Open AccessA Nonparametric Ensemble Postprocessing Approach for Short-Range Visibility Predictions in Data-Sparse Areas
Author(s) -
William R. Ryerson,
Joshua P. Hacker
Publication year - 2018
Publication title -
weather and forecasting
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.393
H-Index - 106
eISSN - 1520-0434
pISSN - 0882-8156
DOI - 10.1175/waf-d-17-0066.1
Subject(s) - weather research and forecasting model , visibility , meteorology , ensemble forecasting , mesoscale meteorology , environmental science , range (aeronautics) , computer science , nonparametric statistics , stability (learning theory) , ensemble learning , statistics , mathematics , machine learning , geography , materials science , composite material
This work develops and tests the viability of obtaining skillful short-range (<20 h) visibility predictions using statistical postprocessing of a 4-km, 10-member Weather Research and Forecasting (WRF) ensemble configured to closely match the U.S. Air Force Mesoscale Ensemble Forecast System. The raw WRF predictions produce excessive forecasts of zero cloud water, which is simultaneously predicted by all ensemble members in 62% of observed fog cases, leading to zero ensemble dispersion and no skill in these cases. Adding dispersion to the clear cases by making upward adjustments to cloud water predictions from individual members not predicting fog on their own provides the best chance to increase the resolution and reliability of the ensemble. The technique leverages traits of a joint parameter space in the predictions and is generally most effective when the space is defined with a moisture parameter and a low-level stability parameter. Cross-validation shows that the method adds significant overnight skill to predictions in valley and coastal regions compared to the raw WRF forecasts, with modest skill increases after sunrise. Postprocessing does not improve the highly skillful raw WRF predictions at the mountain test sites. Since the framework addresses only systematic WRF deficiencies and identifies parameter pairs with a clear, non-site-specific physical mechanism of predictive power, it has geographical transferability with less need for recalibration or observational record compared to other statistical postprocessing approaches.