Physically Based and Data-Driven Models and Propagation of Input Uncertainties in River Flood Prediction

Abstract

The understanding of the model capabilities and inherent uncertainties is vital in river flood prediction systems. This paper addresses the need by considering two conventional models: hydrodynamic (HD) model and Muskingum-Cunge (MC) hydrologic routing model, and two data-driven models: artificial neural network and adaptive network based fuzzy inference system. A major source of uncertainty in all of these models is in input discharge due to the stage-discharge relationship. The study considers the uncertainty by defining fuzzy uncertainty bounds of relationship, which is used for propagation of uncertainties in each of these models. This approach is applied to the Rhine-Neckar river confluence in Germany. The results of the study indicate that all four models are capable of producing good results. While the statistical performance of the MC routing model and two data-driven models are slightly better than the HD model, the HD model is more robust in handling uncertainties. The study therefore suggests that it is important to consider both the performance and uncertainties in these models and it is more appropriate to use more than one model for river flood prediction.