Wind Speed Forecasting Based on Denoising Diffusion Model and Multivariate-Aware Spatiotemporal Transformer

As a renewable resource with significant potential, wind energy presents considerable challenges for its integration into the energy grid due to its inherent randomness. Therefore, accurate wind speed forecasting is crucial for effectively utilizing wind energy. Existing wind speed forecasting methods still fall short in modeling spatiotemporal correlations and lack sufficient consideration of the relationships between wind speed and other auxiliary variables, such as temperature and pressure. This limitation affects their performance in complex wind speed forecasting scenarios. To address these challenges, we propose a multivariate-aware spatiotemporal diffusion model (MASTDM) for wind speed forecasting. This model is a probabilistic forecasting approach based on Diffusion Denoising Probabilistic Models (DDPM), which effectively models temporal and spatial correlations using the self-attention mechanism of transformers. Furthermore, it explicitly models the multivariate correlations and the interdependencies between wind speed and auxiliary temporal information. Experimental results demonstrate that our approach achieves comprehensive and significant performance improvements over various baseline methods.