A Diffusion-Based Framework for High-Resolution Precipitation Forecasting over CONUS

TL;DR

The study introduces a diffusion-based, uncertainty-aware framework for 1 km CONUS precipitation forecasts up to 12 hours, systematically comparing three residual-learning configurations (MRMS-only, HRRR-corrective, and hybrid). By leveraging EDM preconditioning and a dilated-attention U-Net, the approach achieves superior pixel-wise and spatiostatistical skill relative to HRRR, with the HRRR-Corrective configuration excelling at longer lead times and the Hybrid model leading at 1-hour forecasts. An uncertainty quantification strategy combines physics-informed lead-time offsets to deliver calibrated, interpretable bounds without costly ensembles. The work provides a scalable, regionally robust, and practically interpretable forecast framework with broad implications for emergency management and infrastructure planning. It also offers a rigorous, unified evaluation framework that contrasts data-driven and physics-informed inputs across CONUS and climatologically distinct regions.

Abstract

Accurate precipitation forecasting is essential for hydrometeorological risk management, especially for anticipating extreme rainfall that can lead to flash flooding and infrastructure damage. This study introduces a diffusion-based deep learning (DL) framework that systematically compares three residual prediction strategies differing only in their input sources: (1) a fully data-driven model using only past observations from the Multi-Radar Multi-Sensor (MRMS) system, (2) a corrective model using only forecasts from the High-Resolution Rapid Refresh (HRRR) numerical weather prediction system, and (3) a hybrid model integrating both MRMS and selected HRRR forecast variables. By evaluating these approaches under a unified setup, we provide a clearer understanding of how each data source contributes to predictive skill over the Continental United States (CONUS). Forecasts are produced at 1-km spatial resolution, beginning with direct 1-hour predictions and extending to 12 hours using autoregressive rollouts. Performance is evaluated using both CONUS-wide and region-specific metrics that assess overall performance and skill at extreme rainfall thresholds. Across all lead times, our DL framework consistently outperforms the HRRR baseline in pixel-wise and spatiostatistical metrics. The hybrid model performs best at the shortest lead time, while the HRRR-corrective model outperforms others at longer lead times, maintaining high skill through 12 hours. To assess reliability, we incorporate calibrated uncertainty quantification tailored to the residual learning setup. These gains, particularly at longer lead times, are critical for emergency preparedness, where modest increases in forecast horizon can improve decision-making. This work advances DL-based precipitation forecasting by enhancing predictive skill, reliability, and applicability across regions.

Figures (12)

• Figure 1: Regional divisions across the CONUS used for evaluation. These eight regions are adopted from Hill_Schumacher2021 to reflect climatological variability in convective hazards and support region-specific analysis of model performance.
• Figure 2: Overview of the proposed diffusion-based residual forecasting architecture. A dilated attention U-Net denoiser is trained under EDM preconditioning KarrasEA2022. Inputs concatenate HRRR predictors (condition), a Gaussian latent field scaled by the sampled noise level $\sigma$, and $\log \sigma$. The network predicts a single-channel residual correction map (data prediction, $x_{0}$-prediction).
• Figure 3: Pixel-wise evaluation metrics across independent testing months for lead times from 1h to 4h, covering the period from March 2023 to February 2024. Each letter on the x-axis corresponds to the first letter of the month (e.g., M = March, A = April, etc.). For lead times 1h to 3h, we compare the performance of the three proposed models: Data-Driven, HRRR-Corrective, and Hybrid—against the HRRR. For the 4h lead time, we focus solely on the HRRR-Corrective model, as it demonstrated the most consistent performance across earlier lead times. Shaded regions represent 95% confidence intervals. Metrics include MAE, FSS (27×27), CSI, and POD.
• Figure 4: Similar to Figure \ref{['fig_pixel_eval_1to4']}, but for lead times from 5 h to 8 h. Results show the best-performing longer-lead-time model (HRRR-Corrective) compared against the HRRR baseline at the 50th and 90th percentiles. Shaded regions denote 95% confidence intervals.
• Figure 5: Same as Figure \ref{['fig_pixel_eval_5to8']}, but for lead times from 9 h to 12 h.