Spatio-temporal graph attention network and graph-based transformer architecture for distributed urban wind sequence reconstruction and forecasting

Abstract

Urban wind field forecasting plays an important role in city sustainability. As the urban is highly affected by the terrain complexity and city arrangement and its random and nonlinear features, the flow mode is hard to capture and conclude, which make the wind speed forecasting task difficult. Meanwhile, the quality of the wind speed signal depends on the measurement station in service, which is easy under an unavailable state, such as malfuction. In this study, we propose an innovative framework for wind speed data reconstruction and prediction, integrating the graph attention network (GAT) with graph-based Transformer models. This comprehensive framework effectively facilitates forecasting tasks for five wind speed measurement stations, even with a limited number of operational stations. Through comparative analysis with other specific methodologies, our experimental results demonstrate that the predictive accuracy using reconstructed data is comparable to that of the original data, maintaining a low prediction loss with a mean absolute error of 2.218 and a reconstruction accuracy exceeding 84%. In predictive tasks, the graph-based Transformer model excels in long-term forecasting tasks, such as 30-step predictions, effectively capturing future trends and reducing execution time. However, it does not exhibit significant advantages in short-term forecasting tasks, such as 1-step and 10-step predictions, and requires substantial training time.