Improved Tropical Cyclone Track Simulation over the Western North Pacific using the WRF Model and a Machine Learning Method

Abstract

Accurate tropical cyclone (TC) track simulations are required to mitigate property damage and casualties. Previous studies have generally simulated TC tracks using numerical models, which tend to experience systematic errors due to model imperfections, although the model accuracy has improved over time. Recently, machine-learning methods have been applied to correct such errors. In this study, we used an artificial neural network (ANN) to correct TC tracks hindcasted by the Weather Research and Forecasting (WRF) model from 2006 to 2018 over the western North Pacific. TC categories that are stronger than tropical depressions (i.e., tropical storms, severe tropical storms, and typhoons) were selected from June to November, and a bias correction was made to target TC positions at 72 h. The WRF-simulated tracks were used as input variables for training and testing the ANN using the best track and reanalysis data. To obtain a reliable corrected result, the number of neurons in the ANN structure was optimized for TCs during 2006–2015, and the optimized ANN was verified for TCs from 2016–2018. Because the performance of the numerical model differed according to the TC track, the ANN was assessed by cluster analysis. The results of the ANN were analyzed using k-means clustering to classify TCs into eight clusters. Overall, ANN with post-processing improved the WRF performance by 4.34%. The WRF error was corrected by 8.81% for clusters where the ANN was most applicable.