A Dynamic Graph Fusion Model for Ultra-Short-Term Turbine-Level Wind Power Forecasting

Abstract

Accurate ultra-short-term wind power forecasting at the turbine level is important for grid stability and dispatching. To address the time-varying spatial and temporal correlations among multiple turbines in a single wind farm, we build dynamic spatio-temporal graphs to model dynamic spatial dependencies, propose a parallel multi-scale temporal convolutional encoder to combine short-term and long-term dependencies, and propose a graph fusion layer to achieve weight fusion of different graph sources. Experiments demonstrate that GraphFusionGRU achieves lower overall error in short-term forecasting and achieves competitive average performance relative to other baseline models on longer horizons. The results confirm that the model’s robustness and interpretability are enhanced in complex wind-farm environments.