Wavelet-Enhanced Temporal Fusion Transformer for Spatio-Temporal Climate Risk Forecasting

Climate change has intensified the need for accurate and interpretable models to assess spatio-temporal climate risks. Existing approaches often fall short in capturing complex interactions across spatial and temporal scales, limiting their flexibility and predictive performance. This study introduces the Wavelet-Enhanced Temporal Fusion Transformer (W-TFT), a novel deep learning framework that integrates Discrete Wavelet Transformation (DWT) with the Temporal Fusion Transformer (TFT) to improve climate risk prediction. DWT decomposes climate time-series data into multi-resolution components, enabling the model to distinguish between short-term variations and long-term trends. The architecture employs attention mechanisms, gated residual networks, and interpretable covariates—including physical and socio-environmental factors such as land use and population density—to selectively focus on significant features. Additionally, W-TFT incorporates uncertainty quantification, allowing for reliable multi-horizon forecasting essential for informed decision-making under climate uncertainty. Experimental evaluations show that W-TFT outperforms conventional models, achieving low RMSE (0.1254), MAE (0.0073), and MSE (0.0157), while effectively modeling complex spatiotemporal dependencies. The proposed framework offers scalable and adaptive capabilities for applications in climate monitoring, early warning systems, and sustainable policy development, making it a valuable tool for mitigating climate risks across diverse geographic contexts.