Towards adaptive reservoir management: insights into policy and input variable selection for the Grand Ethiopian Renaissance Dam (GERD)

Recent advances in multi-objective computation have focused on the direct parametrization of operation policies for multi-purpose reservoirs, where the decision variables - the releases - are a function of a set of inputs typically including storage and time of year (a proxy for expected inflows). In spite of this, general guidelines concerning what parametrization to choose depending on the characteristics of the problem, the policy objectives, and the available inputs, are yet to emerge. Such guidelines are needed for the real world implementation of these policies in situations where they would help balance conflicting objectives. This work proposes a case-dependent benchmarking framework that enables to test different policy formulations and the possibility of including additional input parameters that would make policy selection better adapted to the context. This framework would complement existing benchmarking frameworks regarding the choice of the multi-objective evolutionary algorithm and its high-performance cluster implementation, by looking into 1) which formulations achieve the highest performance, and 2) the computational costs required to obtain that level of performance. It is applied to the GERD, a major reservoir in construction on the Blue Nile. Two policy formulations are compared, piecewise linear operating rules vs. radial basis functions, and the inclusion of indicators of past releases is selected as an additional input in this situation where a drought can be felt for years on reservoir levels, and where regional stability depends on downstream riparian's water supply.