Particulate material accumulates over time as cohesive layers on internal pipeline surfaces in water distribution systems (WDS). When mobilised, this material can cause discolouration. This paper explores factors expected to be involved in this accumulation process. Two complementary machine learning methodologies are applied to significant amounts of real world field data from both a qualitative and a quantitative perspective. First, Kohonen self-organising maps were used for integrative and interpretative multivariate data mining of potential factors affecting accumulation. Second, evolutionary polynomial regression (EPR), a hybrid data-driven technique, was applied that combines genetic algorithms with numerical regression for developing easily interpretable mathematical model expressions. EPR was used to explore producing novel simple expressions to highlight important accumulation factors. Three case studies are presented: UK national and two Dutch local studies. The results highlight bulk water iron concentration, pipe material and looped network areas as key descriptive parameters for the UK study. At the local level, a significantly increased third data set allowed K-fold cross validation. The mean cross validation coefficient of determination was 0.945 for training data and 0.930 for testing data for an equation utilising amount of material mobilised and soil temperature for estimating daily regeneration rate. The approach shows promise for developing transferable expressions usable for pro-active WDS management.
Multivariate data mining for estimating the rate of discolouration material accumulation in drinking water distribution systems
