Precipitation Forecast Verification from Operational Weather Models

This paper verifies precipitation forecasts from operational numerical models against NCEP Stage IV gridded (~4 km) radar and rain gauge observations. The models that are examined are the Global Forecast System (GFS), North American Mesoscale Forecast System (NAM), Rapid Refresh (RAP), High Resolution NAM, High-Resolution Rapid Refresh (HRRR), and NCEP High-Resolution Windows (HIRESW) since August 2014, and the GFS Ensemble and Short Range Ensemble Forecast (SREF) systems since mid-January 2015. We utilize 6 h precipitation summation intervals. Verification includes the computation of fractions skill score (FSS) and bias score, or frequency bias, for various precipitation thresholds, model cycles (e.g., 0000 UTC and 1200 UTC initiations), forecast lengths, regions, and seasonal periods. FSS is a neighborhood verification method, particularly well-suited for verifying the new generation of high-resolution operational forecasts with convection-permitting horizontal grid spacings (≤ 5 km). These simulations provide tremendous detail, but often the model has poor skill in exactly placing the location of an individual storm, resulting in low equitable threat scores. Recognizing this inherent limitation in predictability, our use of a radius of influence of 60 km around each verification grid point allows the consideration of neighboring grid cells and preserves the valuable precipitation amount information provided by high-resolution models, resulting in a predicted fractional coverage exceeding some precipitation threshold. By removing detail, fields of fractional coverage also typically are easier for forecasters to utilize. Identifying the strengths and weaknesses of these various models may be of great value to weather forecasters and the operational modeling community for future improvements.