Increasing frequency and intensity of extreme weather events due to climate change necessitate accurate predictive models for mitigation and adaptation. This paper proposes a novel deep learning (DL) framework integrating Convolutional Long Short-Term Memory (ConvLSTM) networks and attention mechanisms for spatio-temporal prediction of extreme events (heatwaves, floods, hurricanes). Leveraging multi-source climate reanalysis data (ERA5, CMIP6 projections) and remote sensing imagery, the model captures complex non-linear patterns and teleconnections often missed by traditional Numerical Weather Prediction (NWP) and statistical methods. Evaluated on global datasets spanning 1980-2023, our approach reduces Root Mean Squared Error (RMSE) by 32% for heatwave intensity prediction and improves hurricane trajectory accuracy by 28% compared to ECMWF-IFS benchmarks. The model demonstrates robust skill in 2050 climate projections under RCP 8.5, highlighting its potential for climate resilience planning. Implementation challenges and scalability solutions are discussed.
Vengatesh et al. (Wed,) studied this question.