Data Driven Statistical Model for Manganese Concentration Prediction in Drinking Water Reservoirs

Abstract

Continuously monitoring and managing manganese (Mn) concentrations in drinking water supply reservoirs are paramount for water suppliers, as high concentrations create discoloration of potable water supplied to the customers. Traditional Mn management approaches typically involve manual sampling and laboratory testing of raw water from supply reservoirs on a regular basis (typically weekly) and then treatment decisions are made based on the soluble Mn level exceeding an allowable threshold level; for the reservoir in this study the threshold level for treatment is 0.02 mg/L. Often Mn testing is conducted all year, but in the sub-tropical regions, such as the Gold Coast, Australia, where the reservoir of interest for this study (Hinze Dam) is located, high Mn concentrations only occur for a brief period during the dam destratification process which occurs at the beginning of winter. High concentrations of Mn, resulting from the destratification event, in water entering the water treatment plant are usually treated through pre filter chlorination for concentrations < 0.18 mg/L, or with addition of potassium permanganate for higher concentrations. Recently, a vertical profiling system (VPS) has enabled the data collection of many water parameters, such as water temperature, dissolved oxygen, pH, conductivity and redox potential every 3 hours. Despite the abundance of physical and water quality data collected by the VPS, it cannot directly measure a range of water quality parameters such as Mn, thus manual sampling and testing are still required. Since previous studies have shown significant links between the physicochemical parameters collected by VPS and Mn concentrations, a data driven model can be developed to predict Mn values accurately. A Multiple Linear Regression (MLR) with empirical equations for Hinze dam was trained using data from 2008 to 2011, and tested with an independent dataset from 2012. The model was able to predict one week ahead the average Mn concentration in the epilimnion, where the water is drawn, with a correlation coefficient higher than 0.83. The output is also displayed in form of probabilities of exceeding certain thresholds, for instance 0.02 mg/L (namely Mn treatment needed). Successfully achieving the development of an autonomous and accurate tool for the data mining of VPS parameter datasets to predict levels of Mn provides several benefits for treatment operators: such a decision support system (DSS) would significantly reduce laboratory costs while concurrently enhancing treatment adaption response times.