Performance of the Global Forecast System (GFS) in the Northern South America Region and Its Impact on the Overall Skill of an Operational Regional Weather Forecast Strategy Using WRF

Global numerical weather prediction models are fundamental for the generation of high-resolution limited-area weather forecasts for multiple applications (i.e., risk management, water resource planning), providing the necessary atmospheric boundary conditions. The NCEP Global Forecast System (GFS) is one of the most well-known global weather models, and it is often used for operational purposes in regional atmospheric model systems, being freely-available in near-real-time, with forecasts every 6 hours. As part of the regional forecast assessment process, it is important to quantify how much the performance of the model depends on the configuration of the model itself (parameterizations used, vertical levels, resolution, number of domains, among others), how much on the initialization strategies, and how much on the accuracy of the boundary conditions. In this work, we evaluate the quality of the GFS forecasts based on the NCEP Final Operational Global Analysis (FNL) product for a tropical region, in northern South America. Particularly, the region of interest includes the boundaries of the synoptic domain (10.2°S-19.9°N, and 90.3-59.7°W) from a three-nested-domain configuration of an operational forecast application using the WRF/ARW model, developed for the use of the Medellín (Colombia) Early Alert System (SIATA, www.siata.gov.co). The northern boundary includes the Caribbean Sea, the eastern boundary Amazon Basin, and the Caribbean Sea, the southern boundary a portion of Amazon basin, the Andes Cordillera over Peru, and the East Pacific Ocean; and finally, the western boundary is over the East Pacific Ocean. The operational forecast is carried out one time per day, using the 0.5° GFS 12 UTC as a boundary and initial conditions, with a forecast horizon of 120 hours. We evaluated the spatio-temporal performance of the model in the entire forecast horizon, hourly, using RMSE (Root Mean Square Error), MAE (Mean Absolute Error), BIAS (mean relative difference) and Pearson correlation of the WRF input variables: temperature, relative humidity, U and V wind components, specific humidity, potential temperature, and moisture advection. The verification period is from May 2013 to July 2019. During that time NCEP released five GFS updates (August 20, 2013; January 14, 2015; May 11, 2016; July 19, 2017; and June 12, 2019), and we identify dramatic changes in the performance of GFS starting from some of those dates, especially in January 2015 update. Results suggest that relative humidity is the variable that has been improving the most in the time. The results also show, as expected, that the errors increase sharply with the forecast lead time, with the most prominent errors in all variables (except potential temperature and Specific humidity) located in the upper troposphere between 400 to 100 hPa, and over the continental areas, especially close to the steepest topography of the Andes. Also, there exist considerable atmospheric heating biases associated with the diurnal cycle, particularly marked in the levels close to the surface, with more heating/cooling during the day/night in GFS than FNL. The skill of the WRF precipitation forecasts strongly depends on the quality of the GFS forecasts, more so in most cases than on the selection of the WRF parameterizations.