Optimization of numerical models for operational weather forecasting in Uganda

Abstract

The major focus of this study was on improving the predictability of precipitation, in the form of rainfall, which is one of the major weather and climatic elements of interest in the tropics, using numerical weather prediction (NWP) models. Globally, increasing frequencies and severity of the extreme weather events and associated negative impacts on the lives and infrastructure have been reported. However, the non-linear atmospheric processes, sparse observation network and the uncertainty in initial conditions compound to make quantitative rainfall prediction a challenge. NWP models are one of the tools to objectively guide quantitative rainfall prediction and have been used in this study with the following objectives: (i) to compare the performance of the Consortium for Small-scale modeling (COSMO) model and the Weather Research and Forecasting (WRF) model; (ii) to assess the performance of the convective parameterization schemes; and (iii) to examine the performance of ensemble methods.

This study first compared the rainfall prediction skill by the NWP models namely the COSMO model and the WRF model using the rainfall data observed at 21 Ugandan weather stations for the period 21st April to 10th May 2013 using the root mean square error (RMSE), mean error (ME) and scores obtained from the contingency table namely the probability of detection and the false alarm ratio. This period was chosen because it experienced heavy rainfall over different areas in Uganda. The models were initialized using lateral boundary conditions from the National Centers for Environmental Prediction Final Analysis and the Germany weather Service for running the WRF model and the COSMO model respectively.

This study found the WRF model performing comparatively better than the COSMO model) over Uganda in simulating extreme rainfall events at 5% significance level. For the light rainfall events, the COSMO model presented a comparatively smaller magnitude of error than the WRF model but in spite of this COSMO performance, the study considered extreme rainfall events as being of greater concern and therefore decided to adopt the WRF model in further analyses.

The study then assessed the performance of six convective parameterization schemes under the WRF model in simulating rainfall namely: the Kain-Fritsch (KF) scheme; the Betts–Miller–Janjić (BMJ) scheme; the Grell–Freitas (GF) scheme; the Grell 3D ensemble (G3) scheme; the New–Tiedtke (NT) scheme and the Grell–Devenyi (GD) scheme in simulating rainfall over Uganda for an extended period, the rainy period from 1 st March to 31 st May 2013. It further examined the performance of the ensemble rainfall prediction methods (ensemble mean, analogue ensemble mean and multi-member analogue ensemble mean) and compared their performance to the performance of the convective parameterization schemes using the Student’s t-test.

The root mean square error results of the WRF convective parameterization schemes (KF = 23.96; BMJ = 26.04; GF = 25.85; G3 = 24.07; NT = 29.13 and GD = 26.27) and the mean error results (KF = -0.82; BMJ = -2.05; GF = -1.98; G3 = -1.84; NT = -2.75 and GD= -2.07) showed that all the schemes generally presented a negative bias due to under-prediction of the extreme rainfall events. The comparison of the performance of ensemble methods and that of the three best convective parameterization schemes (KF, GF and G3 schemes) showed that the ensemble mean presented a significant improvement in quantitative rainfall prediction compared to the individual convective parameterization schemes (KF: t = 4.73, p<0.001; GF: t = 5.14, p < 0.001 and G3: t =5.41, p < 0.001). The ensemble mean analogue also presented better performance compared to the individual convective schemes (KF: t = 4.94, p < 0.001; GF: t = 5.38, p < 0.001 and G3: t = 5.66, p < 0.001) and the multi-member analogue ensemble mean presented the best performance compared to all the ensemble methods studied (KF: t = 5.00, p < 0.001; GF: t = 5.44, p < 0.001 and G3: t = 5.73, p < 0.001).

This study, therefore, recommends the WRF model for operational weather prediction over Uganda using the KF parameterization scheme for deterministic quantitative rainfall prediction. It also recommends the multi-member analogue ensemble mean for ensemble quantitative rainfall prediction. Due to the high temporal and spatial rainfall variability, the study further recommends the assimilation of the observational weather data to improve deterministic rainfall prediction especially the mesoscale convective systems. Since the models generally under-predicted the extreme rainfall events, the study further investigation of the sources of the great bias in simulating such events.