A Customized Multi‐Scale Deep Learning Framework for Storm Nowcasting

Abstract

Storm nowcasting is critical and urgently needed. Recent advances in deep learning (DL) have shown potential for improving nowcasting accuracy and predicting general low‐intensity precipitation events. However, DL models yield poor performance on high‐impact storms due to insufficient extraction and characterization of complex multi‐scale spatiotemporal variations of storms. To tackle this challenge, we propose a novel customized multi‐scale (CM) DL framework, including a flexible attention module capturing scale variations and a customized loss function ensuring multi‐scale spatiotemporal consistency. The CM framework was applied to the storm event imagery data set (SEVIR). Representative cases indicate that the CM framework preserves the shape of storms and adequately forecasts intense storms even for longer predictions. The quantitative evaluation shows that all models applying our framework can improve skill scores by 8.5%–42.6% for 1‐hr nowcasting. This work highlights the importance of modeling multi‐scale spatiotemporal characteristics of meteorological variables when using DL.

Reliable precipitation nowcasting using probabilistic diffusion models

“…State-of-the-art data-driven precipitation nowcasting approaches take advantage of deep generative models to yield probabilistic forecast. However, these methods (Ravuri et al 2021, Yang andYuan 2023), mostly based on generative adversarial nets (Goodfellow et al 2014), often encounter severe approximation/optimization errors, making it difficult to achieve a probability forecasts. This challenge is evident in the metrics discussed in section 5.2, where DGMR performs relatively worse than the diffusion model in terms of CSI and CRPS scores.…”