Intelligent data mining of vertical profiler readings to predict manganese concentrations in water reservoirs

Continuously monitoring and managing manganese (Mn) concentrations in drinking water supply reservoirs are paramount for water suppliers since high soluble Mn loads lead to discoloration of potable water. Despite the Mn level currently being manually sampled throughout the year, in subtropical monomictic lakes such as Hinze Dam, critical Mn concentrations in the epilimnion, where the water is drawn, are typically recorded only during winter lake circulation. A vertical profiling system (VPS) installed can continuously collect physical parameters that determine the transport process of Mn in the lake. Therefore, a long-term historical database gives opportunities for the development of a Mn prediction model. In the present study, VPS and sampling data were collected and analysed, and prediction models applying nonlinear regression techniques and data-driven equations were developed and assessed. They were able to accurately forecast future Mn concentrations from 1 to 7 days ahead and in particular the critical peak concentrations in the epilimnion during the lake destratification. The model also displays the probabilities of the Mn to exceed certain key-thresholds, thus assisting operators in Mn treatment decision-making. Such a tool is very beneficial for the water supplier, since costly and time-consuming water samplings for monitoring Mn concentrations can be avoided, thus relying only on the real time VPS-based model outputs.