Chemosphere, Vol. 19, Nos. 1-6, pp 337-344, 1989
Relations between different chlorinated aromatics have been evaluated in 66 samples from various industrial activities. For municipal waste combustion partial least squares modelling with latent variables (PLS) can explain 86 % of the variance in PCDD/PCDF from the isomerspecific analytical data for chlorinated benzenes and phenols.
Indicator parameters; Chlorinated aromatics; Polychlorinated dioxins; Polychlorinated dibenzofurans; Statistical analysis; Multivariate modelling.
Sampling and analysis of micro-pollutants in flue gases increase in complexity and cost as the detection limit is lowered; in particular this applies to ultratrace components like polychlorinated dioxins (PCDD) and dibenzofurans (PCDF). With suitable indicator parameters it may be feasible both to enhance the precision in the measurements as well as to reduce the costs. The correlation between chlorinated benzenes and PCDD/PCDFs, as reported by us and others (1, 2, 3), can form the basis for indirect measurements.
We have previously reported data from combustion studies with different types of waste demonstrating the possibility not only of predicting emission levels, but also the chlorination pattern from one group of chlorinated aromatics to another (4). We have also reported measurement data for brominated aromatics suggesting that indicator parameters may be the only feasible approach to control halogenated aromatics in flue gases (5).
The sampling and analytical methods have been described elsewhere (5,6,7).
Laboratory data were evaluated using different descriptive statistical techniques, regression analysis and partial least square modelling with latent variables (PLS). The statistical analysis and multivariate modelling of data were carried out using the data programs STATGRAPHICS 2.1 from STSC, Inc. (license no 206187) and SIMCA-3B from Sepanova AB (license no 861210-SR1). The analysis and modelling of multivariate data with SIMCA has been reviewed by Wold et al (8).
All data used in this investigation originated from our own laboratory. MILJÖKONSULTERNA also performed most of the sampling work. The 66 laboratory samples chosen for further evaluation represent various industrial activities, table 1.
Table 1. Laboratory samples and industrial activities.
| Activity | No | % |
| Municipal waste combustion | 33 | 50 |
| Municipal & wood waste combustion | 8 | 12 |
| Hazardous waste combustion | 8 | 12 |
| Hospital & special waste combustion | 5 | 8 |
| Metallurgical industry | 9 | 14 |
| Others | 3 | 5 |
The flue gas samples were collected both upstream (raw gas) and downstream of flue gas cleaning systems with different separation efficiencies.
Each sample is described by 41 variables representing the analytical results for PCDD/PCDF, chlorobenzenes and chlorophenols. Measurement results below the detection limit were assigned a value of half this limit and all data were log-transformed.
The emission of specific isomers of PCDD/PCDF is closely related to the total production of chlorinated aromatics, but is also influenced by the chlorine in-put, the plant design and the flue gas cleaning system. These three factors determine the chlorination pattern of both PCDD/PCDF and other chlorinated aromatics (benzenes and phenols). The separation efficiency in the flue gas cleaning system is related through those to the chlorination pattern, since highly chlorinated isomers have lower volatility than those with fewer chlorine atoms.
A multiple linear regression model for 2378-TCDD will be used to illustrate the reasoning concerning the chlorination pattern. Dichlorobenzenes and hexchlorobenzene are used as independent variables and both contribute significantly to the regression, p <0.0001. In figure 1 we show the plot of the predictions, n = 63. The independent variables are not intercorrelated, r = 0.04.
Figure 1. Analytical vs predicted results for 2378-TCDD, ln ng.
The regression, ln 2378-TCDD = -2.38 + 0.38 ln HCB + 0.63 ln DCB, can be visualized as a response surface, figure 2. Large amounts of TCDD are found in samples with large amounts of both hexachlorobenzene and dichlorobenzenes.
Figure 2. Response surface, 2378-TCDD vs hexachlorobenzene and dichlorobenzenes.
The prediction error (or residuals) in the above multiple linear regression model, figure 1, is in our view not small enough for the purpose of using chlorinated benzenes as indicator parameters. However, using all analytical results and not only a small fraction might have an impact.
Multiple linear regression cannot be used to solve the prediction problem as a whole due to the multicollinearity situation, i.e. high correlation between different chlorobenzenes and chlorophenols. We have therefore used PLS to develop a coherent prediction model for PCDD/PCDF. With PLS a projection is calculated that both approximates the independent variables and models the dependent variables to give the best possible correlation.
Chlorinated benzenes and phenols were defined as the block of independent variables (19 different isomers), and PCDD/PCDF were defined as the block of dependent variables (14 different isomers, 8 different groups of congeners and TCDD-equivalents).
Two significant components were found, as checked by cross-validation. This PLS-model explains 67 % of the variance in the dependent variables. In figure 3 we show the predicted versus the analytical results for 12378-PeCDF (64 samples).
Figure 3. Predicted vs analytical results for 12378-PeCDF, log ng.
The samples used in this model originate from a variety of industrial activities. To further reduce the prediction error it is necessary to develop more limited models for samples similar to each other.
Samples from municipal waste combustion (MWC) can be used as an example. A PLS-model with two significant components explains 86 % of the variance in the dependent variables, 10 different 2,3,7,8-substituted tetra-, penta- and hexachlorinated dibenzodioxins and dibensofurans, with a variation between 80 - 92 % for the individual variables. The second component indicates differences in the chlorination pattern. In figures 4 - 7 we show prediction results for two of these toxic isomers.
Figure 4. Predicted vs analytical results for 2378-TCDF, ng.
Figure 5. Predicted vs analytical results for 12378-PeCDD, ng.
Figure 6. Predicted vs analytical results for 12378-PeCDF, ng.
Figure 7. Predicted vs analytical results for 234678-HxCDF, ng.
These plots show that specific isomers of PCDD/PCDF from chlorinated benzenes and phenols can be predicted, with high accuracy over a concentration range of more than four orders of magnitude. The deviation between actual measurements and predictions is seldom more than a factor of three. The residuals are normally distributed.
Finally we divided the MWC-data set in two parts, training set and test set, for additional validation of the PLS-modelling for chlorinated aromatics in flue gases. The training set consisted of 16 samples from 10 different plants. A PLS-model was calculated with TCDD-equivalents (Eadon) as the dependent variable, and chlorinated benzenes and phenols as independent variables. The first two components were found to be significant, explaining 90 % of the variance in the training set.
The test set (17 samples from 11 different plants) was not included while building up the model. These test objects were all classified as similar to the training set, i.e. inside the class tolerance interval of the model. The prediction results for the test objects are presented in figure 8.
Figure 8. Predicted vs analytical results for TCDD-equivalents in the test set, ng.
The amounts and isomer distribution of polychlorinated dibenzo-p-dioxins (PCDD) and dibenzofurans (PCDF) vary substantially in flue gas samples from industrial processes. These variations can be predicted from chlorinated benzenes and phenols. A multivariate calibration model for samples from municipal waste combustion shows high accuracy over a concentration range of four orders of magnitude.
Indirect measurements of PCDD/PCDF with chlorinated benzenes and phenols as indicator parameters are useful for several purposes.
In all these cases the indirect measurements can be performed at lower cost and also with higher precision. A reliable calibration set is however a necessary prerequisite.
To make optimal use of these advantages a simplified sampling train for chlorinated benzenes should be developed and validated against existing equipment.
The potential of this approach for the control of other chlorinated, brominated and mixed halogenated aromatics should also be evaluated.
The work described in this report has been funded by the Swedish National Energy Administration.
We gratefully acknowledge the use of data from measurements carried out for various Swedish industries.
Reprinted from Chemosphere, Volume 19, Öberg, T., Bergström, J. Indicator parameters for PCDD/PCDF, Pages No. 337-344, Copyright (1989), with permission from Elsevier Science. Single copies of the article can be downloaded and printed for the reader's personal research and study.
|
Swedish homepage