Tropical cyclone track prediction model for multidimensional features and time differences series observation

Abstract

Tropical Cyclones (TCs) are highly destructive weather phenomena that can cause significant social and economic damage. With the development of meteorological monitoring technology and the updating of database, accurately forecasting the track of TC movement is one of the effective ways to minimize losses. However, traditional movement track forecasting methods suffer the disadvantages of low efficiency and low accuracy. To address the these problems, a novel Convolutional Neural Network-Temporal Convolutional Network (CNN-TCN) model based on Multidimensional Features and Time Difference Series (MT-CNN-TCN) is presented in this paper. First, different types of meteorological data are processed and then the feature differences between adjoining moments are extracted. Second, a two-branch structure based on Two Dimensional Convolutional Neural Network (2DCNN), 3DCNN and TCN is taken to effectively integrate different types of meteorological features to strengthen its forecasting effect. Finally, experiments are conducted using Northwest Pacific TC data from years 2000–2019. Test results show that the proposed model MT-CNN-TCN can perform well at all three forecast periods (12 h, 24 h, and 48 h), with a significant improvement in accuracy by 7 %, 13 %, and 16 % respectively, compared with current forecasting methods such as Long Short Term Memory (LSTM).