Improvement of Tropical Cyclone Track Forecast over the Western North Pacific Using a Machine Learning Method

 The accurate tropical cyclone (TC) track forecast is necessary to mitigate and prepare significant damage. TC has been predicted by the numerical models, statistical models, and machine learning methods in previous researches. However, those models are separately used for TC track forecast, and historical data with satellite images were used as input variables for machine learning without forecast data from numerical models. In this study, we corrected the TC track forecast of a numerical model by artificial neural network (ANN). TCs that occurred from 2006 to 2015 over the western North Pacific were hindcasted by the Weather Research and Forecasting (WRF) model, and all categories of TCs except for tropical depression (i.e., tropical storm, severe tropical storm, and typhoon) from June to November were included in this study. We evaluated the performance of TC track forecast in terms of duration, translation speed, and direction compared with the best track data. The simulated positions of TCs at 24-hour, 48-hour, and 72-hour forecast lead time were used as variables for training and testing ANN. To optimize the number of neurons in ANN, simulated TCs were divided into two parts; TCs in 2006-2014 for ANN optimization and those in 2015 for a blind test. Also, the output selection method based on the forecast error of the WRF was applied to exclude the outlier of ANN results. By applying the output selection, the forecast error of ANN was further reduced than that of the WRF. As a result, ANN with the output selection method could improve TC track forecast by about 15% compared to the WRF. Also, the effect of ANN tended to increase when the forecast error of the WRF was large. The output selection method was particularly effective by excluding outliers of ANN results when the forecast error of the WRF was small.

※ This research was supported by Next-Generation Information Computing Development Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Science, ICT (NRF-2016M3C4A7952637).