Generalised Machine Learning Model for Prediction of Heavy Metals in Stormwater

The dynamics of the processes shaping the quality of rainwater discharged by sewer systems is very complex. The use of hydrodynamic models to simulate surface runoff and the dynamics of changes in pollutants, including heavy metal (HM) concentrations, requires the collection of a lot of data that is difficult to obtain, and model calibration is complex and time-consuming. This paper presents a machine learning model and investigates the possibility of applying data mining methods to simulate changes in the concentrations of selected heavy metals (Ni, Cu, Cr, Zn and Pb) based on rainwater quality studies conducted in three urban catchments located in Kielce, southern Poland, with the aim of developing a model with broader applicability. Simulations of HM content in rainwater were performed using regression and classification trees (RF), neural networks (MLP) and support vector machines (SVMs). The MLP (MAPE ≤ 21.6) and SVM (MAPE ≤ 23.5) methods were shown to have the highest accuracy in simulating HM content. These models produced satisfactory simulation results based on rainfall amount and meteorological conditions, and they had relatively simple model structures and short simulation time. The study demonstrated that the proposed approach provides a transferable tool for estimating HM content in rainwater based on air quality, expressed in terms of visibility, and the type of catchment development.