Short-term prediction of the intensity and track of tropical cyclone via ConvLSTM model

Abstract

As one of the most destructive natural disasters, Tropical Cyclone (TC) can cause severe casualties and economic losses almost every year in TC-prone areas. For the prevention and reduction of TC-induced disasters in wind engineering, it is of vital importance to forecast TC's intensity and track both accurately and efficiently. Although great achievements have been made in this field, it remains a challenge to balance accuracy and efficiency during the forecast. In recent years, deep learning techniques have gained fast development and demonstrated huge application potential in various fields. This article presents a study on the short-term prediction of TCs in Northwest Pacific basin via the newly developed Convolutional Long and Short Term Memory (ConvLSTM) network which is able to extract both time-related correlations (involved in the sequence of a single parameter) and parameter-related correlations (among different parameters) from input information of different feature parameters. Results demonstrate that the proposed ConvLSTM model has good performance in terms of prediction accuracy and working stability. Through comparison, it is also suggested the proposed model works better than the standard LSTM model. Meanwhile, unlike most dynamic methods which require expensive computational resources, the proposed model can be operated conveniently and economically. Thus, it can be readily applied for engineering practices.