Mixture Density Networks for Tropical Cyclone Tracks Prediction in South China Sea

The forecasting of Tropical Cyclone (TC) Tracks in the South China Sea is important to cope with the associated disasters. But increasing its forecasting accuracy is a hard thing due to many factors. The main objective in the presented study is to develop models to deliver more accurate forecasts of TC Tracks over the South China Sea. The model proposed in this study is the TC Tracks Probability Forecasting Framework based on the Mixture Density Network (MDN). The TC Tracks Probability Forecasting Framework calculates the joint probability distribution of latitude and longitude and consists of latitude MDN and longitude MDN. MDN is a method that models the conditional probability distribution of the target data by the conventional neural network and mixture model. Forecast error is measured by calculating the distance between the real position and forecast position of TC. A decrease of 19.49 km in mean forecast error is obtained by our proposed model compared to the stepwise regression model, which is widely used in TC Tracks forecast. What's more, the model gives the probability distribution of each track prediction. This message can be used well in the TC track forecast cone.