Six-hourly Probabilistic Precipitation Forecasts Created From Hourly Probabilities
Abstract
A lightweight neural network is described that produces six-hourly, spatially detailed probabilistic quantitative precipitation forecasts (PQPF) from a sequence of hourly spatially detailed probabilistic forecasts, for lead times from 6 to 72\,h. The hourly forecasts were generated by a ``zero-inflated'' Attention Residual U-Net that outputs the parameters of a probability distribution including a probability of no precipitation and a weighted mixture of two Gamma distributions at each grid point for each hourly forecast lead time. The six-hourly network passes the six consecutive hourly parameter sets, together with summary statistics of the spatial texture of the raw deterministic forecast, through a small recurrent (GRU) layer and a multi-layer perceptron, and maps them to a corresponding zero-inflated mixture of three Gamma distributions describing the six-hourly accumulated precipitation. Training uses sets of hourly probabilistic forecasts, drawn preferentially from wet and heavy-precipitation cases (importance sampling), and matched Multi-Radar Multi-Sensor (MRMS) six-hourly precipitation accumulations, subject to a quality threshold. The resulting six-hourly probabilistic forecasts are well calibrated and skillful, as demonstrated by reliability diagrams, performance diagrams and Brier Skill Scores evaluated over an out-of-sample period spanning day 12 through the end of every month of 2025. The Brier Skill Scores are consistently and significantly higher than those of a baseline that sums independent draws from the hourly distributions. Forecasts show realistic modulation of precipitation related to terrain variations
Related
- High-resolution Calibrated Probabilistic Hourly Precipitation from a Deterministic Forecast
- Methodological Changes to the Attention ResUNet Hourly Precipitation Postprocessor
- Spatio-temporal precipitation downscaling at hourly resolution based on km-scale regional simulations
- Contrasting drivers of summer extreme hourly precipitation over the southeastern and northeastern Tibetan Plateau
- Climatology and trends of observed daily and hourly extreme precipitation in the French Alps
BibTeX
@misc{hamill2026sixhourly,
title = {{Six-hourly Probabilistic Precipitation Forecasts Created From Hourly Probabilities}},
author = {Thomas M Hamill},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.03326},
eprint = {2611.03326},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.03326},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Six-hourly Probabilistic Precipitation Forecasts Created From Hourly Probabilities
AU - Thomas M Hamill
PY - 2026
DA - 2026/11/02
PB - arXiv
DO - 10.48550/arxiv.2611.03326
UR - https://arxiv.org/abs/2611.03326
AB - A lightweight neural network is described that produces six-hourly, spatially detailed probabilistic quantitative precipitation forecasts (PQPF) from a sequence of hourly spatially detailed probabilistic forecasts, for lead times from 6 to 72\,h. The hourly forecasts were generated by a ``zero-inflated'' Attention Residual U-Net that outputs the parameters of a probability distribution including a probability of no precipitation and a weighted mixture of two Gamma distributions at each grid point for each hourly forecast lead time. The six-hourly network passes the six consecutive hourly parameter sets, together with summary statistics of the spatial texture of the raw deterministic forecast, through a small recurrent (GRU) layer and a multi-layer perceptron, and maps them to a corresponding zero-inflated mixture of three Gamma distributions describing the six-hourly accumulated precipitation. Training uses sets of hourly probabilistic forecasts, drawn preferentially from wet and heavy-precipitation cases (importance sampling), and matched Multi-Radar Multi-Sensor (MRMS) six-hourly precipitation accumulations, subject to a quality threshold. The resulting six-hourly probabilistic forecasts are well calibrated and skillful, as demonstrated by reliability diagrams, performance diagrams and Brier Skill Scores evaluated over an out-of-sample period spanning day 12 through the end of every month of 2025. The Brier Skill Scores are consistently and significantly higher than those of a baseline that sums independent draws from the hourly distributions. Forecasts show realistic modulation of precipitation related to terrain variations
ER -