Empirical mathematical models and artificial neural networks for the determination of alum doses for treatment of southern Australian surface waters

The potential for predicting alum doses far surface waters from southern Au stralia based on physico-chemical parameters of the raw waters was studied. These parameters included dissolved organic carbon (DOC), absorbance at 25 4 nm, turbidity and alkalinity. Procedures used for assessing the predictab ility of alum dosing were empirical mathematical models and artificial neur al networks. Alum doses determined by jar tests were selected on the basis of target val ues for settled and filtered turbidities, colour and residual aluminium. Regression equations which incorporated the parameters of DOG, UV absorbanc e (254 nm/cm), turbidity, alkalinity and pH gave correlation coefficients o f greater than 0.9. These equations gave a high frequency of prediction wit hin +/- 10 mg/L alum of actual doses. Similarly, 86% of alum doses predicte d by artificial neural networks were within 10 mg/L of the actual doses. Al though a good prediction of coagulant dosing was achieved, it is likely tha t the models generated are specific for the types of waters studied and the criteria for alum dose selection.