Machine Learning-Based Estimation of Peak Wind Gusts Utilizing a Global Reanalysis Dataset

Estimating peak wind gusts (Wp) is crucial for various meteorological, wind energy, and wind engineering applications. However, there is often a scarcity of on-site observed Wp data. To address this issue, a recent study [Kartal et al. (2023), Wind Energy Science] proposed decision tree-based approaches, including random forest and gradient boosting machines, within a measure-correlate-predict framework. Using these machine learning (ML) techniques, the authors successfully estimated long-term wind gust records with the ERA-5 global reanalysis dataset. However, the performance of these ML approaches was somewhat limited under extreme conditions (i.e., Wp > 20 m/s).

In the present study, we extend the work of Kartal et al. (2023). In addition to decision tree-based approaches, we utilize a state-of-the-art attention-based deep learning framework, known as TabNet. Furthermore, we perform novel feature engineering by extracting various stability indices (e.g., Lifted Index, K-Index) and other parameters from ERA-5-generated vertical profiles. The usefulness of these additional features under extreme conditions is demonstrated using the SHAP approach. Multi-year observational data from various stations within the New York State Mesonet are used both for training the ML models and for rigorous validation.

Keywords: Deep Learning, Feature Importance, Peak Wind Gust, SHAP, TabNet