A wind power forecasting model based on data decomposition and cross-attention mechanism with cosine similarity

Abstract

In the context of societal advancement, escalating energy demands underscore the critical significance of exploring renewable resources. The accurate prediction of wind power is crucial for maximizing the utilization of wind energy and ensuring the smooth operation of wind farms. Currently, the majority of models are built upon architectures such as Transformer and Long-Short Term Memory (LSTM). However, Transformer-type models are not well-suited for time series forecasting, and LSTM often focuses on historical sequence information, exhibiting limited capability in capturing the relevant information between weather factors and power. Recognizing this inadequacy, this study proposes a novel model—a linear network architecture featuring a cosine-related cross-attention mechanism. The signal decomposition is carried out utilizing variational mode decomposition, with the determination of the optimal decomposition count guided by sample entropy analysis, named Improved Variational Mode Decomposition (IVMD). Subsequently, the conventional self-attention mechanism is replaced with a cross-attention mechanism. Contemporary mainstream models are primarily based on improved versions of core algorithms such as LSTM, Transformer, and Extreme Learning Machines. Experimental results show significant performance enhancements compared to these models across different timescales. In short-term, medium-term, and long-term forecasts, accuracy improves by more than 40 %, 20 %, and 5 %, respectively.