Local-scale wind forecasting for tropical cyclone early warnings

This paper introduces LiveCyc, a machine learning approach that can augment any tropical cyclone (TC) track and intensity forecast with a probabilistic local-scale wind forecast. The authors note that existing forecasting products are critical for informing safety warnings and evacuation measures at the county scale but are not designed to represent experience on the ground at the scale of individual neighbourhoods. LiveCyc transforms TC forecasts into probabilistic, local-scale (1-km) wind predictions that account for topography and surface roughness and generates 1,000 realizations in real time to quantify forecast uncertainty. 

The study evaluates LiveCyc using an extensive back-testing dataset covering 16 major U.S. hurricane landfalls between 2017 and 2024 and 1,840 surface wind observations. The analysis shows that the modelled wind distributions are well calibrated and without important systematic biases, with 87% of observations falling within the 90% confidence interval for the mid-range forecast window. The authors demonstrate that LiveCyc can support objective optimization of protective action triggers through cost-loss analysis, allowing users to calibrate warning thresholds based on the relative costs of preparedness and potential damage. They conclude that LiveCyc can help design automated alert systems for applications such as protection of high-value assets, infrastructure shutdowns and deployment of first responders, while future work should incorporate additional storm structure parameters and further investigate sources of uncertainty.