AI-driven ensemble forecasting of extreme wind gusts: Random Forest modeling and case studies from the western Mediterranean

Abstract

Forecasting extreme wind-gust (WG) is challenging because the driving processes evolve rapidly and non-linearly. We tailor a stochastic random forest (RF) classifier that predicts extreme WGs occurrences along the western Mediterranean coast and ranks the meteorological factors that trigger WG intensification. Uniquely, each atmospheric variable—pressure, humidity, wind-direction and temperature—enters the model twice, both as an instantaneous value and as its trailing 24-h mean, allowing the RF to learn momentary forcing and short-term evolution in a single step. These dual-time predictors are evaluated at 11 height levels (15–500 meters), yet the single (15 meters) version already delivers robust skill. Local and regional models attain 85% mean precision across all lead-times, maintaining robust skill (~80% precision at the 48-h horizon) with false-alarm ratios below 20% and well-calibrated probabilistic reliability. The ensemble’s internal spread provides a first-order measure of forecast uncertainty, making the framework risk-oriented and ready for early-warning forecasting workflows, with direct value for civil-protection agencies and regional adaptation planning. Variable-importance analysis highlights barometric-pressure tendencies and humidity as the dominant early-warning cues. The lightweight, interpretable, and data-efficient design is well poised for broader use across other wind-prone regions—subject to local data availability and validation—thereby strengthening climate-resilience and disaster-risk management.