Development of a Neural Network Predictive Emission Monitoring System for Flue Gas

Measurement Sharifuddin M Zain#1, Kien Kek Chua*2

1 Chemistry Department, University Malaya, 50603 Kuala Lumpur, Malaysia. 2 PETRONAS Group Technical Services, Dayabumi, 50050 Kuala Lumpur, Malaysia

1smzain@um.edu.my 2chuakienkek@petronas.com.my

Abstract— Department of Environment in most countries is increasingly tightening clean Air regulation to mandate heavy industries to comply with stack emission limits. One of the latest measures is to enforce the installation of analytical instrumentation known as Continuous Emission Monitoring System (CEMS) to report emission level online to DOE office. CEMS being hardware based analyzer is expensive and maintenance intensive and often unreliable. Therefore, the need for more economical, reliable and accurate software-based predictive techniques is a feasible equivalent alternative for regulatory compliance. This study has successfully developed a neural network software-based Predictive Emissions Monitoring System (PEMS) to accurately determine stack emission level which can correlate closely with hardware analyzer measurement.

Keywords- Neural network, emission, analyzer

I. INTRODUCTION Furnaces and boilers in industries are constantly pursuing for more efficient and economical combustion control systems in order to comply with increasingly stringent environmental regulations and cost optimization. With the increasing use of fuels containing high CO2 as well as sulphur, there is a growing concern in the oil & gas industry over excessive release of emissions such as SO2, NO2 and CO2. Accurate measurement of these emissions is essential for better control to minimize possible pollution to our environment. Traditionally, the measurement of emission is by online analyzer or manual laboratory sampling analysis. However, due to some limitations of these methods such as hardware sensors fouling by the dirty product of combustion, sensor drift, technical constraint in obtaining representative sample and matrix interference, alternative technology such as predictive measurement using software algorithm to model the emission has been widely studied [1], [2]. By utilizing a more reliable predictive emission monitoring as an alternative to the conventional emissions monitoring, on-line prediction of the emissions can be achieved and control loop can become more responsive, leading to improved control system performance and ultimately reduced emissions.

In this study, a software algorithm has been developed utilizing an artificial neural network to predict combustion emissions. Digital signal processing techniques are applied to extract features from the combustion parameters to establish their relationships with the emission levels. As such relationships are very complex and non-linear, the incorporation of a neural network is the most suitable method. Neural networks have the ability to learn and store any non-linear, complex relationships between its inputs and outputs. To develop the neural network as part of the emission monitoring system, the features extracted from combustion parameters and the corresponding emissions data are used to train the neural network, which is then applied to predict emissions. To evaluate the effectiveness of the system, a set of raw data from incinerator was tested and the preliminary result correlate closely with the hardware analyzer readings.

II. METHODOLOGY

A. Data Acquisition system Data acquisition system forms the front end of the model where combustion data such as fuel/air ratio, pressure, temperature and emission are captured into the storage bank of the data acquisition system. The signal is then further conditioned and passed on to the next block of the software algorithm to extract useful features among the complex data gathered.

B. Features extraction Feature extraction provides a means to identify the emission unique characteristics of each variance of combustion, a process similar to finger-printing. Digital signal processing algorithms have been developed to extract these features from the complex signals in both time and frequency domains. The time-domain features consist of DC mean, AC variance, standard deviation, kurtosis, skewness while the frequency-domain features include entropy and shape factor of the power spectral density (PSD) distribution. These features together with the corresponding emissions data can be used to train a neural network to establish the quantitative relationship between the features and emission

2011 IEEE 7th International Colloquium on Signal Processing and its Applications

314978-1-61284-413-8/11/$26.00 ©2011 IEEE

levels. Once the network has been properly trained, it can be applied to predict emissions.

C. Artificial Neural Network This is the heart of the model. The neural network developed was a Multi-Layer Perceptrons (MLP) trained using back propagation algorithm. Previous research [3] has shown that MLP can accurately represent any continuous non-linear function. This type of network is hence believed to be most suitable for the emissions monitoring purpose where non-linear relationships are expected between the combustion properties and emissions. The architecture of the network consists of one input layer, two hidden layers and one output layer. Each hidden layer has five neurons configured with non-linear transfer function known as log-sigmoid (LOGSIG) and the output layer comprises one neuron with a pure linear (PURELIN) transfer function which gives a continuous output. The learning function selected for the network is the back propagation gradient descent with momentum weight and bias learning (LEARNGDM) and its training performance function uses the differentiable performance function of Mean Square Error (MSE). The network is trained in a supervised mode with a set of inputs and their targets provided for the network to establish the relationship between the inputs and the outputs.

III. SYSTEM SET UP The concept of applying neural network software sensor to monitor emission from an incinerator is shown in Figure 1. The incinerator consists of a box-type radiant section with a vertical up-shot burner. Flue gas flow entering the incinerator is monitored using a flow transmitter (FT). The combustion air is supplied with the help of a forced draft fan and the flow is measured by another flow transmitter (FT) . Emission gases (CO2, O2, CO, NO2) are measured by conventional analysers installed at the top outlet section of the incinerator. The combustion temperature is monitored using an online

Fig.1 Predictive emission monitoring system for flue gas at incinerator temperature sensing element (TT). All signals of the combustion parameters are connected to a plant Distributed Control System (DCS) where data is kept in the historian file of the system. The data in the DCS is then extracted by laptop computer utilizing software algorithm running in MATLAB environment. Digital signal processing techniques are used to extract features from the signals to yield the DC mean, skewness, variance, power spectrum density, shape factor, number of zero crossing and kurtosis. These features, which relate combustion parameters and emission, are used to train a neural network to predict emissions under similar combustion conditions.

IV. RESILTS AND DISCUSSION Combustion parameters with different air fuel ratio and analyzer readings are captured by the neural network to derive correlation pattern. Figure 2 and Figure 3 show the power spectrum density (PSD) plots of the combustion behavior at different air-fuel (A/F) ratio read by feature extraction algorithm of this software model. At low air-fuel ratio of 14, 16, 18, as shown in Figure 2, lower frequency (<100 Hz) components dominate the spectra but as the air-fuel ratio increases (Air fuel ratio of 28, 34, 44), as illustrated in Figure 3, spectrum distribution changes with increased amplitude (arbitrary unit) occurring at the higher frequency region between 350 and 800 Hz. The relationship data of spectrum pattern and air fuel ratio can be used by neural network to establish a correlation to the emission level given by online analyzer reading.

2011 IEEE 7th International Colloquium on Signal Processing and its Applications

315

Fig. 2 Spectrum distribution of combustion at lower air-fuel ratio

Fig. 3 Higher air-fuel ratio produces higher power signal spectrum

The emission trending of CO2 and O2 was further investigated under different air fuel ratio condition as shown in Figure 4. The result indicates a complex relationship, non-linear pattern as expected. This is the inherent challenge faced with combustion process which has wide dynamic and non-linear behavior. Hence manual interpretation and evaluation to establish any relationship between the combustion parameters with emissions can be very inefficient and inaccurate. The incorporation of neural network to handle such tasks is thus necessary as neural network has the ability to relate non-linear relationship between varying combustion parameters and emissions. To implement this, the signals features extracted from combustion parameters were used as the training data for the neural network to predict combustion emissions.

Fig. 4 Combustion emissions at various air-fuel ratios

Figure 5 shows the result of CO2 emission predicted by the neural network after it has been trained with the trained data. It can be seen that the neural network performs perfectly well where the predicted results correspond exactly to the actual measurement obtained by the hardware CO2 analyser.

Fig. 5 Predicted CO2 concentrations using training data

A regression analysis was performed to evaluate the performance of the network by comparing its predicted outputs against the expected outputs as indicated in Figure 6. The ideal results are indicated by a dashed line (45° line) and the best fitted line of the predicted data points (circles) is shown by a solid line. It can be seen that there is no deviation between the two lines illustrating the efficient performance of neural network in predicting known data.

Fig. 6 Prediction using trained data

Subsequent prediction was performed by the neural network on new unknown data and the results are shown in Figure 7. It can be seen that the emissions level predicted by the neural network is very close to the actual hardware analyzer measurement.

Expected output

Pred

icte

d ou

tput

Frequency (Hz)

PSD

(arb

.uni

ts)

x10-4

Frequency (Hz)

PSD

(arb

.uni

ts)

2011 IEEE 7th International Colloquium on Signal Processing and its Applications

316

Further evaluation of its performance is presented in Figure 8 by comparing its predicted outputs with the expected output. The result indicates that the emission prediction achieved by the neural network is consistent with the expected value with accuracy of prediction close to 98%. The overall results from this study have demonstrated that the neural network developed is capable of predicting unknown emissions data accurately and is a versatile tool in quantifying non-linear relationship between the neural input and output

Fig. 7 CO2 readings predicted using new unknown data

Fig. 8 Prediction using unknown new data

V. CONCLUSIONS A predictive emission monitoring system based on neural

network software algorithm has been successfully developed to predict flue gas emission from an incinerator with encouraging results. Preliminary results obtained by the neural network show good correlation with readings of the online hardware analyzers. The accuracy of the predicted emission is about 98%.

The neural networks developed play an important role in the establishment of the complex non-linear relationships between combustion parameters and emissions level. With further future work to enhance and widen the learning envelop, the system has the potential to predict real time emission online This would open up potential alternative

solution for measuring combustion emissions level which has been traditionally performed by expensive and maintenance intensive hardware analyzers.

ACKNOWLEDGMENT The authors wish to acknowledge staff and management

of University Malaya and PETRONAS for their support in completing this study.

REFERENCES [1] Soteris A. Kalogirou. J. Progress in energy and combustion science

29 (2003) 515-566 [2] Anker Jensen and Jan Erik Johnson. J. Chemical Engineering

Science, Vol. 52, No.11, pp. 1715-1731, 1997 [3] Cybenko, G. (1989) “ Approximations by superpositions of a

sigmoidal function”. Mathematics of control, signal and system, 2, 303-14.

Pred

icte

d ou

tput

Expected output

2011 IEEE 7th International Colloquium on Signal Processing and its Applications

317