Table 2 reveals that in the simplest models the iron concentration and pipe material are key factors. Models with reduced parsimony, such as those with greater than seven terms could provide CoD greater than 0.8; however, due to the data set size for this case study overfitting is inevitable. By incorporating Flush2 as an additional input, the best single variable equation was still the first listed equation in Table 2, but the second equation then incorporated Flush2, along with FeConc with a CoD of 0.65. The multi-case strategy (MCS) variant of EPR utilises splitting data into subsets according to, for example, failure history, and it has been used to develop distinct models for different subsets of pipes (Giustolisi & Berardi 2009). An additional EPR study was conducted on cast iron material only for the case study. Table 3 provides two of these models (the simplest). Note that it was possible to get CoD greater than 0.9 for only five terms in this case, however caution over the subset data size needs to be emphasised since the data set was reduced to 44% of the original size. However, CoD for the equation with only the iron concentration term is improved over that in Table 2. Bulk water iron concentration could potentially be a single measure capturing the dominant influence of a number of other water quality factors: source water, coagulation treatment processes and quantity of unlined upstream iron.
Selected Pareto optimal regeneration rate estimation models (cast iron material) identified by EPR
| Model structure | CoD |
|---|---|
| 0.46 | |
| 0.79 |
| Model structure | CoD |
|---|---|
| 0.46 | |
| 0.79 |