Enhancing Extreme Weather Forecasting via Dynamically Weighted MSE

ICLR 2026 Conference Submission15922 Authors

Keywords: Extreme Climate Forecasting, Time-series Forescasting

Abstract: Data-driven weather forecasting empowered by deep learning has shown superior performance compared to traditional physics-based dynamical models. However, conventional training objectives (like Root Mean Squared Error (RMSE)) primarily focus on minimizing average prediction errors, often resulting in oversmoothed forecasts that fail to capture critical extreme weather phenomena, including heavy precipitation, hurricanes, and other high-impact events. To overcome this limitation, we propose a robust loss function, named Dynamically Weighted MSE (DW-MSE), that adaptively reweights training samples to better learning on extreme weather events. Specifically, we introduce a dual-branch meta-network alongside the prediction network to dynamically generate sample weights: one branch captures spatiotemporal dependencies across climate variables, while the other learns from training losses. Guided by a small set of validation samples, the meta-network can be jointly optimized with the prediction network via an efficient bi-level optimization strategy, which provides the fast convergence with the approximated first-order information. Overall, our framework is able to accurately identify and assign greater importance to extreme weather samples without manually designing reweight function and any prior knowledge. Extensive experiments in both training-from-scratch and fine-tuning climate models demonstrate that Meta-MSE consistently outperforms existing approaches in forecasting extreme weather.

Primary Area: learning on time series and dynamical systems

Submission Number: 15922