Predictive Monitoring of Extreme Events for Enhanced Bridge Safety Using Hyperlocal ML-Powered Wind Nowcasting

This study introduces a hyperlocal tool for short-term extreme wind nowcasting (i.e., less than 6 h forecasting), designed to enhance decision-making and risk mitigation for bridge operations. Long-span bridges are highly susceptible to the adverse effects of extreme weather and require reliable short-term forecasts to address potential hazards proactively. The proposed framework integrates advanced machine learning (ML) techniques with a blend of hyperlocal and global sensor data to deliver real-time, high-accuracy meteorological predictions. Real-time measurements from local sensors installed on the target bridge - the 1,624 m main-span Great Belt Bridge suspension in Denmark - are combined with data from a global weather station to provide a comprehensive view of impending meteorological conditions. A supervised ML algorithm, trained on historical data and tailored for hyperlocal meteorological prediction, processes these multi-source inputs to generate probabilistic wind forecasts in the 0 - 180 min horizon. The model is implemented using open-source tools in Python and employs an ensemble of state-of-the-art time-series forecasting architectures to produce both deterministic and probabilistic predictions, including calibrated confidence intervals. Validation against operational forecasts from the Danish Meteorological Institute (DMI) demonstrates that the proposed tool achieves substantially lower root mean squared error (RMSE), higher precision and F1-scores for extreme wind event detection, and markedly improved wind direction accuracy. These results highlight the potential of ML-powered hyperlocal nowcasting as an operational decision-support tool for predictive monitoring and enhanced safety of windexposed bridge infrastructure.