Using Machine Learning Methods to Bias Correct Tropical Cyclone Intensity Forecasts

Abstract: In recent decades forecasts of Tropical Cyclone (TC) tracks have improved significantly. This has enabled earlier warnings, allowing local populations to make better preparations ahead of landfall, and ultimately saving lives. While TC track forecasts have improved dramatically, TC intensity has proved more challenging, and progress has been slower. The resolution of operational numerical weather prediction (NWP) models (10-25km) is not sufficient to capture the mesoscale processes that drive intensification of TCs. A recent example of this was Hurricane Otis which rapidly intensified prior to making landfall as a category 5 storm in Acapulco, Mexico. All global NWP (and ensembles) failed to capture the rapid intensification of Otis, predicting landfall as a weak category 1 hurricane or tropical storm (even at short lead times of 24-48 hours). Failures of NWP forecasts to capture TC intensification reduce the amount of time available to implement readiness and evacuation strategies, leading to more widespread and damaging impacts on vulnerable communities.

As a small group of scientists and software engineers at the Met Office, we have recently been working on using an extreme-gradient boosting (XGBoost) algorithm to bias correct TC intensity forecasts and provide more useful forecast information. We use the International Best Track archive for climate stewardship (IBTracs) database of observed TCs in combination with ERA-5 reanalysis data to train an ML model that we can use to apply a bias correction to TC forecasts. Initial benchmarking of our ML bias corrected output shows a reduction in intensity errors compared with our operational global NWP model. The ML model is trained using physically relevant predictors such as environmental wind shear, sea surface temperatures and environmental moisture. We further explore how best to utilise probabilistic ML techniques to show uncertainty in our tropical cyclone ML predictions.
 

Biography: Richard completed his PhD at the University of East Anglia, focussing on weather and climate in West Antarctica. He joined the Met Office Regional Model Evaluation and Development team in 2020, first focussing on assessing the latest regional model configuration – which will soon become operational. His recent work has utilised high-resolution simulations across very-large tropical domains – assessing the changes in model behaviour we are likely to see as we move to finer resolution. Richard also has a keen interest in tropical cyclones, both through exploiting high-resolution ensembles and utilising machine learning methodologies to improve predictions of tropical cyclone intensity