Wind direction prediction combined with wind speed in a wind farm

This paper presents a two-step method to predict the incoming wind conditions (wind speed and direction) of a wind turbine, considering the mutual influence of topography and wake, the impact of mesoscale meteorological conditions, and the time varying characteristics. Firstly, the incoming wind conditions of turbines are calculated using wind conditions obtained from the anemometer tower, which takes into consideration the coupling effect exerted by the terrain and wake effects between anemometer tower and the turbines on the wind conditions. Secondly, incoming wind conditions are predicted using an encoder-decoder cross-attention module. This module extracts spatial features tied to mesoscale meteorological conditions while simultaneously capturing time-varying characteristics of the wind. These features are obtained through the combined application of LightGBM and dilated causal convolutional networks. Thirdly, LiDAR experiments conducted in a China's wind farms substantiated the efficacy of the method for forecasting incoming conditions. The results show that, compared to LiDAR, the proposed method has a RMSE of 20.8° for wind direction and 0.792 m/s for wind speed. This represents an improvement in the accuracy of the wind direction and wind speed by 38.6 % and 38.2 %, respectively, in comparison with SCADA. This method can provide a reference for wind power predictions. • The proposed model predicts incoming wind speed and direction for turbines. • An encoder-decoder TCN model is used to fuse the spatial and time-series features. • LiDAR experiments validate the accuracy of the wind speed and direction prediction. • The method improves wind direction accuracy by 38.6 % compared to SCADA.