Probabilistic Model for Real-Time Flood Operation of a Dam Based on a Deterministic Optimization Model

Abstract

This paper presents a real-time flood control model for dams with gate-controlled spillways that brings together the advantages of an optimization model based on mixed integer linear programming (MILP) and a case-based learning scheme using Bayesian Networks (BNets). A BNet model was designed to reproduce the causal relationship between inflows, outflows and reservoir storage. The model was trained with synthetic events generated with the use of the MILP model. The BNet model produces a probabilistic description of recommended dam outflows over a time horizon of 1 to 5 h for the Talave reservoir in Spain. The results of implementing the BNet recommendation were compared against the results obtained while applying two conventional models: the MILP model, which assumes full knowledge of the inflow hydrograph, and the Volumetric Evaluation Method (VEM), a method widely used in Spain that works in real-time, but without any knowledge of future inflows. In order to compare the results of the three methods, the global risk index (Ir) was computed for each method, based on the simulated behavior for an ensemble of hydrograph inflows. The Ir values associated to the 2 h-forecast BNet model are lower than those obtained for VEM, which suggests improvement over standard practice. In conclusion, the BNet arises as a suitable and efficient model to support dam operators for the decision making process during flood events.