Neural Network Approximation of a Hydrodynamic Model in Optimizing Reservoir Operation

An approach of models approximation, applicable in the model-based optimization of water resources, is described. It was applied to the optimization of a system of the three reservoirs located in the Apure river basin in Venezuela. Its development plan requires to increase the period when the river is navigable along the certain reach. The problem was posed as a multi-criterial decision making (MCDM) problem for which a set of quasi-optimal policies may be found by solving a series of dynamic programming problems with the energy criterion and the navigability constraint. The hydrodynamics and hydrology of the basin was modelled using MIKE-11 modelling package. In order to run it in the optimization loop, the model of the basin was approximated using artificial neural network generator NNN. Such compact and fast representation of the hydrodynamic/hydrologic model, albeit approximate, easily allowed to include the modelling component into the optimization routines. The presented approach to model approximation may be used in various schemes of water resources optimization.