Methodological Changes to the Attention ResUNet Hourly Precipitation Postprocessor

Thomas M. Hamill

physics.ao-phSep 29, 2026

Methodological Changes to the Attention ResUNet Hourly Precipitation Postprocessor

Authors: Thomas M. Hamill

Organizations: The Weather Company, Atlanta, Georgia

Abstract

This note is a technical companion to a previously published preprint describing an Attention Residual U-Net that postprocesses deterministic forecasts from The Weather Company's Global and Regional Atmospheric Forecast (GRAF) model into probabilistic hourly precipitation forecasts. It documents what has changed in that method since publication. Feature-wise Linear Modulation conditioning on calendar season and forecast lead time is used to produce a single trained model for each season, replacing 192 separately trained per-month, per-lead checkpoints. Lead time is extended from 48 to 72 h. Two new input channels are used, per-pixel local solar hour and a static, monthly-varying precipitation climatology. During verification, the climatological reference against which the Brier Skill Score is computed now has an added diurnal dimension, on top of the monthly resolution it already had. Brier Skill Score and reliability are compared between the new vs. the previous training. Forecasts generated with the new training show a modest, consistent improvement of the current training over the original.

Figures & tables

Axis	Original manuscript	Current
Input channels	7: GRAF precip, terrain deviation, GFS RH, precip $\times$ terrain, precip $\times$ RH, terrain gradient (zonal, meridional)	10: the same 7, plus per-pixel local-solar-hour sine/cosine and a static precipitation climatology channel
Lead/date conditioning	None; lead and calendar dependence handled implicitly by training a separate checkpoint for each	FiLM conditioning vector [sin(day-of-year), cos(day-of-year), lead/72h] applied at each encoder/decoder stage
Checkpoint granularity	One checkpoint per (calendar month $\times$ 3-h lead step), or 192 checkpoints; nearest-lead weights used for leads that are not even multiples of 3 h	One checkpoint per calendar season (DJF/MAM/JJA/SON), 4 total
Lead-time coverage	+3 h to +48 h	+3 h to +72 h, continuously via FiLM
Training-data pipeline	One precomputed patch set per (month, lead) checkpoint, drawn from four 60-day recency/seasonal date windows	Precomputed, season-and-lead-pooled zarr patch archives spanning all years and all leads within a season, read with a chunk-aware shuffle sampler
Inference procedure	Fully-convolutional single forward pass over the padded CONUS domain	Same, but only achievable for the FiLM-conditioned model because solar-hour is a per-pixel input rather than a spatially-varying FiLM term (§ 2.4 )

Table 1: Summary of methodological changes since the original manuscript. See subsections 2.1 – 2.4 for details.

Figure 1: Brier Skill Score vs. lead time for the Attention ResUNet’s own postprocessed (Gamma-mixture) probabilities, original training (orange) vs. current training (blue), stratified by terrain roughness as in the original manuscript’s own BSS figure. Shaded bands are 90% Hodges-Lehmann confidence intervals (see text). Sample: 2025 March/June/September/December, 6-hourly, matching the original manuscript’s own verification sample.

Figure 2: Reliability diagram for the Attention ResUNet’s own postprocessed (Gamma-mixture) probabilities, original training (orange) vs. current training (blue), at +12 h lead, for exceedance of 0.25, 1.0, and 5.0 mm. Inset panels show forecast-probability usage frequency. Sample as in Figure 1 .

Figure 3: As Figure 2 , but at +48 h lead.

Figure 4: Washington-state atmospheric river case, IC 2026031112, +12 h. (a) Raw GRAF 1-h accumulated precipitation forecast, (b) Attention ResUNet probability of exceeding 0.25 mm, original training, and (c) the same, current training.

Explore similar work

Stable Attention Response for Reliable Precipitation Nowcasting
May 13, 2026Penghui Wen, Zexin Hu, Sen Zhang +6Precipitation Nowcasting Atmospheric Dynamics
School of Computer Science, The University of Sydney Australia · Independent Researcher Australia · School of Life and Environmental Science, The University of Sydney Australia +3
Diffusion-Based Refinement for Kilometer-Scale Probabilistic Precipitation Nowcasting
Aug 31, 2026Dohyun Park, Changhoon Song, Tengyuan Chang +2Precipitation Nowcasting Rainfall
Department of Mathematical Sciences, Seoul National University, Seoul, Republic of Korea. · Research Institute of Mathematics, Seoul National University, Seoul,2026 Republic of Korea. · Department of Environmental Management, Seoul National University, Seoul, Republic of Korea.
SwAIther-Precip: Lead-Time-Aware Bias Correction Enables Kilometer-Scale Downscaling of Global AI Precipitation Forecasts over Switzerland
May 15, 2026Dan Assouline, Erwan Koch, Federico Amato +5Dynamical Downscaling Artificial Intelligence Weather Models
Swiss Data Science Center, ETH, Zürich, Switzerland · Expertise Center for Climate Extremes, University of Lausanne, Lausanne, Switzerland · Faculty of Geosciences and Environment, University of Lausanne, Lausanne, Switzerland +1

May 13, 2026, cs.LGJ/K move · Enter open · S save

Stable Attention Response for Reliable Precipitation Nowcasting

Penghui Wen, Zexin Hu, Sen Zhang, Patrick Filippi+5

School of Computer Science, The University of Sydney Australia · Independent Researcher Australia · School of Life and Environmental Science, The University of Sydney Australia+3

Precipitation Nowcasting Atmospheric Dynamics Dynamic Attention Responses

Open paper AskPDF arXiv

Precipitation nowcasting remains challenging due to the highly localized, rapidly evolving, and heterogeneous nature of atmospheric dynamics. Although recent methods increasingly adopt attention-based architectures in both unimodal and multimodal settings, they mainly emphasize stronger representation learning and prediction capacity, while paying less attention to the stability of attention responses across samples. In this work, we show that cross-sample instability of attention-response energy is an important and previously underexplored source of forecasting unreliability. Empirically, inaccurate forecasts are associated with larger attention-response energy variance across heads and layers. Theoretically, we show that cross-sample variability can propagate through self-attention, and enlarge a lower bound on prediction error. Based on this insight, we propose HARECast, a Head-wise Attention Response Energy-regulated framework for precipitation nowcasting. HARECast explicitly models head-wise attention-response energy and stabilizes it through a group-wise regularization objective that reduces cross-sample fluctuations. The proposed formulation is generic and applicable to both unimodal and multimodal nowcasting architectures. We instantiate HARECast in a standard forecasting pipeline with reconstruction branches and a diffusion-based predictor, and evaluate it on commonly used benchmarks--SEVIR and MeteoNet. Experimental results demonstrate that HARECast achieves state-of-the-art performance.

Abstract

Explore similar work

Stable Attention Response for Reliable Precipitation Nowcasting

Diffusion-Based Refinement for Kilometer-Scale Probabilistic Precipitation Nowcasting

SwAIther-Precip: Lead-Time-Aware Bias Correction Enables Kilometer-Scale Downscaling of Global AI Precipitation Forecasts over Switzerland