Contaminant intrusion into water distribution systems (WDS) is drawing increasing concern from the public. To deal with this concern, a contaminant warning system is an efficient way which relies on real-time water quality monitoring. Online sensors monitor surrogate parameters, and are able to identify the presence of contaminants within a WDS. Of considerable importance, however is a procedure to determine the number of sensors needed; this problem is particularly challenging when dealing with nodal demand uncertainty. Consideration of nodal uncertainty typically is involving days of simulation of intrusion events followed by days of optimization. This study fills this gap by means of optimization captured by the leverage of parallel computing.
Sensor placement is formulated as a multiple objective optimizations (MOO) problem, where the two objectives are time delay and sensor detection redundancy, and the decision variables are the sensor locations at nodes of the WDS. The first step involves simulation of contaminant behaviour of possible events under nodal demand uncertainty. This step can consume days for a small WDS (e.g. hundreds of nodes) and years for a medium size WDS (e.g. thousands of nodes) in serial computing in personal computing. Herein, a super-computer, SHARCNET is applied to reduce the computing time significantly. A list of events being alarmed by the designed sensor network is identified as those which can affect 15% of the entire WDS nodes.
Secondly, the MOO is solved by epsilon-NSGA-II algorithm. The two objectives computation is again very time-consuming. Herein, their computations are completed using a structural query language (SQL) sentences against the identified list of events. The functions evaluation in epsilon-NSGA-II is conducted with parallel execution of SQL sentences, which enables a number of processors to compute the two objectives values simultaneously, thus greatly reducing the computing time for optimization. An optimal sensor number for WDS is identified as the point of diminishing marginal return of Pareto front performance improvement by increasing sensor numbers.
The procedure is demonstrated through a case study in the City of Goderich WDS.