재해석 자료를 이용한 전세계 풍속의 회귀분석

Abstract

Although the reanalysis data has low spatial resolution, it can be used for long-term correction of wind resources and high-resolution downscaling in conjunction with numerical weather prediction or computational fluid dynamics. In this study, regression analysis of the global wind speed as a function of topographic factors was carried out, and the possibility of its use in highresolution downscaling was tested. According to the regression results by various morphometric features, the fitness of multiple linear regression, machine learning models, i.e., neural network and random forest models, showed R2 of 0.71, 0.95, and 1.00, respectively. Among the topographic factors, latitude, cell area, terrain elevation, longitude, and terrain openness were found to have the highest explanatory capability in the order. Compared to the conventional neural network, the modified twin neural network model has a performance improvement effect regarding imbalanced dataset. Nevertheless the neural network structure used in this research was not precise to reproduce non-linearity of the given data, however, which was possible with introduction of the random forest model.