A deep learning approach for wind downscaling using spatially correlated global wind data

Windforecasting is an integral part of wind energy management as a crucial instrument for predicting wind patterns in coastal
areas. One commontechniquetopredict the wind field in a specific area is the dynamical downscaling method, which is based
on a physical model and requires a substantial computational cost. Instead, this study proposes a novel approach for wind
downscaling based on deep learning techniques as a substitution for a dynamical downscaling method. Our methodology
starts with generating a high-resolution wind dataset by dynamically downscaling global climate data using RegCM4.7.
Then, we employ a feature selection technique to identify the optimal global wind data points that exhibit a strong spatial
correlation with the local wind data of interest. The selected features from global climate data and the target from the high
resolution wind data are used to develop a machine learning-based model to predict wind variability in a specific location. We
consider various models, namely multilayer perceptron (MLP), AdaBoost, XGBoost, long short-term memory (LSTM), and
bidirectional LSTM (BiLSTM), and conduct performance analysis to find an optimum model. The BiLSTM model has been
shown to be the most optimal algorithm for wind downscaling among various machine learning models. We also evaluated
the model’s performance by conducting a comparative analysis between its predictions and the observed wind data gathered
from Jakarta and Meulaboh. This analysis yields significant insights into the accuracy and applicability of our methodology.
Our approach reveals a strong correlation coefficient of 0.963 and a low root mean square error (RMSE) of 0.476. These
results highlight the efficacy of our method in correctly downscaling wind data