Environ. Sci. Technol. 2008, 42, 5654–5660
Development and Interpretation of Disinfection Byproduct Formation Models Using the Information Collection Rule Database A L E X A O B O L E N S K Y * ,† A N D PHILIP C. SINGER‡ Philadelphia Water Department Bureau of Laboratory Services, 1500 East Hunting Park Ave., Philadelphia, Pennsylvania 19124, and Department of Environmental Sciences and Engineering, CB no. 7431, University of North Carolina, Chapel Hill, North Carolina 27599-7431
Received December 20, 2007. Revised manuscript received March 21, 2008. Accepted April 1, 2008.
Multiple linear regression models were used to examine relationships between water quality, treatment, and disinfection byproduct (DBP) formation in Information Collection Rule field data. Finished water models were specified using a crossvalidation approach based on data for 225 free chlorine treatment plants. Turbidity, bromide, temperature, alkalinity, total organic carbon, ultraviolet absorbance at 254 nm, pH, chlorine residual, chlorine consumed, and chlorine contact time were employed as independent variables. Important trends within the trihalomethane, dihaloacetic acid, and trihaloacetic acid classes were observed. Bromide was a significant predictor for all DBP species and its influence changed in sign and magnitude with the extent of bromine substitution. A similar pattern followed by alkalinity suggested it plays an important role as an indicator of natural organic matter hydrophobicity and reactivity. Chlorine consumed and organic precursor variables were significant predictors in almost all DBP species models, exhibiting trends opposite to those for alkalinity and bromide. Temperature was the most significant variable in chloroform and chloral hydrate models and its significance declined with increasing bromine substitution within the trihalomethane class. pH had a strong positive influence on chloroform formation, a negative influence on trihaloacetic acid formation, and no influence on dihaloacetic acid formation.
Introduction Multiple linear regression models (regression models) were used to study relationships between water quality, treatment processes, and DBP formation employing full-scale water treatment plant data from the Information Collection Rule (ICR) database (1). Regression modeling provides an efficient statistical approach to isolating and quantifying interrelated effects on DBP formation from simultaneously varying factors in a single data set and lends itself to straightforward interpretation. A long history exists of regression model development based on data from controlled laboratory * Corresponding author phone: (215)-685-1450; fax (215) 743-5594; e-mail:
[email protected]. † Philadelphia Water Department Bureau of Laboratory Services. ‡ University of North Carolina 5654
9
ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 42, NO. 15, 2008
experiments (2). The most well-known of these models (3, 4) became the basis for the U.S. Environmental Protection Agency’s Water Treatment Plant Simulation Program (EPA model) that was used extensively to support regulatory development through forecasting DBP production (5, 6). Little work has been done developing such models from observational field data although, at least conceptually, it provides an excellent analytical tool for studying DBP formation (7–9). Limitations include the need for a large data set that includes key water quality and treatment variables with consistent information structure. Availability of the ICR database, containing DBP data and associated water quality and treatment information from a comprehensive survey of large U.S water utilities, provided a unique opportunity to apply regression modeling techniques to the analysis of real water treatment plant data. Summaries of ICR data used in this research and associated data handling methodologies are described elsewhere (10). McGuire et al. provide additional information about the ICR program and uses of ICR data (11). Models developed for the present research are similar in mathematical structure to the EPA model. However, the current focus on using regression models as an investigative tool to probe DBP formation differs fundamentally from previous modeling efforts, which aimed at predicting DBP concentrations. Although forecasting per se was not a goal of the current work, models needed to perform adequately in a predictive capacity so that large error terms would not preclude detecting and comparing effects of interest. The size and structure of the ICR database was considered ample for supporting the research goals. Acknowledging practical limitations to describing DBP formation at a mechanistic level, the models presented here were developed empirically, based on a rational framework. This framework encompasses current knowledge of basic drivers for DBP formation in drinking water and includes the variables characterizing water quality and treatment processes that are easily obtained and commonly monitored. Accordingly, the models can be used to gain a better understanding of how DBPs are related to water quality and process parameters in real, dynamic treatment systems.
Scope of Research Finished water DBP concentrations were modeled for ICR treatment plants using only free chlorine for disinfection prior to distribution (herein termed “chlorine plants”). Separate models were developed for twelve individual DBP compounds (chloroform, CHCl3; bromodichloromethane, CHBrCl2; dibromochloromethane, CHBr2Cl; bromoform, CHBr3, dichloraoacetic acid, Cl2AA; bromochloroacetic acid, BrClAA; dibromoacetic acid, Br2AA; trichloroacetic acid, Cl3AA; bromodichloroacetic acid, BrCl2AA; dibromochloroacetic acid, Br2ClAA; tribromoacetic acid, Br3AA; and trichloroacetaldehyde hydrate, Cl3AH), total organic halogen (TOX), THM4 (sum of four trihalomethanes), X2AA (sum of three dihaloacetic acids), X3AA (sum of four trihaloacetic acids), and HAA9 (sum of nine haloacetic acids). HAA9 models implicitly account for monochloroacetic acid and monobromoacetic acid occurrence but no attempt was made to model these two species due to their consistently low concentrations in the database and the relatively large uncertainties associated with their analytical results (12, 13).
Materials and Methods Statistical Basis for Models. ICR data were collected by large U.S. utilities in 1997-1998. The encompassed domain of 10.1021/es702974f CCC: $40.75
2008 American Chemical Society
Published on Web 06/20/2008
TABLE 1. Dependent Variables: Summary of Data variable
DBP class
Nc
mind
max
meand
median
90th percentile
CVe (%)
MRLf
