Calibrated, Probabilistic Hazard Forecasts From A Time-lagged Ensemble

Abstract

Time-lagged ensembles are a computationally inexpensive substitute for full ensembles, using the “free” uncertainty information provided by different initializations of one or more rapidly cycled, deterministic model(s). The High-Resolution Rapid Refresh (HRRR) Time-Lagged Ensemble (HRRR-TLE) is part of a multi-agency collaborative project to improve probabilistic forecasts of hazardous weather. The HRRR-TLE combines forecasts from the 3 most recent initializations of the experimental HRRR model as ensemble “members”, to predict the chance of heavy rainfall, winter precipitation, severe thunderstorms and aviation-related hazards over the contiguous US, at lead times of up to 24 h. Many time-lagged ensembles are underdispersive, meaning that an envelope of forecasts initialized at successive hours tends to underestimate the range of possible future states. To address this issue in the HRRR-TLE, we employ statistical post-processing methods, such as spatial filtering schemes, to augment the ensemble spread and achieve more reliable probabilistic forecasts. Prior to the application of a spatial filter, the quantitative precipitation forecast from each member of the HRRR-TLE is bias corrected using a threshold frequency technique and a relatively short training dataset, which further improves the reliability of probabilistic heavy rainfall forecasts. In future, similar bias-correction techniques will be applied to other fields using trusted analyses and observation datasets. This presentation will include an overview of the reasoning behind the current HRRR-TLE configuration, a description of hazard identification algorithms, a demonstration of the improvement offered by statistical post-processing, and finally several case-study examples to show typical forecast output.