NN-based implicit stochastic optimization of multi-reservoir systems management

Multi-reservoir systems management is complex because of the uncertainty on future events and the variety of purposes, usually conflicting, of the involved actors. An efficient management of these systems can help improving resource allocation, preventing political crisis and reducing the conflicts between the stakeholders. Bellman stochastic dynamic programming (SDP) is the most famous among the many proposed approaches to solve this optimal control problem. Unfortunately, SDP is affected by the curse of dimensionality: computational effort increases exponentially with the complexity of the considered system (i.e., number of reservoirs), and the problem rapidly becomes intractable. This paper proposes an implicit stochastic optimization approach for the solution of the reservoir management problem. The core idea is using extremely flexible functions, such as artificial neural networks (NN), for designing release rules which approximate the optimal policies obtained by an open-loop approach. These trained NNs can then be used to take decisions in real time. The approach thus requires a sufficiently long series of historical or synthetic inflows, and the definition of a compromise solution to be approximated. This work analyzes with particular emphasis the importance of the information which represents the input of the control laws, investigating the effects of different degrees of completeness. The methodology is applied to the Nile River basin considering the main management objectives (minimization of the irrigation water deficit and maximization of the hydropower production), but can be easily adopted also in other cases.