Prediction of Tropical Cyclone Trajectories using Echo State Networks

In the northwestern Pacific, Tropical Cyclone (TC) is generated around the equator and develops into typhoon while moving north. The resultant typhoons cause huge damage such as heavy precipitation, storm surge, and inundation at coastal regions. Therefore, accurate prediction of TC's trajectory is important to mitigate the coastal damage. The existing methods to predict TC's trajectory can be classified into four types, i.e., numerical, statistical, statistical-numerical, and ensemble. The numerical method solves physical equations numerically while the statistical method is based on historical relations between TC's characteristics. The statistical-numerical method blends both numerical and statistical method and the ensemble method finds an optimal model from a combination of a few simulations. Recently machine learning techniques have been applied for trajectory prediction. Specifically, Recurrent Neural Networks (RNNs) are an effective technique to simulate dynamic systems. Echo State Networks (ESNs) are a branch of RNNs which improve the performance by adopting reservoir computing scheme. In this study, ESNs are used to predict TC's trajectory in the northwestern Pacific. Best track data of total 2011 TCs occurred from 1945s to 2014s is obtained from Joint Typhoon Warning Center. These data provide various TC properties in 6-hourly interval. Among those properties, latitude, longitude, maximum wind speed, and minimum air pressure data are used as inputs. Latitude and longitude are selected as outputs and converted to represent trajectory. The accuracy of 6-hours and 12-hours prediction is evaluated.