Attention mechanism is useful in spatio-temporal wind speed prediction: Evidence from China

Abstract

The spatio-temporal wind speed prediction technology provides the key technical support for the energy management and space allocation of the wind farm. To obtain an accurate spatio-temporal wind speed prediction model, the framework based on the attention mechanism is proposed. The attention mechanism-based model consists of two key steps: In step I, as the main framework of spatial feature extraction, the graph attention network (GAT) can effectively aggregate the wind speed feature data of the target station with other auxiliary stations and extract the feature information. In step II, the spatio-temporal feature data extracted from GAT were transmitted to Informer to train the prediction framework and obtain wind speed prediction results of the target site. Based on the spatio-temporal modeling results of two different wind farms, the following conclusions can be summarized: (1) As the main predictor, the Informer can achieve better prediction performance than the traditional deep learning model. (2) GAT can effectively aggregate wind speed data from different stations and obtain excellent spatio-temporal wind speed features. (3) The proposed attention mechanism-based spatio-temporal wind speed prediction model can achieve satisfactory prediction results. GAT-Informer, whose Mean Absolute Error (MAE) values, Mean Absolute Percentage Error (MAPE) values and Root Mean Square Error (RMSE) values are less than 0.42, 4.1% and 0.54 respectively, works better than other twenty-one models.