| Probability network models (PNM) have emerged in recent years as powerful and efficient tools to assess and predict water quality. Accordingly, this research explores water quality prediction and treatment decision support system based on advancements in graphical modeling, including decomposable Markov networks (DMNs), Bayesian networks (BNs) and Bayesian decision networks (BDNs).;This development of DMN improves the procedures for estimating censored/missing data. It provides an improved alternative to traditional techniques for infilling of censored/missing. The uncertainties in estimates for censored/missing water quality data is demonstrated as being reduced by establishing dependencies among water quality attributes.;A BN model is constructed in this research to simulate different coagulation conditions for enhanced coagulation, a procedure focused on reducing precursors of DBP formation. The causal dependence relationships amongst the variables (e.g. coagulant, pH, temperature, etc.) are encoded in a BN structure, to identify the most advantageous configuration for DBP reduction. A BDN procedure is demonstrated which provides a normative framework for selecting between water treatment processes.;A comparison of probability network modeling with traditional methods such as regression analysis is presented. Regression analysis assumes a Gaussian distribution around a predicted value of a target variable. Probabilistic modeling instead provides the posterior distribution of the target variable without a prior assumption of the probability distribution. In water quality prediction, many constituents have skewed and multimodal distributions; for such cases, probabilistic modeling can closely estimate the posterior probability of the target variable and predict its most probable value, while the regression may associate the target variable with an incorrect probability distribution and, as a consequence, incorrectly predict the value.;It is shown that probabilistic modeling provides more robust predictions of censored or missing water quality parameters. PNM has unique advantages and can perform bidirectional inference, and hence PNM is shown to be a powerful tool for water quality prediction and for dealing with real-time control. Procedures to apply of these general research results to water treatment and quality monitoring practice, and integrate these procedures into operators' operations are described in the dissertation. |
