A hybrid physics and data-driven framework for ultra-short-term wind power forecasting with spatiotemporal learning.

This article focuses on developing an ultra-short-term wind power forecasting method that integrates physical principles and data-driven techniques while explicitly addressing spatiotemporal correlations. The proposed framework employs a dual-layer adaptive graph neural network (AGCN) to capture complex spatial and temporal dependencies in wind power data and introduces a dynamic model switching mechanism triggered by a predefined wind speed mutation threshold to alternate between a data-driven model (DDM) and a physical-data hybrid model (DPHM). Empirical results from a 99 MW wind farm dataset demonstrate that the AGCN outperforms traditional machine learning and deep learning models in prediction accuracy, and the hybrid approach further improves forecasting stability and accuracy during abrupt wind speed fluctuations. The study highlights the importance of meteorological feature selection, adaptive weighted pooling, and threshold-based model switching in enhancing wind power prediction within integrated energy systems, while acknowledging limitations related to data quality and model interpretability.