24AI Advancing Extreme Precipitation Prediction over CONUS Using Diffusion Models with Spatial and Temporal Uncertainty Quantification

Extreme events are rare phenomena that significantly deviate from the norm and have substantial impacts on the environment, infrastructure, and human life. These events include, but are not limited to, extreme precipitation, heatwaves, droughts, hurricanes, and flash floods (McPhillips et al., 2018; Trenberth et al., 2015). However, the exact definition of extreme events, such as extreme precipitation, varies considerably across the Continental United States (CONUS) (Hill & Schumacher, 2021; Gourley et al., 2013; Gourley & Vergara, 2021). Precipitation events may be defined by the amount of total rainfall, but what is extreme in one region may not be in another. Furthermore, within the National Weather Service (NWS) forecast offices, there is no universally standardized criterion for issuing warnings for such extreme rainfall events.(Hill & Schumacher, 2021; Gourley et al., 2013; Marjerison et al., 2016). Given the substantial impacts of extreme weather events and the challenges associated with accurately predicting them, such as the inherent unpredictability of convective processes, the rapid development of storms, and the influence of complex topography, there is a clear need for further research in this area. The accuracy of the HRRR in forecasting extreme rainfall events is limited due to difficulties in modeling small-scale convective processes (James, & Schumacher, 2024). Additionally, the HRRR model has been shown to have certain biases that can lead to the underprediction of precipitation, particularly in regions with complex terrain or rapidly evolving weather systems (Lee et al., 2019, Min et al., 2021). These limitations underscore the necessity for ongoing research to improve the prediction of extreme precipitation events and to develop more robust forecasting tools that can better serve communities at risk.

Various machine learning approaches have been utilized to forecast extreme rainfall events (Sangiorgio et al., 2019; Chkeir et al., 2023; Hill & Schumacher, 2021). However, the majority of these approaches focus on making binary predictions, determining only whether an extreme event will occur. While this binary information is valuable, providing quantitative forecasts of the expected rainfall amounts would offer significant additional benefits for city planners and emergency managers to make more informed decisions, ultimately enhancing preparedness and response strategies for extreme weather events. Recent advancements in deep learning techniques are transforming the field of weather forecasting by demonstrating significant potential to outperform traditional numerical methods (Andrychowicz et al. 2023). In particular, diffusion models, a type of generative AI, are gaining popularity as a promising tool for predicting and generating atmospheric phenomena (Addison et al., 2024; Nai et al., 2024). These models offer new avenues for enhancing the accuracy and reliability of weather forecasts, leading the way for innovative approaches in atmospheric science.

This research introduces the use of diffusion models to predict extreme rainfall events across the CONUS. By leveraging the Multi-Radar Multi-Sensor (MRMS) rainfall dataset as the ground truth and incorporating multiple atmospheric variables as inputs, we developed two distinct models: a CONUS-wide model and a regionally subdivided model. This dual approach offers valuable insights into the atmospheric variables that influence extreme rainfall, thereby deepening our understanding of the physical processes driving extreme precipitation events. Furthermore, a customized loss function specifically designed for extreme rainfall prediction was developed. A key innovation of this research, compared to the current state of the art, is the incorporation of spatial and temporal uncertainty quantification into the predictions. This additional layer of uncertainty information could be valuable for emergency and city managers, as it provides not only a predicted forecast but also a confidence range for both the timing and location of extreme rainfall. Future efforts will focus on integrating these improved rainfall predictions with hydrological models to better predict flooding events and refine the accuracy of both spatial and temporal forecasts. Equipping end users with the tools needed for more informed, timely decisions and ultimately strengthening community resilience against extreme weather events.