<![CDATA[The limited ability of Numerical Weather Prediction (NWP) models to capture nonlinear dynamics and atmospheric uncertainty remains a major challenge in improving tropical cyclone forecasts, particularly over the eastern Indian Ocean. This study evaluates a Multi-Task Learning approach based on several Recurrent Neural Network (RNN) variants, namely LSTM, BiLSTM, GRU, and BiGRU, to simultaneously predict three key cyclone components: position (latitude and longitude), wind intensity, and cyclone category. Historical IBTrACS data from 2000 to 2025 with a 3-hour temporal resolution are used as model input, employing 48-hour sequences to forecast cyclone conditions at lead times of 12, 24, 48, and 72 hours. The results show that all models achieve stable convergence during training. At a 12-hour lead time, the BiLSTM model delivers the best performance, with a mean position error of 83.53 km and a Hit Rate of 0.966, outperforming the other models. For longer lead times (24–72 hours), the BiGRU model demonstrates the most stable positional accuracy, exhibiting the lowest error degradation as the forecast horizon increases. In addition, wind intensity predictions remain robust, with a Mean Absolute Error (MAE) below 4.6 knots up to 72 hours. These findings highlight the potential of multi-output RNN-based models to support more adaptive and efficient tropical cyclone forecasting systems.]]>
Copyrights © 2026
