Stagnation or low turnover of water within water distribution systems may result in water quality issues, even for relatively short durations of stagnation / low turnover if other factors such as deteriorated aging pipe infrastructure are present. As leakage management strategies, including the creation of smaller pressure management zones, are implemented increasingly more dead ends are being created within networks and hence potentially there is an increasing risk to water quality due to stagnation / low turnover. This paper presents results of applying data driven tools to the large corporate databases maintained by UK water companies. These databases include multiple information sources such as asset data, regulatory water quality sampling, customer complaints etc. A range of techniques exist for exploring the interrelationships between various types of variables, with a number of studies successfully using Artificial Neural Networks (ANNs) to probe complex data sets. Self Organising Maps (SOMs), are a class of unsupervised ANN that perform dimensionality reduction of the feature space to yield topologically ordered maps, have been used successfully for similar problems to that posed here. Notably for this application, SOM are trained without classes attached in an unsupervised fashion. Training combines competitive learning (learning the position of a data cloud) and co-operative learning (self-organising of neighbourhoods). Specifically, in this application SOMs performed multidimensional data analysis of a case study area (covering a town for an eight year period). The visual output of the SOM analysis provides a rapid and intuitive means of examining covariance between variables and exploring hypotheses for increased understanding. For example, water age (time from system entry, from hydraulic modelling) in combination with high pipe specific residence time and old cast iron pipe were found to be strong explanatory variables. This derived understanding could ultimately be captured in a tool providing risk based prioritisation scores.
Mounce, Stephen Robert; Sharpe, Rebecca; Speight, Vanessa; Holden, Barrie; and Boxall, Joby, "Self-Organizing Maps For Knowledge Discovery From Corporate Databases To Develop Risk Based Prioritization For Stagnation" (2014). CUNY Academic Works.