Optimal design of water distribution systems including water quality and system uncertainty

Abstract

Genetic Algorithm (GA) optimization is a proven technique for determining the best design for Water Distribution Systems (WDSs). Recent advances in WDS analysis have included water quality issues and uncertainty in system parameters. Water quality modeling requires an extended period hydraulic simulation to be executed, generally requiring a shorter time step and is therefore more computational intensity than steady-state modeling. Similarly, dealing with uncertainty is also more computationally intensive, when using a Monte Carlo (MC)-based method to calculate risk measures, such as reliability and vulnerability. The incorporation of both water quality and uncertainty into GA-based WDS optimization therefore results in impractically long run times. This paper uses metamodels to significantly reduce these run times. A metamodel is an approximation of an existing model, which takes less time to run, making it much more computationally efficient upon repeated use, such as in MC-based analyses or GA-based optimizations. The specific type of metamodel used in this research is an Artificial Neural Network (ANN), as it is capable of approximating any function without specifying, beforehand, its form. Previous research has illustrated the ability of ANN metamodels to approximate pressure heads, chlorine residuals, and both hydraulic and water quality risk measures. This research extends that work by linking those metamodels with a GA to enable risk-based optimization of WDSs by including system uncertainty. The approach developed was applied to an adapted New York Tunnels problem that incorporates water quality in the form of chlorine dosing for disinfection, and uncertainty in the demand. A series of single-objective optimization runs showed that there is a trade-off between minimizing cost, maximizing reliability and minimizing vulnerability. Using ANN metamodels for the risk-based optimization of WDSs that includes water quality issues had run-times of 30 hours, whereas the traditional approach (using EPANET rather than ANN metamodels) is estimated to take in the order of 10,000 hours. Copyright ASCE 2006.