A dual spatio-temporal network for short-term wind power forecasting

Abstract

The intricate spatial and temporal complexities inherent in wind speed data necessitate the utilization of spatio-temporal networks for accurate wind power generation forecasting. But these complexities are made worse by the constant presence of high-frequency disturbances and clear outliers in the spatial and temporal correlations of wind speed, which makes forecasting wind power very difficult. Additionally, the intricate topological architectures that underpin current models often result in superfluous computational operations and susceptibility to overfitting. In response to these formidable challenges, we introduce an innovative solution in the form of the dual spatio-temporal (DST) network. This paradigm, characterized by meticulous design, excels in acquiring precise and resilient spatio-temporal representations essential for wind power forecasting. This is possible because of the careful development of two types of spatial correlation layers: synchronous layers and asynchronous layers. Each type is fine-tuned to recognize different spatial patterns. Recognizing that spatial and temporal correlations change over time, our method explores spatio-temporal representations at many different scales. This makes the system more adaptable and robust. Notably, for effective noise mitigation over extended forecasting horizons, we judiciously integrate wavelet decomposition. Empirical validation, rigorously conducted across twelve wind power stations in Anhui Province, China, firmly attests to the efficacy of our proposed DST network. In summary, our innovation presents a promising avenue for advancing wind power forecasting, characterized by adept noise reduction and improved precision while concurrently optimizing computational efficiency in modeling intricate spatial–temporal dependencies.