Machine Learning Approaches to Reducing Tropical Cyclone Prediction Errors

Tropical cyclones (TCs) in the North Atlantic region are predicted for the June 1-November 30 season using predictors from Atlantic, Indian and Pacific Oceans sea surface temperature anomalies. Here, the aim is to reduce TC seasonal prediction errors for the number of named cyclones, the number of major cyclones and the accumulated cyclone energy. The prediction errors are realized by applying support vector regression (SVR) to an initial predictor pool and using the model prediction errors to iteratively identify additional attributes that reduce those errors. Prediction errors from this approach are compared with those from an existing, statistical seasonal prediction model, developed at Colorado State University (CSU). The SVR approach was optimized using attribute selection with wrapper selection techniques and by testing various kernels over a range of complexity and other model parameters. Results of the comparison between seasonal SVR and the CSU TC model for the number of named storms indicate that proper attribute selection lowers prediction errors significantly. Although the models were trained differently, the CSU predictions were available for a relatively long record, thereby providing a baseline to examine the potential for the machine learning techniques. Compared with the CSU model, the SVR model increases correlations, lowers mean absolute errors (MAE) and root mean square errors (RMSE) (e.g., between prediction and observed annual TC count from 0.62 to 0.83; MAE is reduced from 2.9 to 1.8 and the RMSE drops from 3.8 to 2.7). Furthermore, the approach of using prediction errors to improve machine learning models is flexible and can be adapted readily to other TC basins.