Wavelet-denoised graph-Informer for accurate and stable wind speed prediction

Abstract

As global demand for renewable energy increases, wind power has become increasingly valued as a clean energy source. Effective wind speed forecasting is crucial for wind energy production and power grid's stability. To mitigate high-frequency interference in wind speed signals and improve spatiotemporal feature extraction, we propose a novel short-term wind speed prediction model called WST-Informer. Firstly, Wavelet Decomposition (WD) is fused to erase high-frequency noise in monitored signal series. Secondly, Informer encoder is designed to capture long-term temporal dependencies efficiently, and multiple cities' spatio-temporal maps are constructed through designing Residual Graph Convolutional Network (RS-GCN). Moreover, a new Attentional Feature Fusion (AFF) method is designed to fuse temporal and spatial features. Furthermore, the decoder of the Informer predicts outcomes using fused features. Additionally, outliers are more prone to bigger errors and highly curtail in real-world application, a Kernel Mean Squared Error (KMSE) loss function is introduced to further enhance their prediction. With real datasets from weather stations across five Danish cities and seven Dutch cities, extensive experiments were conducted, and the proposed model demonstrates reduced prediction error across multiple forecasting steps in both datasets, resulting in lower prediction errors during sudden wind speed changes, outperforming current state-of-the-art time series forecasting models.