Evaluation of Three Numerical Weather Prediction Models for Short and Medium Range Agrohydrological Applications

The skill and accuracy of the quantitative precipitation forecasts by CCAM, UM and NCEP-MRF models are verified using various statistical scores at the Mgeni catchment in KwaZulu-Natal, South Africa. The CCAM model is capable of identifying a rainfall event, but with a tendency of under-estimating its magnitude. The UM model is capable of distinguishing rainy days from non-rainy days, but with a significant over-estimation of rainfall amount. There is no significant difference between the 1 and 2 day lead time UM forecasts. Statistical comparisons show that there is an acceptable skill in the CCAM forecasts, but the forecast skill of the UM model is low and unreliable. The role of the initial hydrological conditions in affecting the accuracy of CCAM and UM streamflows forecasts was significant. The results show that the under-estimation of the CCAM forecasts was reduced from −44% to −10%, while the over-estimation in the UM forecasts was reduced from 291% to only 59% when the ACRU agrohydrological model was initialised with observed rainfalls up to the previous day at each forecast run within the study period. The combined use of the CCAM and UM models by a “weighted averaging” had little effect in improving the skill as it is overshadowed more by the over-estimation of the UM forecasts than the under-estimation of the CCAM forecasts. Results obtained for a continuous period of 92 days showed that the NCEP-MRF rainfall forecasts were significantly over-predicted. The NCEP-MRF rainfall forecast is found to be totally unskillful, although the skill was seen to slightly increase with decreasing lead time.