Empirical Models for Chlorination By-products:

Four Years of Pilot Experience in Southern Connecticut

John N. McClellan1, David A. Reckhow1, John E. Tobiason1,

James K. Edzwald1, and Alan F. Hess2

1Department of Civil and Environmental Engineering, University of Massachusetts, Amherst, MA 01003

2South Central Connecticut Regional Water Authority, 99 Sargent Drive,

New Haven, CT 06511

Site-specific empirical models of trihalomethanes (THMs), haloacetic acids (HAAs), and the organic precursors of these compounds were developed based on pilot study data. Either chlorination by-product precursors or chlorination by-products themselves were treated as dependent variables, and either water quality parameters or process parameters comprised the independent variables. Ultraviolet absorbance was found to be the most effective conventional water quality parameter for predicting precursor removal in treatment processes. A first-order model was developed to predict the removal of THM and HAA precursors in waters treated with ozone and subsequent granular activated carbon filtration. In a separate analysis, haloacetic acid concentrations were predicted as a function of THM concentrations. Among other findings, these models have suggested a negative impact of filter aids on chlorination by-product precursor removal at a direct filtration plant.

Trihalomethanes (THMs) and haloacetic acids (HAAs) are by-products of the reactions between chlorine and natural organic matter in drinking water treatment. New rules will lower existing trihalomethane standards for drinking water and impose a standard on haloacetic acids for the first time (1). Measurements of the concentrations of THMs and HAAs and their precursors are necessary to control and assess treatment process performance, but direct measurements are time consuming and require sophisticated equipment. Empirical relationships which allow the estimation of THM and HAA concentrations as functions of other more easily measured parameters are therefore of interest.

Empirical models of chlorine disinfection by-products (DBPs) and their precursors can be used in several ways. First, models can be used for process control. Precursor concentrations during treatment and THM and HAA concentrations in distribution systems can be estimated using easily measured parameters. Information needed for process control can thus be obtained faster and with less expense using models than by direct measurements. Empirical models can also be used to evaluate the relative impact of individual parameters on process performance. Sensitivity analysis employing empirical models can be used for preliminary evaluation of process modification options. Finally, empirical models can provide a convenient way of summarizing large data sets and of comparing the treatability of different waters.

The South Central Connecticut Regional Water Authority (RWA) and the University of Massachusetts (UMass) conducted studies at pilot plants owned and operated by the RWA in which various process modifications aimed at improving the removal of the organic precursors of HAAs and THMs were investigated. A database that includes measurements of THMs, HAAs, and other water quality parameters was developed using results from the pilot studies. This database was used to calibrate and test empirical models for DBP precursors and for the formation of THMs and HAAs.

The objective of this paper is to present the results of these modeling efforts, and to provide a broad discussion of site-specific empirical modeling for control of DBPs.


Previous Chlorination By-product Modeling Efforts. In view of the need to control THM and HAA concentrations and the expense and time required to measure them directly, there has been considerable interest in developing predictive models for these substances. It was recognized early on that chlorine concentration, the concentration of organic precursor compounds, pH, temperature, and contact time all affect THM formation (2,3). Engerholm and Amy (4) proposed a model that predicts chloroform concentration as a function of total organic carbon (TOC), time, and a parameter defined as chlorine dose normalized by TOC concentration. Urano et al. (5) proposed a model which is similar to the Engerholm and Amy (4) model, but which incorporates the effects of pH and temperature. Morrow and Minear (6) explored the effect of bromide on THM formation, and calibrated a model which incorporated bromide concentration.

The three models described above use TOC alone as a surrogate for THM precursor concentration. This is a limitation because TOC-THMFP correlations tend to be site specific and are often poor. Ultraviolet (UV) absorbance at 254 nm is a good surrogate for THMFP (7). Amy et al. (8) found that a parameter defined as the product of TOC (mg/L) and UV absorbance (cm-1) correlated more closely with THMFP than did either parameter individually in experiments with a variety of natural waters. Amy et al. (8) presented a model for THM formation which employs UV absorbance, TOC, chlorine dose, pH, and bromide concentration as predictor variables. The database used for calibration contains over 1000 data points and was developed using nine natural waters in bench scale experiments. The Montgomery Watson (9) models, which are similar in form to the Amy et al. (8) THM models, provide separate equations for each of the THM and HAA species. These models were calibrated using a data set which includes the Amy et al. (8) database, two bench-scale databases developed by Montgomery Watson, and a bench-scale data set provided by the Metropolitan Water District of Southern California.

A difficulty with general empirical models is that they can be unreliable if the value of any independent variable is outside the range found in the calibration data. In addition, the dependent variable could be sensitive to parameters not included in the model. These factors may be less problematic in site-calibrated models. However, general models (e.g. the Montgomery Watson (9) models) have the advantage of large calibration databases. It is unclear which of these modeling approaches is superior in a given situation, or what level of accuracy can be expected from either approach.

The models mentioned above, which can be classified as formation models, predict DBP formation based on reaction conditions and reactant concentrations. Another class of models can be used to predict the concentration of organic DBP precursors. Examples of precursor models from the literature are the equations for THMFP as functions of UV absorbance and TOC given by Edzwald et al. (10), the equations presented by Harrington et al. (12) that predict TOC and UV absorbance removal as functions of alum dose and pH, and the model proposed by Huck et al. (11) for the removal of THMFP in biologically active granular activated carbon (GAC). The Huck et al. (11) model predicts the removal of precursors in a treatment process as a function of process parameters (in this case, empty bed contact time [EBCT]) and influent concentration. Although modeling of precursor removal versus process parameters could be very useful in the design and operation of water treatment facilities, a relatively small amount of work has been done in this area to date.

RWA/UMass Pilot Studies. Pilot studies were conducted at the RWA's West River Treatment Plant (WRTP) in 1992, at the Lake Saltonstall Treatment Plant (LSTP) in 1993 and 1994, and at the Lake Gaillard Treatment Plant (LGTP) in 1993, 1994 and 1995. The effects of various configurations of ozone and granular activated carbon (GAC) on organics removal were examined in these studies. Results from the 1992 WRTP and 1993 LGTP studies were published previously (13). In all these studies, measurements were made of THM and HAA precursor concentrations, UV absorbance, dissolved organic carbon (DOC), and turbidity in samples taken from raw waters, intermediate points in the pilot plants, and full- and pilot-scale treated waters. The data generated in these studies was organized and used for the development and testing of empirical models.

Experimental Methods

Pilot Plants. Lake Gaillard Water Treatment Plant (LGTP) is a direct filtration plant. The process includes coagulant addition (alum and cationic polymer), rapid mix, three-stage flocculation (30 minutes total contact time), and anthracite/sand (A/S) filtration at a hydraulic loading rate of 2.5-5 gpm/sq ft. Pilot Train #1 at LGTP consisted of a direct filtration process which simulated the full scale process, followed by a counter-current ozone contactor (10 minute contact time) and four granular activated carbon (GAC) contactors configured in series. The GAC contactors each contained 20 inches of Calgon Filtrasorb 300 (8 x 30 mesh) media over a gravel support layer and were typically operated at a hydraulic loading rate of 2.5 gpm/sq ft, which resulted in an EBCT of 5 minutes per contactor. Thus, samples could be collected at EBCTs of 5, 10, 15, and 20 minutes. This configuration is referred to as "post-ozone/GAC." Pilot Train #2 at LGTP consisted of a counter-current contactor (10 minute detention time for pre-ozonation of raw water); rapid mix, coagulant addition, and flocculation similar to the full scale plant; and GAC/sand (GAC/S) and A/S filters in parallel. The A/S filter contained 20 inches of 0.9 mm anthracite and 10 inches of 0.45 mm silica sand over a gravel support layer. Calgon Filtrasorb 300 (8 x 20 mesh) was substituted for anthracite in the GAC/S filter. On sampling dates, the mean alum and polymer doses applied in the pilot trains were 8.6 mg/L and 1.0 mg/L respectively, and the mean coagulation pH was 6.6.

The West River Treatment Plant (WRTP) is an in-line direct filtration plant. The treatment process includes pre-oxidation with potassium permanganate, coagulation with ferric chloride and a cationic polymer, and A/S filtration at a hydraulic loading rate of 1.5-3 gpm/sq ft. The WRTP pilot train consisted of a counter current ozone contactor (10 minute contact time), coagulation similar to the full scale plant, and GAC/S filtration at 3 gpm/sq ft. On sampling dates, the mean ferric chloride and polymer doses applied in the pilot train were 7.6 mg/L and 1.3 mg/L respectively, and the mean coagulation pH was 6.6.

The Lake Saltonstall Water Treatment Plant (LSTP) is a conventional treatment plant which includes pre-chlorination, coagulation with alum, rapid mixing, 2-stage flocculation, dual layer sedimentation, and A/S filters which have a design hydraulic loading of 3.5 gpm/sq ft. The LSTP pilot plant was similar to the full scale plant except that ozone was substituted for chlorine as a pre-oxidant, a plate settler was substituted for conventional sedimentation, and GAC media were substituted for anthracite in one of the two parallel filters. On sampling dates, the mean alum dose applied in the pilot train was 35 mg/L and the mean coagulation pH was 7.1. Schematics of the RWA pilot plants are presented in Figure 1. Mean values of raw water quality parameters are presented in Table I. Table II contains mean raw and treated water values for DOC, HAAFP, and THMFP for full scale and pilot scale RWA plants.

Analytical Methods. Formation potentials (FPs) were used to quantify the concentrations of THM and HAA precursors. Unchlorinated samples were transported to UMass and chlorinated within 24 hours of being collected. The standard chlorine dose for FP samples was 20 mg/L, followed by incubation in the dark at 20 {Special Char 176 in Font "Symbol"}C for 72 hours. Samples were incubated headspace-free in 300 mL BOD bottles. A phosphate buffer was added to maintain the pH at 7.0. Simulated distribution system (SDS) analysis was used to estimate distribution system THM and HAA concentrations. For SDS tests, a chlorine dose identical to the dose used in the full scale plants at the time of sampling (typically about 3 mg/L) was applied. Samples were then incubated headspace-free in the dark at 20 {Special Char 176 in Font "Symbol"}C for 48 hours at pH 7.0.

Table I. Raw Water Quality

Observed Mean Value at:
















Pt-Co Units









UV absorbance1





Temperature1{Special Char 176 in Font "Symbol"}






mg/L as CaCO3









1. Mean or range of values observed on sampling dates

2. Values for 1993 as reported by Tobiason (13)

3. Value not available

4. Accurate value not available. Value is assumed to be {Special Char 187 in Font "Symbol"} 30-50 {Special Char 109 in Font "Symbol"}g/L based on typical values for region.

The concentrations of HAAs and THMs were measured by gas chromatography with electron capture detection after quenching and appropriate sample preparation. Two HAA sample preparation methods were used: a modified version of Standard Method 6322 (14), and an alternate method developed by Xie et al. (15). Both methods involve micro-extractions and methylation. In Standard Method 6322, methylation is accomplished by addition of diazomethane. The method of Xie (15) avoids the use of diazomethane and employs acidic methanol instead. A micro-extraction method similar to the method of Koch et al. (16) was used for THM sample preparation. In order to minimize volatilization, extractions were performed immediately after quenching the THM samples. Replicate samples were analyzed for THMs and HAAs; the relative standard deviation was typically less than 5% for DCAA, TCAA, and CHCl3 measurements.

Table II. Mean DOC, HAAFP, and THMFP Levels in Raw and Treated Waters




({Special Char 109 in Font "Symbol"}g/L)


({Special Char 109 in Font "Symbol"}g/L)









Direct Filtration3















Direct Filtration







Pre-O3 with GAC/S Filt.5








Conventional with Pre-O35







1. HAAFP = sum of DCAAFP and TCAAFP

2. CHCl3FP {Special Char 187 in Font "Symbol"} THMFP in RWA waters due to low bromide concentration

3. Full scale data

4. Pilot scale data, 10 minute EBCT in GAC contactors

5. Pilot scale data

Dissolved organic carbon (DOC) and ultra-violet absorbance at 254 nm (UV) were also routinely measured. Samples were prepared for DOC and UV measurements by filtration through pre-washed 0.45 {Special Char 109 in Font "Symbol"}m glass fiber filters. DOC concentrations were measured using the UV-persulfate oxidation method after samples were acidified and sparged with nitrogen.

Modeling Chlorination By-products and Their Precursors

In low-bromide systems, three DBP species dominate: trichloracetic acid (TCAA), dichloroacetic acid (DCAA), and chloroform (CHCl3). In RWA waters, these three species typically account for over 90% of the HAAs and THMs which form as a result of chlorination. Although four THM species and five HAA species (six in 1995) were measured, only DCAA, TCAA, and CHCl3 are considered here. Chloroform can be considered approximately equivalent to total THMs and the sum of DCAA and TCAA can be considered approximately equivalent to total HAAs.

Precursor Models. Precursor models can be used to predict the concentrations of DBP precursors, or the extent to which precursors are removed in treatment processes. The dependent variable in these models is typically the ratio of effluent to influent precursor concentration. Initial concentrations, process parameters, other water quality parameters, or combinations of these may be used for predictor variables.

Fully Empirical Model for Precursor Removal Across Coagulation/Filtration. Data collected at the LGTP direct-filtration pilot train (no pre-ozone, A/S filtration) on thirteen sampling dates in 1993, 1994, and 1995 were used to calibrate models for predicting FP in filter effluents. The independent variables for these models are initial concentration, temperature, alum dose, cationic polymer coagulant dose, and anionic polymer filter-aid dose. For each model, the combination of variables which resulted in the highest adjusted correlation coefficient ( R2adj) (17) value was selected. The adjusted correlation coefficient compensates for small sample sizes relative to the number of predictor variables. Maximum values for R2adj resulted when two of the five predictor variables were used for the DCAA model, and when three of the five were used for the TCAA and CHCl3 models. Log-log transformations resulted in higher values of R2adj than did untransformed data for all three models. The resulting equations are:


N=13, R2adj = 0.61, P-Value = 0.007

CHCl3FP: (2)

N=11, R2adj = 0.70, P-Value = 0.022


N=10, R2adj = 0.83, P-Value = 0.002

where CE is A/S filter effluent FP in {Special Char 109 in Font "Symbol"}g/L, Co is raw water FP in {Special Char 109 in Font "Symbol"}g/L, A is alum dose in mg/L, F is filter aid dose in mg/L, and T is raw water temperature in degrees C. The P-Value (17) is the observed level of statistical significance. The mathematical form is a "power function," which is widely employed for empirical models. The power function form of equations (1), (2), and (3) resulted from taking anti-logs of both sides of the log-transformed regression equations.

In some cases parameters which have a significant effect on the modeled process do not improve the model fit because of a small range of values in the data. Predictor variables which did not improve the model fits were excluded from these models. For example, alum dose and pH clearly have an effect on the removal of DBPs, but alum dose is not included in equation (1) and pH is not included in any of the above equations because of small ranges of values in the data. The mean values of pH, alum dose, and other absent parameters not explicitly included are necessarily incorporated into the constant in each equation. The equations can only be considered valid for pH, temperature, chemical doses, and initial concentrations within the ranges of these parameters in the data, whether or not they appear as predictor variables. Mean values and ranges of operating parameters are given in Table III.

Table III. LGTP Direct Filtration Pilot Plant Operating Parameters

Values observed at

LGTP Pilot Plant-

Feb. 93 - July 95





Raw Water DCAAFP{Special Char 109 in Font "Symbol"}




DCAAFP Removal




Raw Water TCAAFP{Special Char 109 in Font "Symbol"}




TCAAFP Removal




Raw Water CHCl3FP{Special Char 109 in Font "Symbol"}




CHCl3FP Removal




Raw Water Temperature{Special Char 176 in Font "Symbol"}




Alum dose




Filter aid dose{Special Char 109 in Font "Symbol"}




Raw Water pH



Filter Effluent pH



Fully empirical models based on process parameters such as those given in equations (1), (2), and (3) may not be very useful for predicting process performance because of the limitations stated above. However, these models can be used for comparing the relative effects of processes parameters. Estimates of the change in the fraction of raw water FP removed per unit change of the indicated parameter at the LGTP direct filtration pilot train are presented in Table IV. These values were computed by evaluating partial derivatives of equations (1), (2) and (3) at the mean values of the parameters.

Table IV. Correlations Between Operating Parameters and DBP Precursor Removal

Estimated change in fraction raw water FP removed per unit increase of parameter






Raw Water TCAAFP{Special Char 109 in Font "Symbol"}





Raw Water CHCl3FP{Special Char 109 in Font "Symbol"}





Temperature{Special Char 176 in Font "Symbol"}





Alum dose





Filter aid dose{Special Char 109 in Font "Symbol"}





1. Parameter not included because it did not significantly improve model fit

An interesting result shown in Table IV is that higher doses of anionic polymer filter aid correlated with poorer removal of DCAAFP and CHCl3FP. These results suggest that anionic polymer addition may have the undesirable side effect of restabilizing DBP precursor material, although this conclusion cannot be made with certainty based on the results of this study. A controlled experiment is recommended to determine whether filter aid dose and the extent of DBP precursor removal are directly related or whether both are related to some other water quality or operational parameter.

Semi-Empirical Model for Precursor Removal in GAC. Huck and co-workers (11) reported a linear relationship between the extent of THMFP removal (and the removal of other organic carbon parameters) and influent concentration in biologically active GAC. To see if a similar relationship exists for THM and HAA precursors in the post-ozone GAC pilot train at the LGTP, the change in FP across each of the GAC contactors was plotted versus influent concentration. The pilot train had been in continuous operation for about 6 months at the LGTP at the time the of the earliest sampling for this data set. In addition, the GAC contactors were operated at the WRTP for about a year before being moved to the LGTP without changing the media. It is therefore assumed that the adsorptive capacity of the GAC was exhausted, and that the removal of DBP precursors in the GAC media was due to biological activity The plots, presented in Figure 2, show significant linear relationships, although correlations are not as good as the correlations observed by Huck et al. (11) for THMFP removal (r2=0.991 in a post-filtration deep bed GAC contactor, r2=0.748 in a GAC/sand filter with pre-ozone). The GAC contactors at LGTP are configured in series. This configuration is analogous to a single deep bed contactor, with "intermediate depth" sampling at the effluents of each contactor. A linear relationship between removal and influent concentration is apparent for all four contactors, suggesting that biological activity occurs throughout the depth of the media. As can

be seen in Figure 2, a linear extrapolation of the data shows a positive X intercept, which implies that there are influent concentrations below which no removal occurs. It is also assumed that only a fraction of the THM and HAA precursor material is biodegradable. A first-order model for utilization of THM and HAA precursors in the GAC media is proposed which incorporates the above features. The concentration profile with respect to depth in the proposed model is given by


C = Co when z = 0 (4b)

where C is concentration, z is bed depth penetrated, K is an apparent first order constant, Co is influent concentration, {Special Char 108 in Font "Symbol"} is the biodegradable fraction of Co, and Cmin is the influent concentration of biodegradable material below which no further biodegradation occurs. A graphic representation of the model parameters is presented in Figure 3. The integrated form is:


The parameters K and Cmin can be determined from plots of removal versus influent concentration, such as those shown in Figure 2, using the following relations:



where m is the slope of the regression line and b is the y intercept. Bed depth penetrated, z, is used in place of the more common but closely related parameter EBCT in the above equations. The reason is that changes in EBCT (due to flow changes) at the LGTP pilot plant seemed to have little effect on process performance at a given bed depth on the few occasions when the hydraulic loading rate was changed. Bed depth penetrated was therefore considered to be the more meaningful parameter. However, it should be noted that the hydraulic loading rate was 2.5 gpm/sq ft. for most of the calibration data and the model should only be considered valid for flows close to this value.

Values of K and Cmin were computed using the slopes and intercepts of the regression lines shown in Figure 2. Values of {Special Char 108 in Font "Symbol"} were selected to optimize the model fit. These {Special Char 108 in Font "Symbol"} values are in reasonable agreement with the values for biodegradability (approximately 65% for CHCl3FP and 80% for DCAAFP) given by Miltner et al. (18). Table V contains values for the model parameters K, Cmin and {Special Char 108 in Font "Symbol"}.

Table V. Constants for First Order Utilization Model

Precursor Type

K (ft-1)

Cmin ({Special Char 109 in Font "Symbol"}g/L){Special Char 108 in Font "Symbol"}













HAA Precursors vs. THM Precursors. Trihalomethanes have been regulated in U.S. drinking waters since 1979, and many U.S. utilities have already conducted studies aimed at adjusting their processes to meet existing THM regulations. A relatively

small amount of work to date has focused on HAAs. For this reason and because HAAs are more difficult to measure than THMs, modeling of HAA precursors using THMFP as a predictor parameter was explored. The simplest model of this type is the mean HAAFP:THMFP ratio. In Table VI, HAAFP:THMFP ratios for various raw and treated waters are presented.

These results illustrate the changes in the relative distribution of HAA and THM precursors during treatment processes. A decrease in the TCAAFP:CHCl3FP ratio can be observed between raw and treated waters for all cases. The preferential removal of TCAA precursors compared to CHCl3 and DCAA precursors was observed in bench scale studies (19), and is thought to result from the association of TCAA precursors with humic acids. The humic acids are characterized by large molecular size and a high level of hydrophobicity, and are thus easily removed by coagulation processes (19). The speciation is also different in the different raw waters. The general implication is that empirical models tend to be site specific, and models calibrated using raw waters may not make accurate predictions in treated waters.

It is interesting to note that the lowest treated water TCAAFP:CHCl3FP ratio is seen in the data from the conventional LSTP pilot plant, where a relatively high alum dose, without a cationic polymer, was employed (typically about 13 mg alum/mg DOC). Cationic polymer in addition to a relatively low primary coagulant dose (typically about 2.9 mg alum/mg DOC at the LGTP and about 2.5 mg ferric chloride/mg DOC at the WRTP) were added at the direct filtration plants.

Table VI. Mean HAAFP:THMFP Ratios in Raw and Treated Waters












Direct Filtration4







Dir. Filt./Post-O3/GAC5








Direct Filtration6







Pre-O3 with GAC/S Filt.7








Conventional with Pre-O38







1. None of the raw and treated DCAAFP:CHCl3FP ratios are significantly different at {Special Char 97 in Font "Symbol"} = 0.05.

2. All of the raw and treated TCAAFP:CHCl3FP ratios are significantly different at {Special Char 97 in Font "Symbol"} = 0.05.

3. HAAFP = DCAAFP plus TCAAFP; CHCl3FP {Special Char 187 in Font "Symbol"} total THMFP. The raw and treated HAAFP:CHCl3FP ratios are significantly different at {Special Char 97 in Font "Symbol"} = 0.05 except for the WRTP Direct Filtration and Pre-O3 ratios.

4. Full scale data, N=25

5. Pilot scale data, 10 minute EBCT in GAC contactors, N=18

6. Full scale data, N=12

7. Pilot scale data, N=12

8. Pilot scale data, N=7

Chlorination By-product Precursors vs. Conventional Water Quality Parameters. Empirical models of chlorination byproduct precursors based on other water quality parameters provide a convenient way of comparing, optimizing, and controlling treatment processes. Turbidity has traditionally been used in drinking water treatment as a surrogate parameter for particle concentrations, but other parameters may be preferable when optimizing and monitoring processes for HAA and THM precursor removal. A comparison was made between UV absorbance, dissolved organic carbon (DOC), and turbidity as predictor parameters for precursor removal models. Fractions of raw water CHCl3FP, DCAAFP, and TCAAFP remaining were plotted against fractions of raw water turbidity, DOC, and UV absorbance remaining in 60 filter and GAC contactor effluent samples from LGTP. The plots are presented in Figure 4 and Figure 5. These results indicate that UV absorbance is superior to DOC or turbidity as a predictor of THM and HAA precursor removal. Regressions (forced through zero) of these data yield the following equations:

CHCl3FP: (8)


TCAAFP: (10)

where subscript O and E denote treatment process influent and effluent concentrations respectively. Although most of the data used in developing equations (8), (9), and (10) are from a conventional direct filtration process with no ozone applied, some data from the pre- and post-ozone pilot plants at the LGTP are included. In the pre-ozone pilot train, the ozone dose was typically 2 mg/L. In the post-ozone pilot train, a dose of 0.5-0.7 mg/L was applied after filtration but upstream of the GAC contactors. As can be seen in Figure 4, the inclusion of data from the pre- and post-ozone processes did not seem to adversely effect the correlation between the removal of UV absorbance and DBP precursor removal. However, it should be noted that these processes included filtration or adsorption after the application of ozone. The correlation between UV absorbance removal and DBP precursor removal over the ozone contactor alone is be expected to be poor.

Chlorination By-product Formation Models. These models predict concentrations of the DBP species as opposed to precursor concentrations. The important factors that influence the formation of these compounds are chlorine concentration, precursor concentration, contact time, pH, and temperature. Bromide has an important effect on speciation if present in significant concentration. Models that predict THM and HAA formation must therefore take all these factors into account. Two approaches can be envisioned: either all of the above factors can be incorporated into the model, or a predictor parameter which reacts in a similar way to the factors can be employed. As discussed above, models have been published which utilize the first approach, using TOC and UV absorbance as predictors of precursor concentration. An example of the second approach is the use of one DBP species as a predictor of the others. Both of these approaches are discussed below.

Formation potential could also be used in multi-parameter models to represent initial precursor concentration. Models which employ an initial FP measurement, chlorine dose, time, pH, and bromide concentration as predictor variables might ultimately be the most useful for accurately predicting distribution system DBP concentrations. The data necessary to develop this type of model was not collected as part of the RWA/UMass pilot studies.

HAAs vs. THMs. Models which predict HAA concentrations as functions of chloroform concentration were calibrated and tested using SDS data from RWA/UMass pilot studies. All samples used for calibration and testing were treated water samples. The calibration data set consisted of 13 samples collected at LGTP in 1994. To test the models, independent data sets (i.e. data not used for calibration) were employed which included measurements of the parameters being modeled and the predictor parameters. Model predictions were compared to measured concentrations, and the predictions were considered successful if the differences fell within a specified range. Absolute ranges were used instead of relative ranges because the average variance of observed values from the model predictions remained fairly constant over the range of values in the test data. Results using two "success ranges" are reported. The equations were tested on 1995 LGTP data (N=12), and on data collected at LSTP in 1993 and 1994 (N=17). The chloroform model equations and prediction success rates are presented in Table VII.

The HAA vs THM models predicted DCAA and TCAA values from the LGTP (the calibration data was also from LGTP) reasonably well using a success range defined as {Special Char 177 in Font "Symbol"} 7 {Special Char 109 in Font "Symbol"}g/L. Using data from the conventional LSTP, the DCAA predictions were reasonably good, but the HAA vs THM model systematically overpredicted TCAA concentrations. This result illustrates that the extent of TCAAFP removal compared to the removal of CHCl3FP or DCAAFP can be quite different for different coagulation conditions, as discussed above. The TCAA vs. THM model is therefore particularly sensitive to site-specific conditions.

Table VII. Site Calibrated HAA vs THM Models1

Percent of predictions within:

{Special Char 177 in Font "Symbol"}

7 {Special Char 109 in Font "Symbol"}g/L of

measured value at{Special Char 177 in Font "Symbol"}

3 {Special Char 109 in Font "Symbol"}g/L of

measured value at














1. These models were calibrated and tested using SDS data. Means and ranges of test data values ({Special Char 109 in Font "Symbol"}g/L): CHCl3: 29 (14-59), DCAA: 19 (8-37); TCAA: 22 (11-40).

2. Units: {Special Char 109 in Font "Symbol"}g/L

Multi-parameter THM and HAA Formation Models. The Montgomery Watson models (9) for CHCl3, DCAA, and TCAA formation were tested against treated water SDS data from the LGTP and the LSTP. The models were tested on the same data set used to test the HAA/THM models discussed above. Table VIII shows the Montgomery Watson equations (9) for CHCl3, DCAA, and TCAA and their prediction success rates.

Table VIII. Montgomery Watson Equations1 for THM and HAA Formation

Percent of predictions within:

{Special Char 177 in Font "Symbol"}

7 {Special Char 109 in Font "Symbol"}g/L of

measured value3 at{Special Char 177 in Font "Symbol"}

3 {Special Char 109 in Font "Symbol"}g/L of

measured value3 at


























1. Reference (9)

2. UV = ultraviolet absorbance at 254 nm in cm-1. T = temperature in {Special Char 176 in Font "Symbol"}C t = time in hours. CHCl3, DCAA, and TCAA in {Special Char 109 in Font "Symbol"}g/L. TOC, Cl2, and Br-

in mg/L

3. Means and ranges of test data values ({Special Char 109 in Font "Symbol"}g/L): CHCl3: 29 (14- 59); DCAA: 19 (8-37); TCAA: 22 (11-40).

The Montgomery Watson (9) equations were quite successful at predicting HAA concentrations using data from both the LGTP and the LSTP where the success range was defined as {Special Char 177 in Font "Symbol"} 7 {Special Char 109 in Font "Symbol"}g/L, even though most of the UV absorbance values in the test data are outside the "boundary conditions range" (the range of values in the calibration data) given by Montgomery Watson (9). The test data mean UV absorbance value is 0.022 cm-1 while the lower limit of the boundary conditions range for the DCAA and TCAA equations is 0.05 cm-1.

The Montgomery Watson (9) equation systematically underpredicted CHCl3 concentrations, and the success rate for this equation was consequently very poor. Some of the test data UV absorbance values are outside the boundary conditions range for the CHCl3 equation (the lower limit of the boundary conditions range is 0.029 cm-1). In addition, all of the test data in this study are from treated water samples, but a substantial fraction of the database used to develop the CHCl3 model is from raw water samples. The Montgomery Watson equation tended to underpredict CHCl3 concentrations in treated waters during model validation (9).

In this study, the performance of the site-calibrated HAA vs. THM and the general Montgomery Watson (9) models for HAA formation was comparable. These models may be useful for making rough estimates of HAA concentrations, but they cannot be considered reliable if a high degree of accuracy is required.


Empirical models for DBPs and DBP precursors are useful for summarizing large data sets. They may help elucidate effects of process variables that would otherwise be overlooked, and they may be useful for predicting aspects of water quality, under specific treatment conditions, that have not been directly measured. The use of empirical models, developed from data collected over a four-year period at three RWA plants has led to the following conclusions:

(1) Removal of turbidity in water treatment processes may not indicate good removal of THM and HAA precursors. Ultraviolet absorbance at 254 nm is a superior parameter for process monitoring and optimization for the control of DBPs.

(2) The extent of THM and HAA precursor removal in biologically active GAC media with ozone applied upstream is proportional to influent concentration, and can be predicted using a first order model of the form

where z is the depth of penetration in the GAC bed, K is an apparent first order constant, {Special Char 108 in Font "Symbol"} is the biodegradable THMFP or HAAFP fraction, and Cmin is the concentration of biodegradable material below which no further biodegradation will occur.

(3) Haloacetic acid concentrations can be predicted as linear functions of THM concentrations. Using a small calibration data set (N=13), these models were successful at making rough predictions of haloacetic acid concentrations ({Special Char 177 in Font "Symbol"}7 {Special Char 109 in Font "Symbol"}g/L). The performance of the models for DCAA and TCAA formation developed by Montgomery Watson (9) was comparable to the performance of the site-calibrated HAA vs. THM models, although some of the test data values were outside the range of values in the calibration data. The Montgomery Watson model for chloroform was less successful than their HAA models.

(4) The HAAFP:THMFP ratio is lower in treated waters than in raw waters because of the preferential removal of TCAA precursors in coagulation/filtration processes. This effect seems to be more pronounced in conventional treatment than in direct filtration.

(5) A correlation between increased anionic polymer filter aid dose and poorer removal of DBP precursors was observed. It may be that filter aid has a direct inhibiting effect on DBP removal or it may be that filter aid is related to some other water quality or operational parameter that adversely effects removal.



A/S Anthracite/sand

CHCl3 Chloroform

DBP Chlorine disinfection by-product

DCAA Dichloroacetic acid

DOC Dissolved organic carbon

FP Formation potential

GAC Granular activated carbon

GAC/S GAC/sand

HAA Haloacetic acid

LGTP Lake Gaillard Treatment Plant

LSTP Lake Saltonstall Treatment Plant

RWA South Central Connecticut Regional Water Authority

SDS Simulated distribution system

TCAA Trichloroacetic acid

THM Trihalomethane

TOC Total organic carbon

UV absorbance Ultraviolet absorbance at 254 nm

WRTP West River Treatment Plant

Symbols:{Special Char 108 in Font "Symbol"}

Biodegradable fraction

A Alum dose

b Regression line Y-intercept

CE Effluent concentration

Co Influent concentration

Cmin Minimum biodegradable concentration

EBCT Empty bed contact time

F Filter aid dose

K Apparent first-order constant

m Regression line slope

P-Value Observed level of statistical significance

R2adj Adjusted correlation coefficient

T Temperature

UVE Effluent UV absorbance

UVo Influent UV absorbance

z Penetration depth in adsorption bed


Data collected by UMass graduate students James Ayers, Raashina Humayan, Paul Schmidt, Denise Springborg, Nagaraju Vinod, Jonathan Weiner, and Qing-wen Zhu were used in this work. Operation of the pilot plants was supervised by Howard Dunn and Gary Kaminski, formerly of the South Central Connecticut Regional Water Authority. The funding provided by the South Central Connecticut Regional Water Authority is gratefully acknowledged.


1. Pontius, F. W.; J. AWWA, 1995, 87(2), 48.

2. Stevens, A. A. et al.; J. AWWA, 1976, 68, 615.

3. Kavanaugh, M.C.; Trussell, A.R.; Cromer, J.; Trussell, R.R.; J. AWWA, 1980, 72, 578.

4. Engerholm, B. A.; Amy, G. A.; J. AWWA, 1983, 75, 418.

5. Urano, K.; Wada, H.; Takemasa, T.; Water Res., 1983, 17, 1797.

6. Morrow, C. M.; Minear, R. A.; Water Res., 1987, 21, 41.

7. Edzwald, J. K.; Organic Carcinogens in Drinking Water, Ram, N. M.; Calabrese, E. J.; Christman, R. F.; Eds.; John Wiley and Sons: New York, NY, 1986; pp 199-236.

8. Amy, G. L.; Chadik, P. A.; Chowdhury, Z.; J. AWWA, 1987, 79(7), 89.

9. Montgomery Watson; Final Report: Mathematical Modeling of the Formation of THMs and HAAs in Chlorinated Waters; American Water Works Association: Denver, CO, 1993.

10. Edzwald, J. K.; Becker, W. K; Wattier, K. L.; J. AWWA, 1985, 77(4), 22.

11. Huck, P. M.; Zhang, S.; Price, M. L.; J. AWWA, 1994, 86(6), 61.

12. Harrington, G. W.; Chowdhury, Z. K.; Owen, D. M.; J. AWWA, 1992, 84(11), 78.

13. Tobiason, J. E.; Chemical Water and Wastewater Treatment III, Klute, R.; Hahn, H. H; Eds.; Springer-Verlag: Berlin, 1994; pp 155-171.

14. Standard Methods for the Examination of Water and Wastewater, APHA, AWWA, WPC, Washington, D.C., 1992, 18th Ed.

15. Xie, Y. D.; Springborg, D. C.; Reckhow, D. A.; Proceedings of AWWA Water Quality Technology Conference, Miami, FL, 1993.

16. Koch, B.; Crofts, E. W.; Schimpff, W. K.; Davis, M. K.; Proceedings of the AWWA Water Quality Technology Conference, 1988.

17. Neter, J.; Wasserman, W.; Kutner, Michael H.; Applied Linear Statistical Models; Irwin: Burr Ridge, IL, 1990.

18. Miltner, R. J.; Shukairy, H. M.; Summers, R. S; J.AWWA, 1992, 84(11), 53.

19. Reckhow, D. A.; Singer, P. C.; J.AWWA, 1990, 82(4), 173.