Western North Pacific tropical cyclone track forecasts by a machine learning model

Abstract

An ensemble machine learning model for tropical cyclone (TC) track forecasts over the Western North Pacific was developed and evaluated in this study. First, we investigated predictors including TC climatology and persistence factors which were extracted from TC best-track dataset and storm’s surrounding atmospheric conditions which were extracted from ERA-Interim reanalysis. Then, we built a Gradient Boosting Decision Tree (GBDT) nonlinear model for TC track forecasts, in which 30-year data was used. Finally, using tenfold cross-validation method, the GBDT model was compared with a frequently used technique: climatology and persistence (CLIPER) model. The experimental results show that the GBDT model performs well in three forecast times (24 h, 48 h, and 72 h) with relatively small forecast error of 138, 264, and 363.5 km, respectively. The model obtains excellent TC moving direction aspects. However, the model is still insufficient to produce aspects of storm acceleration and deceleration, with mean moving velocity sensitivities all less than 60%. Nevertheless, the model obtains much more robust and accurate TC tracks relative to CLIPER model, where the forecast skills are 17.5%, 26.3%, and 32.1% at three forecast times, respectively. The presented study demonstrates that the GBDT model could provide reliable evidence and guidance for operational TC track forecasts.