Improving Wind Gust Prediction with the Combination of WRF and Machine Learning Algorithms

Wind gusts, one of the most unpredictably occurring meteorological elements, have a significant impact on the generation of wind energy and the occurrence of power outages. Numerical weather prediction models (NWP) such as Weather Research and Forecasting (WRF) model use post processing tools and parameterizations to estimate wind gust. However, it has been observed that gusts obtained from such post processing techniques often significantly differ from reported gusts. The objective of this study is to evaluate how machine learning (ML) algorithms perform when compared to WRF in terms of gust prediction. Forty eight storms exhibiting high wind were simulated using WRF 3.8, over the northeastern United States. Hourly WRF model outputs such as wind speed and temperature gradient at different levels, planetary boundary layer height (PBLH), wind direction and frictional velocity have been used as predictors and observed hourly wind gust was the target variable. We trained machine learning (ML) models using two tree-based methods, Random Forest (RF) and XGBoost (XGB), and both models outperformed the gust prediction of the WRF unified post processing (UPP) tool. Results from a leave-storm-out cross-validation performed to prevent overfitting showed that even generalized linear model (GLM) performed better than WRF and showed similar predictive accuracy as RF and XGB. Feature importance plots showed that wind speed at different levels were the most influential predictors for gust. However, the performance of these ML models was still not satisfactory when wind gust exceeded 15 m/s due to limited number of gust observations in the higher range. Resampling technique SMOTE was used to account for right skewed gust observations and showed slight improvement in model performance for higher gust values (>15 m/s). Finally, learning curves have been constructed to quantify sensitivity of ML models' prediction accuracy to different training sample sizes.