ANNs and inflow forecast to aid stochastic optimization of reservoir operation

Abstract

Implicit stochastic reservoir optimization (ISO) typically utilizes nonlinear regression to correlate release as a function of initial storage plus inflow forecasted for the month. This study shows that improved ISO-based policies can be derived by replacing current-month forecast and regression for long-term mean inflow forecast (LTF) and artificial neural networks (ANN), respectively. The ISO-LTF-ANN approach is applied to the Aswan High Dam reservoir, Egypt. First, perfect-forecast deterministic optimization (PFDO) defines operation strategies for 100 scenarios of 100-year inflows. Then, optimal release and storage data are grouped into databases corresponding to different forecast horizons. Next, ANNs are trained for each database to serve as release policies. Later, the policies are used to operate the system under other scenarios. Operations by the standard operation policy (SOP), stochastic dynamic programming (SDP) and PFDO are employed for comparison. ISO-LTFANN performs near PFDO and better than SOP, SDP and a regression-based ISO-LTF approach.