Denoising Diffusion Probabilistic Model (DDPM). This model uses an enhanced self-attention mechanism to design a neural network architecture tailored for ultra-short-term output scenario generation of renewable energy plants better to capture the correlation of renewable energy output time series and fit its probability distribution. Subsequently, addressing the issue of insufficient historical data for newly built renewable energy plants, a framework for output scenario generation of newly built renewable energy plants based on model-based transfer learning is proposed to complete the construction of the scenario generation model under the condition of small sample. Finally, case studies are conducted on the wind and solar output dataset open-sourced by the National Renewable Energy Laboratory (NREL) in the United States. Combined with the proposed evaluation metrics, the results of case studies indicate that the model proposed in this paper significantly outperforms generative adversarial networks, variational autoencoders and models without model-based transfer learning.