Assessment of a Probabilistic Scheme for Flood Prediction

This study presents the development of a probabilistic discharge prediction scheme based on an uncertainty framework called generalized likelihood uncertainty estimation (GLUE). By being explicit about a hydrologic model’s parameter uncertainty, historical data is used adaptively on a storm-to-storm basis to derive ensembles of representative parameter sets, along with the corresponding likelihood weights of discharge prediction quantiles. The quantile with highest likelihood weight represents the most probable discharge hydrograph, with upper/lower uncertainty limits represented by the various upper/lower likelihood weight quantiles. On the basis of new data, the Bayesian theorem is used to update for the posterior representative parameter sets and likelihood weights of prediction quantiles. The probabilistic scheme is evaluated using 15 flood-inducing storms over a medium-sized watershed in northern Italy. The scheme’s discharge predictions on the basis of its highest likelihood quantile are evaluated comparatively to the conventional single optimum parameter set prediction. It is observed that the two methods have comparable accuracy in terms of the overall hydrograph prediction, but the probabilistic scheme is subject to 50% less variability in time to peak error. The probabilistic scheme has an added value important to decision making and risk assessment, which is its ability to provide consistent assessment of uncertainty in such major flood parameters as peak runoff and time-to-peak. The procedure is simple in design, model-independent, and can be easily implemented in real-time for computationally efficient rainfall-runoff models.