The TC track correction algorithm based on the depth series model (conv-lstm) (Figure 1) has been preliminarily developed, and the deep learning method based on ECMWF ensemble prediction has been designed. The learning data uses the typhoon prediction path of 51 members of EC ensemble prediction and 52 models of deterministic prediction, as well as the real-time typhoon positioning of the National Meteorological Centre. Based on the 6-hour prediction error feedback mechanism, the typhoon track AI prediction model (AI-TYTEC) is established. The model scrolls and corrects the prediction weight of each member every 6 hours, and recalculates the corresponding objective path prediction according to formula 1. Based on the evaluate result, for all typhoons from 2012 to 2019, the depth learning model is used to compared with the numerical model (or ensemble average prediction) with the best performance in the same period. The result shows that the 24-hour prediction error of the depth learning model is about 15 ~ 20% less than that of the best numerical model. Furthermore, the AI-TYTEC can extend the TC track forecast up to 7 days.

                     (1)

Fig 1. Process of the AI-TYTEC

Fig 2. Average TC track forecast error of AI-TYTEC in 2020