Improving Ensemble Model Tropical Cyclone Track Forecast Using Machine Learning

Track forecast errors for tropical cyclones have been decreasing over the years, but there is still need for improvement, especially given recent challenging forecasts, such as hurricane Ian’s Florida landfall in 2022. The Global Ensemble Forecast System (GFS) and the European Center for Medium-Range Weather Forecast (ECMWF) models are used to predict the weather from days to weeks. Due to uncertainties in its initial conditions and physics, one approach for tropical cyclone track prediction is an average of the ensemble members; however, there can be relatively large biases that increase the predicted track error, especially for longer forecast lead times (3-5 days). This research aims to use machine learning to correct the bias and errors in the ensemble models for the 0-5 day forecast period over the western Atlantic and Gulf of Mexico regions.

The 30 GEFS ensemble members and 50 ECMWF ensemble members were obtained every 6 hours out to day 5 for the years 2008-2022. The observed tropical cyclones tracks from 2008 to 2022 were retrieved from the National Hurricane Center (NHC). The GEFS TC ensemble storm tracks were obtained from NHC (ftp://ftp.nhc.noaa.gov/atcf/archive/). The ECMWF ensemble storm tracks were obtained from THORPEX Interactive Grand Global Ensemble (at http://rda.ucar.edu/datasets/ds330.3/). The fifth generation ECMWF atmospheric reanalysis (ERA5) atmospheric reanalysis data was obtained to help relate the track errors to any spatial patterns in the winds, temperatures, and moisture at four levels (300, 500, 700, and 850 hPa). Next, a randomly selected 90% of the resulting dataset was used to train a CNN-LSTM machine learning (ML) model through k-fold cross validation to predict a 5-day track forecast. The training process was repeated 10 times with a different set of random storms (90% training and 10% validation) from the training dataset. The CNN-LSTM model was also trained to understand how the GEFS predicted track changes over previous three 6-hour forecasting cycles. Finally, model evaluation tests (using the remaining 10% independent set of storms) and permutation feature importance analysis to identify important variables were performed.

Using the GEFS across all forecast lead times, the average total track error of the CNN-LSTM model was less than that of the combined ensemble mean. The CNN-LSTM model is also better than the ensemble mean for a larger percentage of occurrences for the day 0-4 forecasts. The 300 hPa v component of wind, 500 hPa u component of wind, and 850 hPa u component of wind variables contributed most to the success of the CNN-LSTM model, which is consistent with the importance of the winds controlling the steering of the cyclone. Results with both the GEFS and ECMWF will also be shown. This model can eventually be set up to run operationally to help forecasters better predict these storms and thus provide more useful warning information to various stakeholders in the coastal regions.