Prediction of Sulfur Dioxide (SO2) Concentration Levels from the Mina Al- Fahal Refinery in Oman Using Artificial Neural Networks

In this investigation, two Artificial Neural Network (ANN) models were applied for predicting ground-level sulfur dioxide (SO2) in the Sultanate of Oman in order to provide an early warning advisory for the protection of public health. The objective of the first model (Model I) was to use ANN to predict sulfur dioxide (SO2) levels at certain receptors from the Mina Al-Fahal refinery in Oman. The artificial neural network was also used for predicting the first 3 maximum SO2 concentrations and their corresponding locations with respect to the refinery (Model II). The models were used to determine meteorological conditions that most affect SO2 concentrations. In assessing this aspect, five meteorological parameters that are expected to affect the SO2 concentrations were explored. They include wind speed, atmospheric stability class, wind direction, mixing height, and ambient temperature. The developed models showed good predictive success with, R-squared values above 0.96 indicating high accuracy for both the models development and generalization capability. The meteorological variables with the greatest influence on SO2 concentrations were also identified. It was found that wind direction was the variable most important to Model I while wind direction, stability, and wind speed were the highest contributing variables in Model II. The investigation indicated that the ANN models were well-suited for modelling SO2 levels. Additionally, the ANN models can be extended for other applications in which non-linear relationships are observed.


INTRODUCTION
Sulfur dioxide (SO 2 ) is considered one of the indicators of air quality. It is formed primarily from the combustion of sulphur-containing fuels and can affect the health of people and have a negative impact on the environment. Modelling of environmental parameters is a basis for a better understanding and prevention of air pollution.
However, modelling SO 2 is a complex task which has drawn the attention of many scientists all over the world since the early 1960s. The literature on this subject showed that there was no universal approach for modelling SO 2 and, hence, several modelling approaches were proposed. One approach is to use the atmospheric diffusion model to predict future pollutant concentrations [1] . This approach simulates the airflow pattern and pollutant concentration by solving a highly coupled, non-linear, partial differential equation set. Such an approach demands huge computing costs which sometimes cause difficulties in computational convergence, especially for treating large space cases, and experimental validation, which is even more expensive and difficult to achieve due to the scaling inconsistency [2] .
A second approach in facilitating the prediction of pollutant concentrations is by developing statistical models. In recent years, the general trend was to use more statistical methods instead of traditional deterministic modelling. Statistical models were used to determine the underlying relationship between a set of input variables and the targets [3][4][5] . These models assume that the relationship between the variables is statistical in nature. Such models require information about the distribution of data which is generally not known previously [6] . The open literature showed that statistical approaches are frequently considered for short-term forecasting applied to real-time control of emissions or to air quality assessment [7,8] . These methods have some advantages over deterministic approaches. First, they do not need data about emissions as they are based only on the air quality and meteorological measurements. Second, they are often simpler in their structure than deterministic models, and they can more easily be implemented and used by nonexperts. However, the statistical models are not portable from site to site since they are developed and calibrated on local data.
By their unique structure, artificial neural networks possess the ability to learn non-linear relationships with limited prior knowledge about the process structure [18,19] . These models provide a better alternative than regression statistical models because of their computational efficiency. Additionally, the ANN methods can also be used in combination with traditional deterministic modelling techniques [20] . The capabilities and disadvantages of neural networks were described by Boznar and Mlakar [21] for the cases of wind forecasting, reconstruction, and SO 2 pollution forecasting.
Many researchers in the literature have used the ANN approach for tropospheric ozone forecasting. The most widespread ANN design is a Multilayer Perceptron (MLP) with a learning procedure based on the backpropagation algorithm. Abdul-Wahab and Al-Alawi used backpropagation MLP models to test the relationship between ozone and other variables. Their investigation indicated the potential of the MLP approach for capturing non-linear interactions. On the other hand, there are limited applications of ANN models for SO 2 prediction that are cited in the literature. Boznar et al. [9] applied a neural network model to predict SO 2 concentrations at Sostanj. The used inputs included historical data on observed SO 2 concentrations and meteorological parameters such as wind speed and direction.
A feedforward network with a backpropagation learning algorithm was applied. The major finding was that ANN models can be applied to predict ground-level concentrations of pollutants in complex terrains. However, the results were not compared with those obtained using any other modelling techniques. For the same area, Mlakar and Boznar [22] used an artificial neural network as a method of short-term air pollution prediction. The method was applied to the complex orography around the coal-fired thermal power plant at Sostanj. Gardner and Dorling [12] presented a review on the application of artificial neural networks in the atmospheric sciences. The applications of multilayer perceptions in atmospheric sciences were discussed in detail. Gardner and Dorling [12] concluded that ANN generally give as good or better results compared with the statistical linear method, especially where the problem being analyzed included non-linear behaviour [23] .
Chelani et al. [24] predicted SO 2 concentration by using an ANN, and the predicted values were compared with the measured concentrations at three sites in Delhi. A multivariate regression model was also used for comparison with the results obtained by using the neural network model. The study results indicated that the neural network was able to give better predictions with less residual mean square error than those given by multivariate regression models.
Nunnari et al. [25] inter-compared several statistical techniques for modelling SO 2 concentration at a point by using neural networks, fuzzy logic, generalized additive techniques, and other recently proposed statistical approaches. They found that artificial neural network-based models and neuro-fuzzy models were the most promising. The results of the modelling inter-comparison was planned to be used to give guidelines for designing a warning system for air quality assessment.
The control of SO 2 and the associated prediction of its levels are needed to take the preventive actions during the episodes of high SO 2 concentrations. In this paper, an artificial neural network was used for predicting SO 2 levels at certain receptors from the Mina Al-Fahal refinery in the Sultanate of Oman (Model I). The particular aim was to relate the SO 2 concentrations to meteorological parameters.
In this study, five meteorological parameters that are expected to affect the SO 2 concentrations were investigated. These include wind speed, atmospheric stability class, wind direction, mixing height, and ambient temperature. The artificial neural network was also used for predicting the first 3 maximum SO 2 concentrations and their corresponding locations with respect to the refinery (Model II).

Area's description: The Mina Al-Fahal refinery in
Oman is used as a case study in this paper. The refinery, constructed in 1982, is situated on a plot area of 170,000 m 2 located in the industrial hub of Mina Al-Fahal. The location of the site was selected based on its proximity to both the natural resources supply and the main consumers. The refinery was originally built to process 50,000 BPD of Oman's crude oil to meet the domestic requirements of petroleum products. However, the capacity was enhanced to 80,000 BPD in 1987 through an extensive revamp to meet the growing demand of petroleum products in the country.
The refinery complex, as a whole, contains the refinery itself, its tank farm, containing 24 vertical storage tanks, wastewater treatment plant, and gas turbine electricity generating sets. The refinery's offices and laboratory are located approximately 400 m to the east of the refinery. The location of the surrounding residential areas in the vicinity of the Mina Al-Fahal area is shown in Fig. 1. Residential areas to the west and south of the refinery are within 500 m of the site boundary and therefore could be at risk from atmospheric emissions. More details about the site can be found elsewhere [19] .
The air emissions in the refinery are mainly due to the quantities of fuel gas burned which results in the emission of various pollutants to the atmosphere, including SO 2 . The main sources for SO 2 generation in the study area included five gas turbine generators, two steam generators, six heaters, three reboilers and one flare.
In all these systems, gaseous fuels were combusted and the resulting flue gases that contain various pollutants were released into the atmosphere through a total of 17 stacks. The composition of the fuel gas used in the combustion contained 78.9 % H 2 , 3.7 % CH 4 , 5.1 % C 2 H 6 , 7.9 % C 3 H 8 , 4.1 % C 4 H 10 , 0.1 % C 5 H 12 , and 0.2 % H 2 S. In the combustion, air was used as the source of oxygen. When this fuel is burned, carbon in the fuel reacted to form CO and/or CO 2 , hydrogen formed H 2 O, and sulphur formed SO 2 . At temperatures greater than approximately 1800 ºC, some of the nitrogen in the air reacted to form oxides of nitrogen. The emission rate of SO 2 that resulted from the combustion from each stack was calculated by using material balances. Detailed information about the stacks and their SO 2 emission rates can be found in Abdul-Wahab et al. [19] .

Development of ANN models:
It was decided that two ANN models would be developed for this work. Table  1 shows the input and output for each model. The first model assessed the factors affecting the SO 2 concentrations at specific selected receptors. The second ANN model was developed for predicting the first 3 maximum SO 2 concentrations and their corresponding locations.  The first and the most critical step in developing an effective ANN model is input and output definition and data preparation. This includes identifying variables of interest, gathering the relevant data and inspecting them for possible errors, missing values, and outliers; hence, the data for developing the two SO 2 models was inspected carefully since data accuracy is vital for the development of efficient models. The Feed Forward network using the Back-Propagation (BP) algorithm was used to develop these two models. The BP algorithm uses the supervised training technique. In this technique, the interlayer connection weights and the processing elements' thresholds are first initialized to small random values. The network is then presented with a set of training patterns, each consisting of an example of the problem to be solved (the input) and the desired solution to this problem (the output).
These training patterns are presented repeatedly to the ANN model and the error between actual and predicted results is calculated. Weights are then adjusted by small amounts that are dictated by the General Delta Rule [26] . This adjustment is performed after each completed iteration whenever the network's computed output is different from the desired output. This process continues until weights converge to the desired error level or the output reaches an acceptable level.

RESULTS AND DISCUSSION
Prior to conducting the network's training operation using the back propagation paradigm, training sets of 9216 cases were obtained from the data set. This data set covers different situations that could possibly take place. For the purpose of model testing and validation, 15% were extracted from the given data set. Therefore, the training set consisted of 7834 cases while the testing set consisted of 1382 cases. As shown in Fig. 2, the ANN Model I used in this work consisted of 7 input nodes representing wind direction, wind speed, temperature, stability, mixing height, X and Y. The output consisted of one node representing the SO 2 concentration. The training process was performed using the NeuroShell® simulator.
After several adjustments to the network parameters, the network converged to a threshold of 0.00001 using 3 hidden nodes. The trained model prediction was in good agreement with the actual results. The R 2 value for the developed model was 0.9693 (Table 2). This indicates that approximately 96.9% of the variation in the SO 2 concentration could be explained by the selected input variables and the data used for model development.
Having trained the network successfully, the next step is to test the network in order to judge its performance and to determine whether the predicted results confirm with the actual results. The trained model is assumed to be successful if the model gives good results for the test set. Using the 1382 cases allocated for the testing set, the model-input parameters were entered consecutively for each case and a prediction for SO 2 concentration values was obtained. The results were then compared with the actual results for these cases. The statistical analysis of these results shown in Table 2 indicates that the R 2 value for the testing set was 0.9666, the mean squared error was 8.902 and the mean absolute error was 1.173. The ANN predicted and actual testing results are also shown in This high generalization capability indicates that the ANN model developed in this work can be used to model and predict the relationship between SO 2 concentration and the given input variables.
Model II, on the other hand, consists of 5 input nodes representing wind direction, wind speed, temperature, stability, and mixing height and 9 output nodes X1,Y1, Maximum 1 of SO 2 concentration, X2,Y2, Maximum 2 of SO 2 concentration, X3,Y3, Maximum 3 of SO 2 concentration, as illustrated in Fig.  4. The training set for ANN Model II consisted also of 7834 cases while the testing set consisted of 1382 cases. Similarly, the training process was performed using the NeuroShell® simulator. The trained model prediction was in good agreement with the actual results.
The R 2 value for the training and testing sets for the output variables X1,Y1, Maximum 1 of SO 2 concentration, X2,Y2, Maximum 2 of SO 2 concentration, X3,Y3, and Maximum 3 of SO 2 concentration is shown in Table 3. All R-squared values were above 0.996, indicating high accuracy for both model development and model generalization capability. Fig. 5 illustrates the comparison of predicted and actual maximum SO 2 concentration levels.
To find the percent contribution of each of the input variables with respect to the output variables, the partitioning method of the connection weights proposed by Garson [27] was used. The method involves partitioning the hidden-output connection weights of each hidden neuron into components associated with each input neuron [28,29] . It should be noted that there are several methods for calculating artificial neural networks' output contributions. Some of these methods are (1) Partial Derivatives or 'PaD'; (2) the "weights method or the partitioning method used in the current work; (3) the "perturb" method; (4) the "Profile" method; (5) the "Classical Stepwise"; (6) the "improved stepwise a"; (7) the "improved stepwise b". A summarized review and comparison of methods to study the contribution of variables in artificial neural network models can be found in Gevrey et al. [30] . Gevrey et al. [30] presented a simplification of the partitioning algorithm as indicated below: • For each hidden neuron h, divide the absolute value of the input-hidden layer connection weight by the sum of the absolute value of the input-hidden layer connection weight of all input neurons, i.e.
For h=1 to nh, • For each input neuron i, divide the sum of the Q ih for each hidden neuron by the sum for each hidden neuron of the sum for each input neuron of Q ih , multiply by 100. The relative importance of all output weights attributable to the given input variable is then obtained. Using this method, for Model I, it was found that the highest contributing variable that affects the SO 2 concentration level is the wind direction; while in Model II, wind direction, stability and wind speed are the highest contributors. The relative importance of the different input variables in the prediction of SO 2 concentration levels are shown in Fig. 6 for Model I, and in Table 4 for Model II.

CONCLUSION
Interest in neural networks and other artificial intelligence techniques is on the increase due to their high potential. Intelligent systems make it possible to obtain expert knowledge and to discover new knowledge to improve data analysis and the decision making process, hence improving the quality of life.
This paper examines the potential of ANN models as a predictive tool for predicting SO 2 concentrations as a function of meteorological conditions. The objective was to use an ANN model to predict SO 2 levels at certain receptors in the Mina Al-Fahal refinery in the Sultanate of Oman (Model I). The results offer an insight into the dependence of SO 2 concentrations on meteorological parameters. Five meteorological parameters that are expected to affect the SO 2 concentrations were investigated. These include wind speed, atmospheric stability class, wind direction, mixing height, and ambient temperature. The artificial neural network was also used for predicting the first 3 maximum SO 2 concentrations and their corresponding locations with respect to the refinery (Model II). The data fed to the neural network was divided into two sets: a training set and a testing set.
It was found that the neural network model consistently gives superior predictions and the results produced were encouraging. The R-squared values were found to be high. The R-squared for the first model was of the order of 97 % and they were of the order of 99.9 % for the second model. The relative importance of the various meteorological variables was also investigated. Clearly, this study indicated the potential of the neural network approach for capturing the non-linear interactions between SO 2 levels and meteorological variables and for the identification of the relative importance of these variables.