Forecasting of Ceiling and Visibility: Blending NWP, LAMP and an Observations-Based Statistical Model (Invited Presentation)

Current and emerging ceiling and visibility (C&V) forecast resources comprise an increasingly diverse and skilled population of forecast tools – including LAMP, SREF, RAP, HRRR, the North American Rapid Refresh Ensemble (NARRE), and others. While this diversity is highly desirable, it raises the question how can complementary strengths within this population of resources be harnessed and used?

This paper focuses on a method to build 1-10 hour probabilistic and deterministic C&V forecasts as a blend of input forecasts from three complementary sources (LAMP, RAP, and an observations-based statistical forecast model). Blending proportions unique to each site, forecast product, projection, initiation time and season are determined in advance through a training/trial process using a four-year data set. When the training/trial process indicates that use of a blend may not or will not yield forecast improvement, or when there are too few target events to enable reliable training, the resulting forecast defaults to LAMP.

Forecasts are made for airports within the contiguous U.S. and frequently (but not always) exceed the skill of the input forecasts. Since the data available for training and independent testing of the blending method is limited, training and verification rely upon a cross-validation technique through which each target forecast day and its two neighboring days are withheld from the corresponding training data set. Verification results characterize and compare the skill of the blended forecast, LAMP and RAP for a matching set of test days over a four-year period.

Although the current instance of this method utilizes the RAP, the method can rather simply be adapted and retrained for use of the HRRR in place of the RAP. Adaptation for HRRR use and evaluation of the resulting skill of the blended forecast method are planned as next steps.