Kejiang Zhang

Machine learning approaches to predict coagulant dosage in water treatment plants

Two machine learning methods, support vector machine and K-Nearest Neighbours (KNN) were investigated in this paper to predict the coagulant dosage in water treatment plants (WTPs). Two types of support vector machine regression techniques, ε-SVR and v-SVR, using two different kernel functions (radial basis function (RBF) and polynomial function), and KNN were investigated in order to predict coagulant dosage in a large, a medium, and two small-sized WTPs. The results show that these two types of support vector machine regression techniques have good predictive capabilities for the large and medium WTPs as compared to small water systems. The performances of ε-SVR with RBF kernel function were compared with that obtained from the KNN algorithm (as baseline) for four WTPs. The comparison shows that the KNN has similar performances as ε-SVR for the large and medium- sized WTPs and performs better for two small-sized WTPs. The results show that different machine learning methods have competing predictive abilities.

An integrated health risk assessment of contaminated sites under aleatory and epistemic uncertainties

An integrated health risk assessment framework including contaminant transport model, exposure assessment models, physiologicallybased pharmacokinetic (PBPK) modelling, and dose-response assessment was developed in this paper to bridge the gap amongst environmental engineers, toxicologists, and researchers in environmental health science and to improve the comprehensive understanding of the health risks posed by contaminated sites. Aleatory and epistemic uncertainties are identified, properly represented and coupled into the presented framework. The results show that: 1) the propagation of aleatory and epistemic uncertainties through the framework is transparent; 2) contaminant plume detection time (DT) or starting exposure time (SET) influence the final risk assessment; 3) for life time risk (e.g., exposure time equals 30 years), the steady state PBPK models are sufficient; 4) the total risk estimated using a point estimation method is higher by one to three orders of magnitude than that obtained from the integrated assessment framework.