An AI-based weather prediction method for wind farms combining global forecast field and wind speed temporal transfer characteristics

The accuracy of wind power forecasting depends critically on the quality of weather prediction, typically derived from Numerical Weather Prediction (NWP) models. Despite NWP is widely used, it faces the following limitations: the long computation time affects the timeliness of prediction, and its mesoscale model ignores wind farm-specific flow field information. To address the problems, an AI-based Weather Prediction (AWP) method that combines large-scale spatial characteristics of global forecast field and wind speed temporal transfer characteristics is proposed to reduce the computation time from hours to seconds, and realize the micro-scale prediction at the turbine-level. Firstly, a large-scale spatial feature calculation module is constructed using Multi-scale Convolutional neural network to extract the spatial correlation features across various scales and depths. Then, a temporal transfer feature calculation module is established utilizing Informer to explore the internal dependencies and detailed fluctuation features. Finally, LightGBM is employed to integrate the features calculated to obtain the microscale wind speed prediction. Data from 10 wind farms are taken to verify the effectiveness and robustness of the method. The proposed method can reduce the wind speed prediction error by 0.72–18.11 % (0.02–0.62 m/s) and can better capture the detailed fluctuation features, compared to three commercial NWPs.