Three-dimensional temporal-spatial attention for tropical cyclone forecast

Abstract

The tropical cyclone (TC) is a strong and highly destructive tropical low-pressure vortex that often brings disasters such as strong wind, heavy rain, and storm surge. The formation and intensity forecast of TC is very important for TC disaster warnings. In this paper, we propose a three-dimensional temporal-spatial (3D-TS) attention TC forecast model based on deep learning, which considers the temporal-spatial relationship between TC variables on the basis of machine learning methods. The model introduces 2D convolutional neural networks (2DCNNs) and 3D convolutional neural networks (3DCNNs) to learn oceanographic variables and atmospheric variables of TC, while utilizing Long Short-Term Memory (LSTM) for capturing the temporal correlation in TC’s evolution process, and introduces 3D-TS to grasp the temporalspatial characteristic of TC and enhance the model’s precision. Through experiments, it was found that the model’s performance in TC formation and intensity forecast is better than many existing numerical forecasting methods, statistical forecasting methods, and machine learning methods. We validated the model on the Western Pacific TC dataset, in the TC formation forecast experiment, the model can achieve a maximum accuracy rate of 86.1% and an AUC value of 92.5%. In the TC intensity forecast experiment, the minimum error is 7.1 kt, showing that we achieved state-of-the-art results.