Criteria for Evaluating Distribution Network Early Warning Systems

The ability to detect and act upon changes in water quality is a critical component in the drive to protect our drinking water supplies from intentional or accidental contamination. The twin motivators of the terrorist threat to water along with consumer demands for safe and potable supplies has lead to a sea change in the drinking water industry. The distribution system represents the last analytical frontier in the water quality industry that has been, to date, overlooked and ignored since the inception of modern drinking water systems. The monitoring of source water and treatment plant processes has progressed to a level at which we can be confident that we are providing good quality water from the plant to the distribution system. Once the water reaches our aging distribution systems, however; our knowledge as to its continued integrity is limited by the quality and amount of available data. From a historical perspective, most monitoring in the distribution system has been relegated to the occasional snapshot provided by grab sampling for a few limited parameters or the infrequent regulatory testing required by mandates such as the Total Coliform Rule. A number of studies conducted since 9/11 have shown that bulk monitoring of basic water quality parameters has the potential to indicate the presence of many harmful agents in water at the levels of interest. (EPA, 2005; Kroll, 2006; Hall et. al., 2007). The realization of the potential of bulk parameter monitoring as a practical tool to detect terrorism related events has lead to the development of a number of sensor packages designed for deployment in the distribution system. Since 9/11, numerous communities have installed multi-parameter monitoring stations in various locations through out their distribution network as early warning systems to detect potential water security threats as well as providing operational data. These continuous on-line systems have recorded large streams of data (at some sites for a number of years) relevant to water quality in the distribution systems in which they have been deployed. These data streams are quite complex and it becomes a Herculean task to differentiate what is normal background noise and fluctuations due to normal everyday events from changes that are indicative of a deviation in water quality deserving further attention. Unless a full-time team of statisticians is to be employed to make sense of this information, the need for computer aided data interpretation becomes obvious. Intelligent algorithms are a necessary adjunct to bulk parameter monitoring if useful decision trees are to be built from this type of monitoring. Intelligent algorithms are necessary to streamline the process of data interpretation. These algorithms should be capable of detecting the subtle changes in bulk parameter readings that are indicative of an incursion into the system without burdening the operators with a constant stream of false or trivial alarms. They should also be capable of differentiating the unique pattern of responses that are elicited by different classes of agent. These differences may be enough to narrow down the cause of events and, possibly, to fingerprint the class of disturbance that caused the event. Over the past several years a number of such algorithms designed for this use have been under development by private industry, government programs, universities and national labs. The question then becomes how to evaluate the effectiveness of these potential early warning system (EWS) solutions. While a number of research studies (McKenna et al, 2006; Umberg, 2008) and programs such as the EPA's ETV program (ETV2005) have attempted to set criteria and means for evaluating such systems, many important functional criteria for such systems have been overlooked in past studies. A number of factors that are mandatory for the success of any such system are outlined below.