Wind Speed Prediction Model for Multiple Sites Using Heterogeneous Spatio-Temporal Learning

Understanding and predicting wind speeds at both multiple sites and steps is significantly important for wind farm power generation prediction. To capture temporal and spatial dynamics inherent in wind speed data, a wind speed prediction model for multiple sites is proposed, which leverages fusion learning of heterogeneous convolutional channels to extract the spatio-temporal characteristics of regional winds. The correlation of wind speeds in different timing scales is diversely exploited and integrated. A three-dimensional convolutional channel is designed to extract the medium-longterm characteristics, and a depthwise separable temporal convolutional channel is designed to extract the fine-grained dynamic characteristics in the short term. Then, wind speeds at multiple sites can be predicted with a fusion strategy on the multi-scale features. The simulation results on Automated Surface Observing System (ASOS) dataset show that the proposed prediction method can achieve significantly higher prediction accuracy when compared with the state-of-the-art similar prediction methods.