Prediction of wind fields in mountains at multiple elevations using deep learning models

Abstract

To optimize the site selection of wind turbines, it is essential to understand the variability of wind speeds at different elevations on mountains and their wind characteristics. The erection of anemometer at a large altitude in mountainous regions poses significant financial and safety challenges. Therefore, it is of utmost importance to design a method that facilitates the estimation of multiple high-elevation wind speeds utilizing low-elevation measurements. This study built a graph data structure by assimilating prior knowledge, including the mountain topography and relative position of the anemometer. A custom-designed graph neural network model was designed to predict wind speed data at high-elevation regions based on wind speed data at a low-elevation region. Moreover, an enhanced UNet model was employed to predict wind field data at higher elevations utilizing predicted wind speed data. To improve the stability and performance of the model, the flux conservation equation was integrated into the loss function of the graph neural network. The results suggest that integrating mountain information and physical loss in the model can significantly improve the accuracy of wind speed data predictions. The graph neural network has remarkably enhanced the performance of the UNet model.