GEV distribution-enhanced Fourier diffusion model for extreme value capture in day-ahead photovoltaic scenario generation

Day-ahead photovoltaic (PV) scenario generation is a critical task in the field of renewable energy, especially for fine-scale temporal tasks such as economic dispatch planning for the day ahead, optimization of unit operation commitments, and participation in electricity markets. However, existing research on PV scenario generation primarily focuses on capturing the global statistical characteristics of the dataset and simulating its temporal characteristics, with significant limitations in modeling extreme events. This paper proposes a novel day-ahead PV scenario generation method based on a Generalized Extreme Value (GEV) enhanced Denoising Diffusion Probabilistic Model (DDPM) combined with Fourier operators. The proposed method is designed to effectively capture extreme value features in PV power data, particularly the generation of peak power values. Firstly, we integrate the GEV distribution into the DDPM framework to guarantee the realism and accuracy of the generated data in both the overall distribution and extreme events. Secondly, to improve the noise prediction accuracy of the reverse denoising network, we introduce a Fourier operator-based denoising network. This network leverages Fourier transforms to extract features in the frequency domain, boosting the model's effectiveness in identifying the temporal dynamics of PV power data. Finally, to conduct an in-depth evaluation of the generated data quality, we propose a multi-faceted evaluation approach that considers both deterministic and uncertainty assessments, providing a thorough evaluation of the accuracy and reliability of the generated data. Experimental results show that the proposed method achieves the lowest MAE (0.12 and 0.05), RMSE (0.23 and 0.08), and the highest R2 (0.95 and 0.94) on the Hebei and DKASC datasets, respectively, outperforming all benchmark models. Moreover, it demonstrates strong robustness in both deterministic and uncertainty evaluations, indicating its superior ability to generate realistic and reliable day-ahead PV scenarios, particularly under extreme conditions.