Implementation of a Convolutional Neural Network for 24-hour tropical cyclone track forecasting in the Bay of Bengal using fused historical and reanalysis data

Tropical cyclones are one of the most destructive natural disasters, inflicting severe damage on human life, infrastructure, and the environment. This study aims to present a framework for 24-hour tropical cyclone track forecasting in the Bay of Bengal (BoB) region. Deep learning algorithms, including Convolutional Neural Network (CNN), Numerical Weather Prediction (NWP) techniques, and models from the European Centre for Medium-Range Weather Forecasts (ECMWF), were utilized to capture the complex non-linear relationships between atmospheric conditions and storm trajectories. A predictive model was developed using a fused dataset that integrates more than 40 years of cyclone history from the IBTrACS database with atmospheric variables from the ERA5 reanalysis data. Through systematic experimentation with model architecture and hyperparameters, the best-performing configuration a shallow CNN, achieved a 24-hour Mean Absolute Error (MAE) of 141.2 km, calculated using the Haversine great-circle distance. This result demonstrates a significant improvement over the standard persistence model benchmark, which yielded an MAE of 186.11 km on the same test set. The model’s predictions for a hypothetical cyclone align closely with the observed climatological behaviour of the region. Forecast error analysis revealed clustering around 150 km with a mean error of 223.69 km, reflecting the multidimensional distribution of this value. This study underscores the potential of deep learning in cyclone prediction, offering a promising tool for real-time cyclone tracking and regional early warning. The proposed framework supports the future development of coastal risk management systems aimed at enhancing preparedness.