Using data mining for event-based prediction of turbidity in a drinking water source

Drinking water treatment plants (DWTP) using surface water as potable source could be particularly vulnerable to short term transient events leading to sudden rise in suspended sediments, organic matter, nutrients, pathogens and other organic and inorganic contaminants. The prediction of source water parameters using early warning systems could be one solution to drinking water operators to manage short term transient water quality contamination events. In the context of climate changes where an intensification of rainfall-runoff events and consecutive pollution episodes is predicted, using data mining techniques could be of particular interest as forecasting tools to adapt efficiently drinking water treatment during transient pollution episodes. This study focuses on the development of data mining techniques using neural networks and trend analysis to forecast turbidity peaks in a drinking water source located in a humid continental climate (Quebec, Canada). The DWTP uses surface water to provide drinking water to almost 300 000 inhabitants. High frequency data from 2012 to 2016, from on-line measurements, are used for source water turbidity. Rainfall indicators (number of dry days, sum of the daily precipitation for 1, 2, 5 and 10 days prior to turbidity event start, days since daily precipitation of at least 10, 20 and 50 mm prior to event start) have been created using five meteorological stations located within the watershed as input parameters for models. The results of this study could help water treatment plant operators to anticipate the variability of key water quality parameters.