Wind speed prediction based on deep learning and time frequency feature

Abstract Wind speed prediction is a crucial issue in meteorology and environmental science, and time-frequency features benefits the understand of various scales of wind variations and further boosting wind speed prediction performance. Therefore, this study designs a wind speed prediction network based on the Gated Recurrent Unit (GRU) network and Time Series Decomposition(TSD), and together with a scale attention mechanism to fuse features of different frequencies adaptively. Firstly, the observed data is decomposed into different periodic components through TSD, reducing the complexity of the prediction task. Then, an attention module is used to precisely weight each predicted component and further fuse them adaptively. After that, GRU networks are constructed to extract features and make predictions for each type of periodic component, relieving the prediction of complex wind speed observations. Finally, extensive experiments are conducted based on wind speed datasets from meteorological stations in seven cities in Netherlands, verifying the effectiveness of the proposed model. Comparative analysis with some state of the arts are also performed to demonstrate the advance of the designed model.