An efficient spatial-temporal graph latent diffusion model for regional ultra-short-term wind speed probabilistic forecasting

Accurate and efficient regional ultra-short-term wind speed probabilistic forecasting is critical for the integration of renewable energy into power grids, which requires the rapid generation of high-precision forecasts within a short period. Diffusion models excel at learning complex data distributions, making them ideally suited for representing the highly stochastic and fluctuating nature of wind speed. However, their slow, iterative sampling process creates a fundamental bottleneck for time-sensitive forecasting applications. To address this, we propose an efficient spatial-temporal graph latent diffusion model (STGLD), which introduces a novel leap-refine sampling framework to achieve both precision and speed. A leap predictor accelerates sampling by jumping to a coarse intermediate state in one step. Subsequently, a refine denoiser governed by a dynamically scheduled sparse attention mechanism performs efficient and progressive denoising, evolving its receptive field from global to local spatial-temporal patterns. This pipeline operates in a latent space, facilitated by an integrated auto-encoder that reduces the computational burden through dimensionality reduction. Experiments conducted on publicly available wind speed data demonstrate the superiority of our approach. Specifically, STGLD reduces the continuous ranked probability score (CRPS) by at least 3.05% compared to all benchmark models. Furthermore, it achieves a 75% reduction in computational cost while improving reliability by 14.5% and narrowing the prediction interval width by 18.2% compared to other diffusion models.