Track Prediction of Tropical Cyclones Using Long Short-Term Memory Network

Accurate track prediction of tropical cyclones is of utmost importance to minimize loss of lives and property. We propose a Long Short-Term Memory network based Recurrent Neural network that uses data of the first few hours of cyclones and predicts their track for the next several hours with high accuracy. The model uses variables like estimated central pressure, maximum sustained surface wind speed, latitude, and longitude, which are available at a regular time interval. The output of the model is a GridID, which has been obtained by a grid function applied on latitude and longitude ranges. The model has been applied to hurricanes in the Atlantic ocean, and tropical cyclones in the north Indian ocean. The model performance has been compared with existing state-of-the-art results. The proposed model outperforms the existing models in terms of mean absolute error, time-span of prediction, and training time.