Optimizing water supply headworks operating rules under stochastic inputs: Assessment of genetic algorithm performance

Realistic optimization of the operation of urban water supply headworks systems requires that the issues of system complexity and stochasticity be addressed. A promising way of achieving this is to couple probabilistic search methods such as genetic algorithms (GAs) with Monte Carlo simulation models. However, the objective function surface, characteristic of this genre of problem, exhibits piecewise flat regions separated by steep slopes. The size of these flat regions is affected by the frequency of droughts sampled in the Monte Carlo simulation. Although an earlier study proposed a GA variant that appears to robustly negotiate such objective function surfaces, the assessment was limited to a simple one-reservoir system. There remains therefore a legitimate concern that for more complex systems the GA may converge prematurely producing decisions of little practical value. This study assesses the ability of the GA to optimize key operating rules for a complex urban headworks system with nine reservoirs and interbasin transfers subject to a highly variable climate. Eight decisions affecting restriction rules, pump marks, and contingent desalination were optimized using an objective function that minimized expected annual costs. It is shown that the GA produced results that are judged consistent with the strategy that minimizes total expected costs. In addition, the sensitivity of the GA results to the length of the Monte Carlo simulation was investigated