MODELING DRINKING WATER QUALITY VIOLATIONS WITH BAYESIAN NETWORKS

摘要

Compliance violations at community water systems are rare but represent significant human health risks. These risks are mediated by the decision schema of human operators at water treatment facilities. However, causal uncertainty among physical and human factors involved in water quality problems complicates assessment of their probability and severity. This study uses a probabilistic Bayesian network modeling approach to explore the causes of compliance violations in a sample of water treatment systems in Pennsylvania. The model presented here is one of several created by treatment system operators during an expert elicitation process. The expert model alone predicts violations poorly, suggesting that experts make inaccurate quantitative estimates. However, Bayesian networks are capable of combining the subjective expertise of treatment system operators with the objective compliance histories of the facilities they manage, and the expert model accurately predicts violations when trained with historical compliance data. Analysis of the trained network reveals those components of the treatment process, including environmental and system characteristics as well as operator decisions, that play the greatest role in determining the likelihood of major violation types. Among operator decisions, coagulant dosing and filter backwash frequency are the most important determinants of violation likelihood.