Modeling of the Polluting Emissions from a Cement Production Plant

Dec 2, 2005 - Soft models, based on linear (partial least-squares, PLS, and principal component regression, PCR) and nonlinear (artificial neural netw...
0 downloads 0 Views 290KB Size
Download PDF
Environ. Sci. Technol. 2006, 40, 272-280

Modeling of the Polluting Emissions from a Cement Production Plant by Partial Least-Squares, Principal Component Regression, and Artificial Neural Networks EMILIO MARENGO,* MARCO BOBBA, ELISA ROBOTTI, AND MARIA CRISTINA LIPAROTA Department of Environmental Sciences and Life, University of Eastern Piedmont, Via Bellini 25/G, 15100 Alessandria, Italy

A Portland cement process was taken into consideration and monitored for one month with respect to polluting emissions, fuel and raw material physical-chemical properties, and operative conditions. Soft models, based on linear (partial least-squares, PLS, and principal component regression, PCR) and nonlinear (artificial neural networks, ANNs) approaches, were employed to predict the polluting emissions. The predictive ability of the three regression methods was evaluated by means of the partition of the dataset by Kohonen self-associative maps into both a training and a test set. Then, a “leave-more-out” approach, based on the use of a training set, a test set, and a production set, was adopted. The training set was used to build the models, the test set was used to select the number of latent variables or the neural network training endpoint, and the production set was used to produce genuine predictions. ANNs proved to be much more effective in prediction with respect to PLS and PCR and, at least in the case of SO2 and dust, provided a predictive ability comparable with the experimental estimated uncertainty of the response. This showed that it is possible to satisfactorily predict the two responses. Such a prediction will result in the prevention of environmental and legal problems connected to the polluting emissions.

Introduction Cement is a basic material used for building construction and for civil engineering projects. Cement industries typically produce Portland cement, which is a fine, typically gray powder, comprised of dicalcium silicate, tricalcium silicate, tricalcium aluminate, and tetracalcium aluminoferrite, with the addition of various forms of calcium sulfate. Cement preparation involves mining, crushing, and grinding of raw materials. The most important step is calcination of calcium carbonate followed by burning the resulting calcium oxide together with silica, alumina, and ferrous oxide at high temperatures to form clinker (sintering). The clinker is then cooled and ground (or milled) with gypsum and other constituents to produce cement. The clinker burning takes place in a rotary kiln which needs a lot of energy to reach the temperature necessary for * Corresponding author phone: +39 0131 360272; fax: +39 0131 360250; e-mail: [email protected]. 272

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 40, NO. 1, 2006

the calcination and sintering processes. Traditionally, the fuels used in the cycle of production of the cement are heavy combustible materials such as oil and coal but the favorable conditions of the kiln allow the use of alternative fuels such as exhausted motor oil, spent solvents, printing inks, paint residues, cleaning fluids, scrap tires, and plastics (1-4). Production of the cement generates a variety of wastes but the major releases are emissions from the kiln system into the atmosphere. Emissions are formed from the combustion of raw materials and fuels in the rotary kiln and are principally nitrogen oxides (NOx), sulfur oxides (SOx), carbon monoxide (CO), carbon dioxide (CO2), and dust (5-9). A more detailed description of cement production is provided in the Supporting Information. The purpose of this paper is to develop a mathematical model to predict the gas emissions from the cement production process. The independent variables considered to build the model are the chemical properties of raw materials, the chemical and physical-chemical properties of fuels (both traditional and alternative), and the kiln conditions. These variables are expected to affect the emissions of polluting agents into the atmosphere. Principal component regression (PCR) (10-12) and partial least-squares regression (PLS) (13-15) are the two regression methods most frequently used in solving similar problems. They assume a linear relationship between x and y variables and provide correct results, as long as deviations from linearity are not too strong. With high nonlinearity, alternative modeling methods must be applied. In this work, artificial neural networks (ANNs) (16, 17) are used. ANNs are mathematical algorithms that simulate the capacity of the biological “brain” to solve complex problems and they have the ability to approximate virtually any function in a stable and efficient way. ANNs have already been applied to many environmental problems (18-22). In this case, they were employed for predicting the amount of gas emissions. Optimal network architecture selection is the principal problem to be solved to obtain a good predictive model. For this reason, the predictive ability of several architectures of back-propagation neural networks was evaluated. The results provided by the neural networks were compared with the results obtained with the classical PCR and PLS algorithms. The selection of a representative training set, used for training the networks and for evaluating their predictive ability, was performed with the unsupervised Kohonen network (23-30).

Releases from Cement Kilns The main releases from the production of cement into the atmosphere are from the kiln exhaust gases, the clinker cooler, and the bypass gases. The rotary kiln used in the Robilante plant has a length of 78.0 m, a diameter of 4.6 m, and a slope of 3%. The oven is able to produce 6 tons per day of clinker. The kiln consists of three principal sections: a thermal exchanger, a rotating oven, and a cooling system. The preheating of raw materials is performed by recycling the gases produced inside the kiln, while the clinker, outgoing from the burner, is cooled by a grate cooler. Polluting releases result from the particular chemical composition of the raw materials and fuels employed. In this work, the emissions of nitrogen oxides, sulfur oxides, and dust were analyzed. Nitrogen Oxides. There are two principal mechanisms acting in any combustion process to form NOx: the oxidation 10.1021/es0517466 CCC: $33.50

 2006 American Chemical Society Published on Web 12/02/2005

TABLE 1. Emission Limits SO2 emission limit (mg/Nm3)

dust emission limit (mg/Nm3)

NO2 emission limit (mg/Nm3)

600

50

1800-3000

of molecular nitrogen in the combustion air (thermal NOx) and the oxidation of nitrogen compounds in the fuel (fuel NOx). In the production of cement, thermal NOx is a significant formation route. Thermal NOx formation is strongly dependent on the combustion temperature, with a marked increase in formation above 1400 °C. When the flame temperature in a kiln, during cement production, is around 2000 °C, it is in the sintering zone where most of the thermal NOx is formed. In the calcination stage, temperatures are 800-900 °C, which is not high enough to form significant thermal NOx compared with fuel NOx. NOx formation is also dependent on the amount of excess air present in the flame, with higher oxygen contents enhancing formation. Different raw material characteristics can influence the amount of NOx produced. For instance, some limestone requires far less burning than others to produce cement clinker and, consequently, less thermal and fuel NOx are released. In addition, some raw materials contain chemically bound nitrogen, which can convert to NOx at temperatures between 300 and 800 °C. However, this source of NOx is usually not significant in the cement production process. Sulfur Oxides. Sulfur oxides, released from cement production, can occur in the kiln exhaust. The release is mainly in the form of SO2 (99%) although some SO3 is produced and, in reducing condition, H2S as well. Sulfur oxides arise due to the sulfur content of both the fuel and raw materials. Raw materials, such as limestone, can contain sulfur in the form of sulfates (e.g. calcium sulfate), sulfides (e.g. pyrites), and organic compounds. The proportion of sulfur released from cement kilns depends on the balance between the absorption and release of SO2 at various stages of the process. The absorption capacity of a kiln varies with chemistry (alkali, sulfate, and chloride balance; oxygen content), temperature, and kiln design. Cement Kiln Dust. Cement kiln dust (CKD) is the powder retrieved from the exiting gases and is either all or partly returned to the operation or removed entirely. The type of system, the chemical makeup of the raw materials and fuel, and the condition of the system operations all affect the chemical configuration of the CKD. Table 1 shows the Italian legal limits for the emissions of SO2, dust, and NO2 (Ministerial Decree, 12 July 1990). The principal problem for the cement industry is to keep the level of the gaseous emissions within the legal limits. Therefore, monitoring the variables which can affect the amount of emission and understanding what influence these variables have on the formation of the emissions is very important in order to establish optimal operative conditions and avoid environmental problems. The aim of this work is to build mathematical models that are able to predict the amounts of SO2, NO2, and dusts which will be produced, in relation to the characteristics of the raw materials and fuels, and operative conditions of the kiln.

Theory Artificial Neural Networks (ANNs). ANNs are mathematical algorithms that mimic the ability of the human brain to solve complex problems. Just as humans apply knowledge gained from past experience to new problems or situations, a neural

network takes previously solved examples to build a system of neurons that make new decisions, classifications, and forecasts. Neural networks look for patterns in training sets of data, learn these patterns, and develop the ability, if the system does not change or drift out of the training set range of the predictors, to correctly classify new patterns or to make forecasts and predictions. ANNs have the ability to approximate virtually any function in a stable and efficient way and, for this reason, they are used to solve problems with the presence of nonlinear relationships. Two types of neural networks are applied in this study: supervised (back-propagation) and unsupervised (Kohonen) networks. Supervised ANNs build models which classify patterns and make predictions according to other patterns of inputs and output they have learned. They give the most reasonable answer based upon the variety of learned patterns. Backpropagation network (31-35) is one of the most used supervised network types. In this case, several architectures were investigated and the most efficient network contained only one hidden layer with five neurons, with a Gaussian transfer function from input to hidden layer and the symmetric logistic function as transfer function from the hidden to the output layer. The best values of learning rate and momentum were 0.3 and 0.1, respectively, and the data were range-scaled in the interval 0.2-0.8. Unsupervised networks can group the objects of a dataset into different classes on the basis of their similarity. Kohonen network is a typical unsupervised network. A total of 500 iterations were used, with the learning rate decreasing linearly from 0.5 to 0.01. At the same time, the range of the weights correction decreased from the maximum range to 0. A more detailed explanation of back-propagation and Kohonen network theory is reported in the Supporting Information. Partial Least-Squares (PLS) and Principal Component Regression (PCR). PLS and PCR are calibration methods based on the decomposition of the dataset into a reduced set of orthogonal latent variables and linear combination of the original x variables. Their algorithms have been widely described elsewhere (10-15). Models Evaluation. The coefficient of multiple determination, R2, for PCR, PLS, and ANN models was calculated as:

R2 ) 1 -

∑ ∑

x

ˆi i)1,n(y

- yi)2

i)1,n(yi

- yj)2

where the two sums run on the number of samples n of the training set, yˆ is the predicted value of the response for the i-th experiment, yi is the experimental value, and yj is the average response. An analogous expression was used to evaluate the predictive ability of the models (R2 PrSet), using the production set experiments (the two sums run on the experiments of the production set). The root-mean-square error (RMSE) between the measured and predicted values is estimated as:

RMSE )

∑ (yˆ - y )

x

i

i

2

i

n

that again can be calculated using either training (RMSEF, root-mean-square error of fitting) or production (RMSEP, root-mean-square error of prediction) set experiments, to achieve information about fitting and predicting ability. VOL. 40, NO. 1, 2006 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

273

FIGURE 1. Cement process. Moreover, the Wilcoxon signed rank test (36) was employed to compare the predictive performance of pairs of models calculated with the same dataset. The test provides a value of probability Φ(z) that, if sufficiently small (usually less than 0.05), suggests that the first model outperforms the second one. The test results in the opposite conclusion if Φ(z) is sufficiently large (usually greater than 0.95).

Experimental Section A representation of the cement production process is shown in Figure 1; the scheme highlights the principal phases of the process, the flow of the materials, and the flow of the gases/ emissions. Furthermore, the diagram shows the different points where the data were collected: (i) Downstream of the mill process of the raw materials (X-ray fluorescence spectrometer XRF). (ii) Downstream of the production process of the charcoal powder (the chemical composition and the calorific value of the charcoal powder are measured by ion chromatography and calorimetry). (iii) Incoming alternative fuel (the chemical composition and the calorific value are measured by ion chromatography and calorimetry). (iv) In the kiln (the chemical composition of the clinker is measured by an X-ray fluorescence spectrometer XRF). (v) In correspondence of the conditioning tower (analysis of gaseous releases by gas chromatography). Dataset. The dataset consisted of 356 samples taken at irregular intervals over 25 successive working days of January 2001 in the cement production plant of Buzzi Unicem (Robilante, Cuneo, Italy). NO2, SO2, and dust emissions are the responses taken into consideration, while the physical-chemical properties of the raw materials and fuels (both traditional and alternative) and the process operative conditions are the descriptors employed for modeling the responses. The independent variable set consists of the following descriptors: (i) Mill Raw Materials. Lime combination factor, silica, and alumina modules were used. (ii) Conventional Kiln Fuels. In this process, petroleum coke and fossil coal are mixed to form the charcoal powder. 274

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 40, NO. 1, 2006

The fuel properties measured are the higher calorific value, expressed as kcal/kg, the percentage of volatile substances, ashes, chlorine (as Chlorine), volatile sulfur (as Sulfur), nitrogen (as Nitrogen), fossil coal, and petroleum coke. (iii) Alternative Fuels. The monitored characteristics of alternative fuels are the higher calorific value (kcal/kg), the percentages of moisture, sulfur as S and as SO3, nitrogen as N, and chlorine as Cl. (iv) Kiln Conditions. The most important variable is the temperature of the material in the kiln, which greatly depends on the raw materials and fuels employed. Other variables are the flow of raw material in the kiln (ton/h) and the flow of charcoal powder fed in the head, and in the precalcination zone of the kiln. The final dataset consists of 356 samples described by 22 variables, 19 descriptors, and 3 responses. Software. Back-propagation ANNs were calculated by NeuroShell2 (release 4.0, Ward Systems Group, Frederick, MD), and PCR and PLS models by Unscrambler (version 7.6, CAMO ASA, Oslo, Norway), while Kohonen SOMs were calculated by a self-developed software written in Visual Basic 6.0 (Microsoft, U.S.A.). Microsoft Excel was used as visualization program.

Results and Discussion The Kohonen network was used to select the training set (37-40). The aim was to find the samples that guaranteed a homogeneous representation of the whole experimental domain. For this purpose the 356 samples were the input for an 8 × 8 × 19 Kohonen network; the network was trained for 500 epochs, which permitted us to achieve convergence so that the configuration of the first layer (top-map) did not change anymore. In the end, the samples were assigned to the cells of the top-map (Figure 2), on the basis of the similarity of the 19 descriptors. Approximately one-third of the samples present in each cell were randomly selected and assigned to the test set. This led to a partition of the dataset into a training set with 237 samples and a test set with 119 samples. The training set was used for training the neural network and, in particular, for optimizing its weights by the backpropagation algorithm. The test set was used to determine

FIGURE 2. Occupation of the 8 × 8 Kohonen top-map.

TABLE 2. Summary of Results for SO2 Emission Obtained with the Three Calibration Methods

ANN PCR (7 PCs) PLS (6 LVs)

R2 TrSet

R2 PrSet

RMSEF TrSet (mg/Nm3)

RMSEP PrSet (mg/Nm3)

0.96 0.69 0.74

0.82 0.51 0.53

1.7 4.1 3.7

3.0 5.1 5.0

the epoch to interrupt the network training. So the selection of when to stop the training was performed using the prediction ability of the network with respect to the test set samples. To evaluate the predictive ability of the network, the so-called production set was defined. This was obtained by eliminating, from both training and test sets, all the samples belonging to the same day, in a sort of “leave-moreout” cross-validation algorithm. The responses for the samples left out during the training of the network were then predicted by the network and optimized using the pruned training and test set. This procedure guaranteed that real genuine predictions were performed, since the samples predicted were neither used for training the network nor for deciding when to interrupt the training. This procedure was repeated for all days, leaving out 1 day at a time. The procedure described above was used for evaluating the predictive ability of all regression models, including PLS and PCR. In the case of PLS and PCR, the training set was used for building the model and the RMSEP of the test set was used for selecting the number of latent variables to be inserted into the models. The number of latent variables corresponding to the minimum of the RMSEP curve calculated on the test set was selected. SO2 Emission. Table 2 shows the coefficients of multiple determination and the RMSE for SO2 obtained with ANN, PCR, and PLS. The artificial neural network provides the best results both in fitting and prediction, with a coefficient of multiple determination of 0.96 for the training set and 0.82 for the production set. PLS provides results slightly better than PCR, but the prediction ability of both methods is not comparable to the prediction ability of ANN. The values of RMSEP reflect this result; in particular it can be stated that the RMSEP obtained by ANN is much better than the RMSEP obtained with the other two algorithms! If genuine predictions of whole days eliminated from both test and training sets are considered and used to evaluate the models’ predictive ability, it can be affirmed that the artificial neural network performs really well for SO2.

FIGURE 3. ANN, PLS, and PCR results: predicted vs measured SO2 emission for the production set. The Wilcoxon test results show that the ANN model, compared with the PCR and PLS models (Φ(z) ) 0.0), has the best predictive ability, while the two other methods are equivalent (Φ(z) ) 0.17 for PCR compared to PLS). Figure 3 shows the predicted vs experimental responses for the production set obtained with ANN, PLS, and PCR. Dust Emission. The coefficients of multiple determination obtained and the RMSE for both the training and production sets for dust emission are shown in Table 3. As in the previous case, the optimum result is obtained with the artificial neural network. The coefficient of multiple determination obtained with the ANNs for the production set is 0.65 with an RMSEP equal to 1.2 mg/Nm3. The RMSEP is comparable with the uncertainty of the experimental response which is estimated around 10%. PCR and PLS provide results that are comparable with each other but much worse than the results obtained by VOL. 40, NO. 1, 2006 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

275

TABLE 3. Summary of Results for Dust Emission Obtained with the Three Calibration Methods

ANN PCR (9 PCs) PLS (4 LVs)

R2 TrSet

R2 PrSet

RMSEF TrSet (mg/Nm3)

RMSEP PrSet (mg/Nm3)

0.80 0.58 0.60

0.65 0.35 0.41

0.9 1.4 1.3

1.2 1.7 1.6

TABLE 4. Summary of Results for NO2 Emission Obtained with the Three Calibration Methods

ANN PCR (9 PCs) PLS (4 LVs)

R2 TrSet

R2 PrSet

RMSEF TrSet (mg/Nm3)

RMSEP PrSet (mg/Nm3)

0.60 0.28 0.36

0.22 -0.07 0.01

88.8 119.6 112.8

124.5 146.3 140.5

ANNs (R2 ) 0.35 and 0.41, respectively, for the production set), but PCR needs many more principal components to obtain this result. The Wilcoxon test results show that the ANN model, compared to the PCR and PLS models (Φ(z) ) 0.0), has the best predictive ability; in this case, the two other methods are not equivalent, PLS being better than PCR (Φ(z) ) 0.99 for PCR compared to PLS).

Figure 4 shows the predicted vs experimental responses of the production set for ANN, PLS, and PCR. NO2 Emission. Table 4 shows the results obtained with the three algorithms in regards to NO2 emissions. Again, ANN provides the best results, both in fitting and prediction. It must be remarked that, in this case, the two linear models do not show any capacity to predict the data left out; in fact R2 is either negative or close to 0. Also, ANN’s predictive ability is not very satisfactory, but at least a trend can be

FIGURE 4. ANN, PLS, and PCR results: predicted vs measured dust emission for the production set.

FIGURE 5. ANN, PLS, and PCR results: predicted vs measured NO2 emission for the production set.

276

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 40, NO. 1, 2006

FIGURE 6. Network derivative for SO2 emissions with respect to the 19 input variables. identified, Figure 5, which shows the experimental vs the predicted SO2 values for the production set can be identified. The Wilcoxon test was not performed in this case given the unsatisfactory predicted results (Φ(z) ) 0.0). ANN has

the best predictive ability, while the two other methods are equivalent (Φ(z) ) 0.91 for PCR compared to PLS). It is clear that the ANN model in this case cannot be used for predicting the NO2 emissions using the kiln conditions VOL. 40, NO. 1, 2006 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

277

FIGURE 7. Network derivative for dust emissions with respect to the 19 input variables. and the physical-chemical properties of raw material and fuels as variables. There is probably a problem in the description of the system so that some relevant factors that greatly affect NO2 emission are not taken into account. 278

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 40, NO. 1, 2006

The ANNs are considered a “black box”. This view stems from the fact that the contribution of the input variables in predicting the value of the output is difficult to disentangle within the network. Consequently, input variables are often

entered into the network and an output value is generated without gaining any understanding of the inter-relationships between the variables, and therefore, providing no explanatory insight into the underlying mechanisms modeled by the network. In this work the calculation of the first derivative of the best network was performed to understand the influence of the input variables on the responses. In this way, it was possible to find the real influence that the input variables have on the SO2 and dust responses; these results can be useful to reduce the risk of exceeding the legal limits because input variables can be varied to control the gas emissions. Figures 6 and 7 report the value of the ANN first derivative (SO2 and dust emissions, respectively) with respect to each input variable vs the scaled value (from 0 to 1) of the corresponding input variable. SO2 Emission. From Figure 6 it is possible to observe which variables have more influence on the SO2 emissions. The network derivative applied to SO2 emissions shows that some input variables do not have a high influence on the output; in fact the derivative value is around zero. These variables are: temperature of the kiln, flow of raw materials, percentage of nitrogen of alternative fuels, and the percentage of ashes, nitrogen, and chlorine of conventional fuels. Some input variables show an approximately constant value of the first derivative instead. Positive constant value means that the variable has a positive linear influence on the response; therefore, if the input variable increases, the output variable also increases. The opposite reasoning can be made for a negative value of the derivative. The variables that show a positive, approximately constant value are the percentage of moisture, of S and of SO3, of alternative fuels, and the percentage of volatile S of conventional fuels, while the variables that have a negative derivative value are the alumina module and higher calorific value of alternative fuels. For the remaining input variables, the calculation of the first derivative shows a more complex trend due to the more complex relations between input and output variables (for example quadratic and cubic relations). Dust Emission. From Figure 7 it is possible to observe which variables have more influence on the dust emissions. Also, in this case, the network derivative applied to dust emissions shows that some input variables have a first derivative around zero. These variables are the lime combination factor, the flow of charcoal powder in the head of the kiln, the percentage of S and SO3 of alternative fuels, and the percentage of ashes, volatile S, and Cl of conventional fuels. These input variables do not have a high influence on the output. The input variables that have a positive constant derivative are the silica module, the percentage of moisture of alternative fuels, and the higher calorific value of alternative and conventional fuels. These variables have a positive linear influence on the response. The remaining input variables show a more complex trend of the first derivative. The back-propagation artificial neural models, for SO2 and dust emissions, were simplified by taking these results into consideration. The models have been recalculated without the input variables that have the first derivative around zero; that is, the input variables that do not have a high influence on the response have been eliminated. The results are shown in Table 5. For the SO2 emissions, the final network has been trained with only 13 variables instead of the 19. In this case, the R2 of the production set is equal to 0.86 and the RMSEP is equal to 2.7 mg/Nm3, which are better with respect to those obtained with all variables (R2 of 0.82 for the production set and RMSEP of 3.0 mg/Nm3), showing that the inclusion of

TABLE 5. Summary of Results for SO2 and Dust Emissions Obtained with the Reduced ANN Models R2 TrSet R2 PrSet SO2 emissions dust emissions

0.98 0.80

0.86 0.66

RMSEP TrSet RMSEP PrSet (mg/Nm3) (mg/Nm3) 1.5 0.9

2.7 1.2

nonrelevant information in the first ANN model caused a loss of predictive ability and consequent overfitting. For the dust emissions, the final ANN model has been calculated with only 12 variables and provides a coefficient of multiple determination for the production set of 0.66 and a RMSEP equal to 1.2 mg/Nm3. In this case, the results obtained with the reduced model are very similar to those obtained with the complete model and offer the advantage of using fewer variables. It is clear that the PCR and PLS algorithms perform, both in fitting and prediction, much worse than the ANN. For SO2 and dust emissions, the ANN models can be successfully used to predict the future amounts of emissions from the cement production process. In the case of NO2, the models are not satisfactory enough to be employed for predictive use. The obtained results must be judged considering that there are some problems connected with both the technological cycle and plant; the principal problems are attributed to the large size of the kiln ( 78 m long and 4.6 m in diameter) and to the difficulty in establishing an exact correspondence between the materials fed into the oven (raw materials and fuels) and the amounts of gases emitted. The gaseous emissions are recycled inside the plant. The results obtained are encouraging with regards to the possibility to predict the emissions to avoid environmental and legal problems, but a more exhaustive description of the system is needed (at least in the case of NO2) and more reliable determinations of both responses and descriptors must be performed. ANN, given the probable intrinsic nonlinearity of the system, seems to be the best choice with respect to the two linear approaches taken into consideration in this preliminary work. In the first part of the study, many independent variables were taken into consideration: the characteristics of the raw materials and fuels, and the operative conditions of the kiln. The second part of the study has been devoted to evaluating the real influence that every variable has on the amount of gas emitted and to using this information to refine the mathematical models achieved so far. This would produce a better understanding of the chemical and physicalchemical mechanisms involved in the emissions production and a more efficient tool to reduce the risk of the industrial plant exceeding the legal limits.

Supporting Information Available Detailed information about the process of Portland cement production and the main characteristics of cement and the theory relative to the ANNs used in this work (backpropagation and Kohonen networks). This material is available free of charge via the Internet at http://pubs.acs.org.

Literature Cited (1) Guo, Q.; Eckert, J. O., Jr. Heavy metal outputs from a cement kiln co-fired with hazardous waste fuels. J. Hazard. Mater. 1996, 51 (1-3), 47-65. (2) Mokrzycki, E.; Uliasz-Bochenczyk, A. Alternative fuels for the cement industry. Appl. Energy 2003, 74 (1-2), 95-100. (3) Trezza, M. A.; Scian, A. N. Burning wastes as an industrial resourcesTheir effect on Portland cement clinker. Cem. Concr. Res. 2000, 30 (1), 137-144. VOL. 40, NO. 1, 2006 / ENVIRONMENTAL SCIENCE & TECHNOLOGY

9

279

(4) Ka¨a¨ntee, U.; Zevenhoven, R.; Backman, R.; Hupa, M. Cement manufacturing using alternative fuels and the advantages of process modelling. Fuel Process. Technol. 2004, 85 (4), 293301. (5) Canpolat, B. R.; Atimtay, A. T.; Munlafalioglu, I.; Kalafatoglu, E.; Ekinci, E. Emission factors of cement industry in Turkey. Water, Air, Soil Pollut. 2002, 138 (1-4), 235-252. (6) Sheinbaum, C.; Ozawa, L. Energy use and CO2 emissions for Mexico’s cement industry. Energy 1998, 23 (9), 725-732. (7) Prisciandaro, M.; Mazziotti, G.; Veglio`, F. Effect of burning supplementary waste fuels on the pollutant emissions by cement plants: a statistical analysis of process data. Resour. Conserv. Recycl. 2003, 39 (2), 161-184. (8) Von Bahr, B.; Hanssen, O. J.; Vold, M.; Pott, G.; StoltenbergHansson, E.; Steen, B. Experiences of environmental performance evaluation in the cement industry. Data quality of environmental performance indicators as a limiting factor for Benchmarking and Rating. J. Cleaner Prod. 2003, 11 (7), 713725. (9) Lemieux, P.; Stewart, E.; Realff, M.; Mulholland, J. A. Emissions study of co-firing waste carpet in a rotary kiln. J. Environ. Manage. 2004, 70 (1), 27-33. (10) Massart, D. L.; Vandegisten, B. G. M.; Deming, S. N.; Michotte, Y.; Kaufman, L. Chemometrics: A Textbook; Elsevier: Amsterdam, 1988. (11) Esbensen, K.; Geladi, P.; Wold, S. Principal component analysis. Chemom. Intell. Lab. Syst. 1987, 2 (1-3), 37-52. (12) Martens, H.; Naes, T. Principal component regression in NIR analysis: viewpoints, background details and selection of components. J. Chemom. 1988, 2, 155-167. (13) Hoskuldsson, A. PLS regression methods. J. Chemom. 1988, 2, 211-228. (14) Haaland, D. M.; Thomas, E. V. Partial least-squares methods for spectral analyses. 1. Relation to other quantitative calibration methods and the extraction of qualitative information. Anal. Chem. 1988, 60, (11), 1193-1202. (15) Geladi, P.; Kowalski, B. R. Partial least-squares regression: a tutorial. Anal. Chim. Acta 1986, 185, 1-17. (16) Zupan, J.; Gasteiger, J. Neural network for chemist: an introduction; John Wiley & Sons: New York, 1993. (17) Hagan, M.; Demuth, H.; Beale, M. Neural Network Design; PWS Publishing: Boston, 1996. (18) Boznar, M.; Lesjak, M.; Mlakar, P. A neural-network-based method for short-term predictions of ambient SO2 concentrations in highly polluted industrial-areas of complex terrain. Atmos. Envrion. 1993, 27 (2), 221-230. (19) Rege, M. A.; Tock, R. W. A simple neural network for estimating emission rates of hydrogen sulfide and ammonia from single point sources. J. Air Waste Manage. Assoc. 1996, 46 (10), 953962. (20) Elkamel, A.; Abdul-Wahab, S.; Bouhamra, W.; Alper, E. Measurement and prediction of ozone levels around a heavily industrialized area: a neural network approach. Adv. Environ. Res. 2001, 5 (1), 47-59. (21) Kolehmainen, M.; Martikainen, H.; Ruuskanen, J. Neural networks and periodic components used in air quality forecasting. Atmos. Environ. 2001, 35 (5), 815-825. (22) Slanvetpan, T.; Barat, R. B.; Stevens, J. G. Process control of a laboratory combustor using artificial neural networks. Comput. Chem. Eng. 2003, 27 (11), 1605-1616. (23) Zupan, J.; Novic, M.; Ruisanchez, I. Kohonen and counterpropagation artificial neural networks in analytical chemistry. Chemom. Intell. Lab. Syst. 1997, 38 (1), 1-23.

280

9

ENVIRONMENTAL SCIENCE & TECHNOLOGY / VOL. 40, NO. 1, 2006

(24) Melssen, W. J.; Smits, J. R. M.; Rolf, G. H.; Kateman, G. 2-dimensional mapping of IR-spectra using a parallel implemented self-organizing feature map. Chemom. Intell. Lab. Syst. 1993, 18 (2), 195-204. (25) Song, X. H.; Hopke, P. K. Kohonen neural network as a pattern recognition method based on the weight interpretation. Anal. Chim. Acta 1996, 334 (1-2), 57-66. (26) Melssen, W. J.; Smits, J. R. M.; Buydens, L. M. C.; Kateman, G. Using artificial neural networks for solving chemical problems. 2. Kohonen self-organizing feature maps and Hopfield networks. Chemom. Intell. Lab. Syst. 1994, 23 (2), 267-291. (27) Wu, W.; Walczak, B.; Massart, D. L.; Prebble, K. A.; Last, I. R. Spectral transformation and wavelength selection in nearinfrared spectra classification. Anal. Chim. Acta 1995 315 (3), 243-255. (28) Kohonen, T. The self-organizing map. Neurocomputing 1998, 21 (1-3), 1-6. (29) Zupan, J.; Novic, M.; Ruisa´nchez, I. Kohonen and counterpropagation artificial neural networks in analytical chemistry. Chemom. Intell. Lab. Syst. 1997, 38 (1), 1-23. (30) Kohonen, T. Self-Organizing maps, 3rd ed.; Springer: Berlin, Germany, 2001. (31) Wythoff, B. J. Backpropagation neural networkssA tutorial. Chemom. Intell. Lab. Syst. 1993, 18 (2), 115-155. (32) Walczak, B. Neuralnetworks with robust backpropagation learning algorithm. Anal. Chim. Acta 1996, 322 (1-2), 21-29. (33) Goh, A. T. C. Backpropagation neural networks for modeling complex-systems. Artif. Intell. Eng. 1995, 9 (3), 143-151. (34) Zhang, L.; Subbarayan, G. An evaluation of back-propagation neural networks for the optimal design of structural systems: Part I. Training procedures. Comput. Methods Appl. Mech. Eng. 2002, 191 (25-26), 2873-2886. (35) Zhang, L.; Subbarayan, G. An evaluation of back-propagation neural networks for the optimal design of structural systems: Part II. Numerical evaluation. Comput. Methods Appl. Mech. Eng. 2002, 191 (25-26), 2887-2904. (36) Thomas, E. V. Non-parametric statistical methods for multivariate calibration model selection and comparison. J. Chemom. 2003, 17 (12), 653-659. (37) Wu, W.; Walczak, B.; Massart, D. L.; Heuerding, S.; Erni, F.; Last, I. R.; Prebble, K. A. Artificial neural networks in classification of NIR spectral data: Design of the training set. Chemom. Intell. Lab. Syst. 1996, 33 (1), 35-46. (38) Rajer-Kanduc, K.; Zupan, J.; Majcen, N. Separation of data on the training and test set for modelling: a case study for modelling of five colour properties of a white pigment. Chemom. Intell. Lab. Syst. 2003, 65 (2), 221-229. (39) Marengo, E.; Soave, C.; Gennaro, M. C.; Robotti, E.; Bobba, M.; Lenti, M. Comparison of different calibration methods for the determination by FT-NIR spectroscopy of the hydroxyl number in polyester resins. Ann. Chim. 2004, 94 (3), 219-228. (40) Marengo, E.; Bobba, M.; Robotti, E.; Lenti, M. Hydroxyl and acid number prediction in polyester resins by near infrared spectroscopy and artificial neural networks. Anal. Chim. Acta 2004, 511 (2), 313-322.

Received for review September 2, 2005. Revised manuscript received September 20, 2005. Accepted September 28, 2005. ES0517466