Neuro-dynamic Programming for the Efficient Integrated Water Resources Management

Design of policies within the integrated water resources management (IWRM) paradigm demands to take into account multiple objectives such as flood protection, water supply for hydropower generation, irrigation and urban use, imposition of minimum environmental flows, etc.. Finding efficient water management policies is a demanding task and Stochastic Dynamic Programming (SDP) can provide an effective solution methodology. The efficiency boost is even more pronounced when the scale of integration of the water management problem spans networks of water resources (i.e. sets of interconnected reservoirs, or wide-area water distribution systems). Unfortunately, SDP suffers from the so-called “curse of dimensionality”, that is, computational times and memory storage requirements increase exponentially with the dimension of the problem (number of reservoirs and water diversions), and rapidly the problem becomes intractable. Neuro-dynamic programming (NDP) can sensibly mitigate both of these problems by approximating the Bellman functions with Artificial Neural Networks (ANNs). In this paper we present the theoretical framework of this approach and some preliminary results obtained on a real world case study, where the NDP approach to IWRM problems has been applied to a multipurpose water reservoir network. Copyright © 2004 IFAC