Deep Learning for Estimating Local Near-surface Wind in Mountainous Areas

A deep learning model using wind velocities obtained from mesoscale meteorological models is a promising approach for quickly and cost-effectively predicting near-surface wind speeds in mountainous areas. In this study, the mesoscale meteorological model (WRF) was implemented in four mountainous regions with complex terrain to obtain horizontal wind speed components at 10 m above ground level. The two horizontal components of wind speed obtained from WRF were used as input data for the convolutional neural network (CNN) model. The observation data in short periods (less than a month) was used in the training data, and the data in the other period (eleven months) as test data. The errors of the wind speed estimated by the trained CNN model become smaller than the wind speed directly estimated from WRF in all areas. The CNN model also improves wind directions that are difficult for WRF to predict. The present CNN model can predict wind speeds for an entire year by training on observation data in short periods. In addition, how the quality of the training date affects the CNN model performance is also investigated. This study is based on results obtained from a project, JPNP21001, commissioned by the New Energy and Industrial Technology Development Organization (NEDO).