Probabilistic Models for Decision Support During Floods
Luis Garrote, Dep. Ingeniería Civil: Hidráulica y Energética , Universidad Politecnica de Madrid
ABSTRACT. In this seminar, an approach to make probabilistic forecasts using physically based deterministic models is presented. The methodology is based on the use of Bayesian networks to learn from synthetic results generated with the physically-based model.
Bayesian networks are a kind of data-driven model where the joint probability distribution of a set of related variables is inferred from observations. The qualitative simulation of hydrologic processes is performed through causal relations quantified with conditional probabilities. The solution algorithm of Bayesian networks allows the computation of the expected probability distribution of output variables conditioned to the probability distribution of input variables. Bayesian networks are computationally more flexible and more efficient than complex physically-based models for real-time use. However, their application to flood forecasting is limited because basins with long data sets for calibration or validation of this type of models are relatively scarce.
To solve the problem, the data set for the calibration and validation of the Bayesian model is obtained through a Monte-Carlo simulation technique, combining a stochastic rainfall generator and a deterministic rainfall-runoff model. The methodology allows making probabilistic discharge forecasts in real time using an uncertain quantitative precipitation forecast. This framework is adequate for basins where the use of deterministic models in Monte Carlo simulations is computationally unfeasible in real time, because the computational burden of Monte Carlo simulations is transferred off-line. The approach has been tested to make predictions in the Spanish Mediterranean region, where flooding problems are caused by intense storms moving over many small basins with short response times. The validation experiments made show that the data-driven model can approximate the probability distribution of future discharges that would be obtained with the physically-based model applying ensemble prediction techniques, but in a much shorter computation time. The computational structure of the Bayesian network also allows for an efficient user interface for real-time decision support.