GA-based implicit stochastic optimization and RNN-based simulation for deriving multi-objective reservoir hedging rules.

Abstract

Management of reservoir systems is a complicated process involving many uncertainties regarding future events and the diversity of purposes these reservoirs serve; therefore, an effective management of these systems could help improve resource utilization and avoid stakeholder disputes. The aim of this paper was to build an optimization-simulation framework based on implicit stochastic optimization (ISO), genetic algorithms (GA), and recurrent neural network (RNN) for addressing the issue of reservoir operation. Inflow scenarios were generated synthetically based on a monthly scale to be used as an input to a multi-objective genetic programming model to construct an optimal operating rules database. Such database was subsequently used simultaneously with the output of the inflow forecasting model to simulate monthly reservoir hedging rules using RNN. Our results demonstrate the effectiveness of the GA-ISO-RNN model for simulating and predicting optimal reservoir release with consistent accuracy. Results from both the training and testing phases clearly proved the usefulness of RNN in predicting optimal reservoir release with relatively higher values of the Nash-Sutcliffe model efficiency coefficient, correlation coefficient, and lower values of root mean squared error and mean absolute deviation. Furthermore, by comparing the historical releases and the output of the proposed model, the results show that the proposed model was less vulnerable than standard operating rules. The proposed methodology was applied to the Bigge reservoir in Germany, as it features an extensive management infrastructure, but this methodology can also be easily adopted in other similar cases.