Data‐driven model for river flood forecasting based on a Bayesian network approach

Abstract

Abstract Uncertainty analysis of hydrological models often requires a large number of model runs, which can be time consuming and computationally intensive. In order to reduce the number of runs required for uncertainty prediction, Bayesian networks (BNs) are used to graphically represent conditional probability dependence between the set of variables characterizing a flood event. Bayesian networks (BNs) are relevant due to their capacity to handle uncertainty, combine statistical data and expertise and introduce evidences in real‐time flood forecasting. In the present study, a runoff–runoff model is considered. The discharge at a gauging station located is estimated at the outlet of a basin catchment based on discharge measurements at the gauging stations upstream. The BN model shows good performances in estimating the discharges at the basin outlet. Another application of the BN model is to be used as a reverse method. Knowing discharges values at the outlet of the basin, we can propagate back these values through the model to estimate discharges at upstream stations. This turns out to be a practical method to fill the missing data in streamflow records which are critical to the sustainable management of water and the development of hydrological models.