On the Rainfall and Temperature Forecast Skill for a Tropical Andean Mountain Area in Northern South America Using Different Operational Weather Forecast Strategies: Role of the Diurnal Cycle of Rainfall on the Success of Data Assimilation

Numerical Weather Prediction models (NWP) have been used extensively since the '40-'50s. Despite the advances in the field, the representation and forecast of the magnitude and variability of tropical processes in models is still a challenge. One of the steps to improve the precipitation forecasts using limited-area models is to evaluate which set of physical schemes and model domain configurations represent in a better way the actual behavior observed in the tropics. We implemented, as a part of a regional risk management strategy, two different operational weather forecast strategies for a complex terrain region in the Andes mountain range in northern South America. Both strategies, together, generate a total of eleven different forecasts every day, using the Weather Research and Forecasting model (WRF) with initial and boundary conditions from the Global Forecast System (GFS). The first configuration, implemented over five years ago and referred to as SYNAPSIS, includes three nested domains (18, 6 and 2 km) and is carried out every day using the 12 UTC GFS run and three different microphysics parametrizations: Eta Ferrier scheme, Purdue Lin Scheme and Thompson Scheme. The forecast lead-time of the latter strategy is 120 hours, and it does not use data assimilation. Since December of 2017, we implemented a second configuration termed RDFS, with two nested domains (12 and 2.4 Km), which carried out four times a day using the 00, 06, 12 and 18 UTC GFS runs. This configuration has a 30-hours lead time with the Thompson microphysics scheme. In RDFS, two WRF forecast runs are performed for each start hour, one assimilating weather radar reflectivity and the other without assimilation as control run, for a total of eight forecast runs daily. In this study, we assess the rainfall and temperature forecasts for all the different configurations using precipitation derived from reflectivity from weather radar, and air temperature at 2m from a network of automatic weather stations. We use 6 hourly and monthly skill scores (RMSE, BIAS, and Correlation coefficient) to quantify the precipitation differences between the SYNAPSIS and the RDFS configurations.

To evaluate the impact of data assimilation in the precipitation forecast, we aggregate the results in a region within the inner domain, and then we calculate the average precipitation forecast between 0 and 36 predicted hours for RDFS with and without data assimilation. The results suggest a strong relationship between the forecast start time and the improve of precipitation forecast accuracy using data assimilation. The diurnal cycle of precipitation in the study region has a minimum in the morning (12 UTC) and a maximum in the afternoon (00 UTC) and during the night (09 UTC). The correspondence between the forecast improvement using data assimilation and the diurnal cycle of precipitation is likely due to the amount of assimilated data. In order to quantify the precipitation differences between the different microphysics parametrizations used in the SYNAPSIS strategy and the observed precipitation, we calculated the model skill scores for the average forecast of precipitation in a region inside the inner domain. Results show that the use of the latter three parametrizations is likely redundant because the results did not show a significant difference in forecasts skill among the different microphysics schemes. According to the results, we conclude that a configuration with the domain distribution of RDFS, with radar reflectivity data assimilation at the hour of maximum precipitation, produces the most accurate and skillfully forecast of precipitation. For surface air temperature, we observe systematic errors as a result of the height differences between the real and the model topography. We use a Model Output Statistics (MOS) technique to improve the forecast results, consisting of a quantile-quantile linear regression, between the observations and the temperature forecasts. This correction works well for mean temperatures, but not for maximum and minimum temperature. In general, the surface temperature forecasts are warmer at noon and colder before sunrise than the observations from an AWS network. We also compare the observed and simulated turbulent heat fluxes inside the urban areas; the difference has a strong correlation with the observed cold or hot bias in the temperature forecast.