This study developed a simulation/optimization model based on artificial
neural networks and genetic algorithm techniques for solving conjunctive
use of groundwater problems. The model simultaneously addresses all
significant flows in a dynamic hydraulically connected stream-aquifer
system. It also addresses multiple objectives. The genetic algorithm is
a search procedure based on the mechanics of natural selection and
genetics. It has been applied almost exclusively to single objective
optimization problems. Water resources projects are generally constructed
to serve multiple objectives. The nondominated sorting genetic algorithm
was adopted for multiobjective optimization. The artificial neural network
was used to represent simulation constraints inside the optimization model.
It predicted groundwater system responses to changes in decision variables.
The linked ANNGA model was able to solve complicated nonlinear reservoir-
stream-aquifer system equations. A fuzzy-penalty function method was used
to implicitly handle constraints on state variables. Three conflicting
objectives were considered: (1) maximization of the sum of unsteady
groundwater pumping and surface water diversions for three consecutive
60-day periods, (2) minimization of operating costs, and (3) maximization of
hydropower production. The three objective functions were constrained by
21 state variables. Water use from one multipurpose water reservoir, two
pumping wells, two stream diversions, and one reservoir diversion was
optimized. Three scenarios were investigated. The first scenario optimized
the first two objectives simultaneously. The second scenario maximized total
water diverted and minimized pumping costs concurrently. The third scenario
was a combination of the first two and optimized the three objective functions.
The genetic algorithm control-parameters were investigated. A crossover
probability of 0.6 and a niche size corresponding to 200 peaks were most
appropriate for all tested scenarios. Mahfoud's population sizing method was
investigated and validated. Population sizes for the three scenarios were
obtained accordingly. The 21 artificial neural networks estimated state
variables within a linear correlation range of 0.85 and 0.99. The final
product of all three scenarios was a set of trade-off curves that allow the
manager to assess the relative importance of each objective. The new
simulation/optimization model was robustly able to simulate water flows and
optimize different water management problems without violating any of the
specified constraints.