Simulation study of the effects of model uncertainty in variational assimilation of radar data on rainfall forecasting

Recent developments in data assimilation techniques make use of cloud models initialized with radar data an attractive alternative for real-time quantitative precipitation forecasting (QPF). Before such approaches are used operationally. there are a number of aspects that need to be addressed to understand better the benefits and drawbacks of the practical applications of cloud models in QPF. One aspect is the effect of various sources of uncertainty on the forecasting performance. Data assimilation formulations based on variational techniques allow accounting for uncertainty in observations but have no efficient mechanism of accounting for uncertainty in the model on which they are based. To investigate the issue. a simulation-based Monte Carlo methodology, suitable for the analysis of complex nonlinear models, is used. A one-dimensional stochastic-dynamic cloud model, derived by considering stochastic terms in a physically based cloud model is used to simulate rainfall and radar reflectivity data. A deterministic version of the model is then initialized by a variational assimilation technique and used for forecasting. The differences between the forecasts and actual realizations of the stochastic cloud model are statistically analyzed to assess the effect of cloud model uncertainty on forecasting. In this paper this methodology is also used to study additional effects of other types of uncertainty, such as those in radar observations and in the description of rain drop size distribution, for a more complete understanding of the impact of uncertainties on rainfall forecasting. Based on the scenarios investigated in the paper, conclusions and recommendations concerning the use of complex cloud models in real-world applications are made. (C) 2000 Elsevier Science B.V. All rights reserved.