ABSTRACT
Throughout the year 2007, 89 treated water samples from three water treatment plants (WTPs) of the Athens Water Supply and Sewerage Company (EYDAP S.A.) and 180 samples from network tanks (NWTs) were analyzed for electrical conductivity (EC), alkalinity (TA), pH, aluminium (Al), total hardness (TH), chloride (Cl(-)), residual chlorine (free Cl), calcium (Ca(2+)) and magnesium (Mg(2+)). The results regarding the WTPs were subjected to a principal component analysis (PCA) with 75% of the total variance being explained. A stepwise linear discriminant analysis (LDA) model constructed from the 89 treated water samples was used to predict class membership of the samples from the NWTs with a view to estimating the propagation of a possible water quality deterioration originating from the WTPs. The model utilized Cl(-), Al and EC and yielded a 96% correct classification of the training dataset, whereas the cross-validation yielded a 94% correct classification. Network tank samples were 95% correctly classified with regard to their theoretically expected origin. The stepwise discriminant analysis based on separate covariance matrices of the canonical discriminant functions yielded a 98% correct classification of both the training dataset and the network tank samples. The classification and regression tree (C&RT) algorithm showed that the main parameters used in the discrimination of the WTP samples were EC and Al. The post-hoc classification of the training dataset was 99%, whereas 88% of NWT samples were correctly classified.
