Post-processing using deep learning to create operational, high-resolution, and probabilistic weather forecasts
In one sentence. Operational multi-task neural network post-processes HRRR forecasts with a CRPS loss, improving RMSE by 3–25% and CRPS by 31–45% over HRRR.
Abstract
We present a machine learning model that post-processes high-resolution, deterministic forecasts to produce short to medium-range probabilistic forecasts for seven core weather variables.We developed and operationally implemented a multi-task neural network with a custom loss function, the Continuous Ranked Probability Score (CRPS), using input data from the High Resolution Rapid Refresh (HRRR) model.This combined high-resolution Numerical Weather Prediction (NWP) modeling with machine learning to generate more accurate deterministic and probabilistic forecasts.The results show deterministic forecast improvements in Root Mean Squared Error (RMSE) from 3% to 25% and probabilistic improvements in CRPS from 31% to 45% over the HRRR model.The multi-task neural network is cheap to train and fast to run operationally on hourly forecasts.SIGNIFICANCE STATEMENT: Weather forecasting is a challenging problem.Operational forecasts use numerical weather prediction (NWP) models that solve physical equations.These weather forecasts are statistically post-processed to correct errors.Traditionally, uncertainties in the forecasts are captured by creating an ensemble of deterministic predictions with variations in the initial conditions and the NWP model.In this study, we developed a post-processing system to correct weather forecasting errors relative to observations and to produce probabilistic ensembles.Our system uses machine learning to generate more accurate deterministic and probabilistic weather forecasts.The post-processing system is cheap, fast, and deployed operationally.
Labels
- Topic
- Post-processing & calibration, Probabilistic & ensemble forecasting, Regional & km-scale models
- Type
- New model or method
- Method
- Other
- Time scale
- 6 h–3 days, 3–15 days
- Domain
- Regional
- Data
- NWP analysis or forecasts (IFS, GFS, ...)
- Status
- Operational
- Identifiers
- DOI:10.22541/essoar.169841550.08343114/v1 OpenAlex:W4387971817
- Added
- 2026-10-22
Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.