Multi-step short-term wind speed predictions employing multi-resolution feature fusion and frequency information mining

Accurately predicting wind speed is crucial towards maximizing the use of wind energy. Unfortunately, due to the non-stationary and intermittent nature of wind speed, reliable prediction has proven to be challenging. Properly exploring the inherent sequence connection between wind speed and its underlying frequency characteristics would be advantageous in enhancing forecast accuracy. Expanding on this concept, innovative models have been developed. Initially, they utilize multiple convolutions to capture the internal features of the time series data. This is followed by the application of the state frequency memory recurrent block to extract the relevant frequency properties. To facilitate the identification of crucial hidden elements, different attention mechanisms are utilized between these two modules which enables multi-resolution feature fusion. Real-world data is used to conduct multi-step short-term wind speed prediction tests, and various classical models are evaluated. The results indicate that our proposed models, which possess time-frequency mining capability, outperform other models in terms of forecast accuracy across four commonly used error indices. Additionally, a parameter sensitivity analysis is conducted on the prediction models to demonstrate the high stability of the proposed model. Furthermore, the working mechanism of our proposed model is thoroughly explored in order to confirm its exceptional prediction accuracy.