Published by European Centre for Research Training and Development UK (

Transcription

1 PREDICTION AND MODELING OF SEASONAL CONCENTRATIONS OF AIR POLLUTANTS IN SEMI-URBAN REGION EMPLOYING ARTIFICIAL NEURAL NETWORK ENSEMBLES B. Yoror, T. G. Leton & J. N. Ugbebor Deartment of Civil and Environmental Engineering, Faculty of Engineering, University of Port Harcourt, Nigeria. ABSTRACT: This study utilizes Artificial Neural Networ (ANN) ensembles to redict seasonal variation of air ollutants in semi-urban region of Eleme, Rivers state, Nigeria. A ten year monthly concentrations of SO2, NO2, CO and CH4 in the region was obtained for dry and rainy seasons. Air ollutant concentrations in semi urban area of Eleme can be attributed mainly to industrial activities, vehicular emissions and some local bacground concentrations influenced by meteorological and geograhical conditions of the area. Training of the networ models was achieved using Neural NetTime Series feature of MATLAB software. Observed concentrations of ollutants and meteorological arameters were used as inut variables for the rognostic models. The develoed ANN rognostic models accurately catured the dynamic relationshis between ollutant concentrations and meteorological redictor variables. The relationshis between redicted and observed values were highly significant at 95% of confidence level for all models as dry and rainy seasons models gave R 2 greater than 0.99 (indicating close relationshis between observed and redicted values). CH4 showed R 2 of and for dry and rainy seasons resectively; CO showed R 2 of and for dry and rainy seasons resectively; NO2 showed R 2 of and for dry and rainy seasons resectively; SO2 showed R 2 of and for dry and rainy seasons resectively. The trend in redicted ollutants indicated that the study area is a major recetor of air ollutants emanating mainly from industrial activities and vehicular exhaust emissions. Further research study is needed to comare ANN model with other modeling aroaches such as with multile linear regression models for the rediction of air ollutants. KEYWORDS: Semi-Urban Region, Air Pollutants, Artificial Neural Networ, Inut layer, Hidden Layer, Outut Layer. INTRODUCTION In many urban and semi-urban areas concentrations of air ollutants are on the increase (Rao and Rao, 2005) mainly due to burning of fossil fuels for energy roduction for domestic, industrial, and transortation uses. Thus, air ollution resents a comlex issue that is driven by multile sources ranging from industrial emissions, vehicular emissions, and other emissions from fossil fuels, construction activities to domestic activities. Hence, air ollution is of ublic health and environmental concern (Davis and Cornwell, 2008), as it imacts widely on the biohysical environment. It is harmful to all forms of life, including lants, animals, and birds. 1

2 LITERATURE REVIEW Artificial Neural Networs (ANN) offer suitable aroach in modeling atmosheric ollutants than conventional deterministic modeling techniques (Elangasinghe et al., 2014). This is because Artificial Neural Networs can cature different inds of relationshis among variables, which otherwise may have been very difficult or imossible with statistical and deterministic models. According to Kuonen et al., (2003) artificial neural networ can be used to model multifactor, nonlinearity and uncertainty. In contrast to stochastic modeling aroach, ANN does not mae rior assumtions on the attern of data set (Ming et al., 2009). Artificial neural networ has been emloyed to a greater extent in the rediction of concentrations of air ollutants such as Nitrogen dioxide, Sulhur dioxide (Sudeshana et al., 2013), PM10 (Hooyberghs et al., 2005), Ozone (Elamel et al., 2001), and Carbon monoxide (Sudhir et al., 2013). Thus, Artificial Neural Networ can be regarded as an effective and intelligent modeling aroach which has received much attention in the field of environmental engineering (Sudeshana et al., 2013), and has been roven to be a owerful tool in Air Quality Modeling and rediction (Dorling et al., 2003; Primoz and Marija, 2011; Amirsasha and Farhad, 2012). Its role is increasing with imroving artificial neural networs algorithms that mae artificial neural networs alications on real systems even more accurate and reliable (Grivas and Chaloulaou, 2006; Singh et al.,, 2012; Elangasinghe et al., 2014). Recent studies have shown that artificial neural networs based air ollution models have better erformances than other deterministic and statistical techniques (Chelani et al., 2002; Grivas and Chaloulaou, 2006; Singh et al., 2012). Dorling et al., (2003) discusses several Air Quality Forecasting using artificial neural networs aroaches and concludes that Neural Networ models are useful for simulating air quality in an environment. MATERIALS AND METHODS Study Area Eleme region (Figure 1) is located within the coastal area of Rivers State in the Niger Delta region between Longitude E and Latitude N. The comlex coastline and lowlying flat toology of the area result in comosite surface wind seed atterns, esecially during low wind flows when land and sea breezes dominate the surface wind of the area. Figure 1: Ma of Eleme region 2

3 Collection of data A 10 year data (January 2006 to October 2015) was collected cororate monitoring stations in Eleme area and used in the study. Meteorological data acquired and adoted in the study are monthly ambient temerature, relative humidity, wind seed and direction, while monthly air ollutants arameters are carbon monoxide, CO, sulhur dioxide, SO2, Nitrogen dioxide, NO2, and hydrocarbon methane, CH4. A total of 944 data set was obtained which forms the basis of data analysis and model building resented in this study. In this study, significance imortance was laced on using a minimal set of meteorological arameters (redictors) that are readily measured or observed in the area to ensure that the models are of ractical use. Determination of Networ Architecture/Toology A three layer ercetron ANN was used as the base architecture or toology for the designed rediction models. The effectiveness of the multi-layer ercetron (MLP) to accurately redict non-linear systems and its ability to generalize is the main reason for its choice. The rediction models develoed are multi-layer ercetron with inut layer units, hidden layer units and an outut layer unit. The inut variables (original data including the target variables) were entered into the inut layer of the networ. The oututs of the inut layer units function as inut to the hidden layer units, and the oututs of the hidden layer units function as inut of the outut layer. The outut of the outut layer unit was obtained as the final outut of the networ, which is the ollutant concentration. The number of atterns and features were used to determine the networ toology or architecture. The three-layer networ architecture model designed for the rediction of ollutants concentrations consists of n neurons in the inut layer, m neurons in the hidden layer and q neurons in the outut layer. The inut features were determined using assembled collected data from 2006 to Both the inut and outut data sets form the database of atterns or ercetual structure of the model. The number of neurons in the hidden layer(s) was determined using Aaie s Information Criterion (AIC) rules (Equations 1 to 3) which validate the number of neurons in the hidden layer of the networ. The most efficient neural networ architectures that accurately redict ollutant concentrations in the area were chosen. The general AIC estimator of Kullbac Leibler information (Gaurang et al.,, 2010) is exressed as: AIC -2(maximum log lielihoodof the model) 2(number of free arametersof the model) AIC 2 ln( lielihood) 2 The best ANN model was determined by calculating the difference between lowest AIC model and the others (Gaurang et al.,, 2010) as: i AICi AIC min Equation (2) is used to calculate the lielihood of a model given the set data and to determine the validity of each model being the best aroximator (Gaurang et al.,, 2010). (1) (2) 1 Ex( i) 2 i 1 Ex( i 2 i) (3) 3

4 Model Develoment Process In develoing the rognostic models, the ten year data was categorized into dry season (November to March) and rainy season (Aril to October). The data set were further divided into training, validation and testing data sets. The training data set was used to modify and change the synatic weights of the ANN models. Testing data set was utilized eriodically to test the model s generalizing caabilities (Primoz and Marija, 2011) during the learning rocess so as to achieve best otimization during learning. The training and validation sets were organized into the learning set. The testing set was used for model verification to determine its ability to redict accurately and the exected associated error. During ANN models design rocess original data was randomly searated and arranged into training data set (70%), validation data set (15%) and test data set (15%) by Matlab software caabilities. The training data set used for each model contains the inut vector variables and their associated observed targets that constitute one class of the test set. Several numbers of such test classes were fed into the neural networ rograms. The design ANN models used selected transfer functions for secific algorithms based on their derivability and their non-linearity. Sigmoid (logistic) function was used as the transfer function between the inut layer units and the hidden layer units, and the hyerbolic tangent function was used between the hidden layer units and the outut layer units. The develoed ANN rognostic models (Figures 3 and 5) were trained and genetically otimized on the training data set. Testing data sets were tested on the models to determine their ability to redict from observed data. Data Normalization Before use, the inut data was initially normalized and made uniform to avoid overflows of the networ due to variations in synatic weights. Normalization also eliminates errors and missing data (which was 3% of the data set). In order to imrove networ sensitivity, the data set was normalized into the range (-1, 1) using Equation (4). After rediction the outut of the networ was transformed and de-normalized bac to the original values. X norm 1 2 X X max max X X i min (4) Where: X norm = normalized value, Xi = original value, Xmax = maximum value of Xi and Xmin = minimum value of Xi The outut of the networ must be de-normalized or transformed bac to the original values after rediction by maing Xi the subject of the above Equation. Networ Training and testing rocess Training of the networ models was achieved using a Nonlinear Autoregressive with External (Exogenous) Inut (NARX) of Neural NetTime Series feature of MATLAB. Alyuda Forecaster software (Alyuda NeuroIntelligence 2.2, htt:// was used as control software to establish the reliability of MATLAB software in training the networ models. Training and testing of the networ models were conducted after the networ toology was determined and the atterns reared. The goal of the training rocess was to obtain a desired outut given a set of inuts vectors and synatic weights. MATLAB Conjugate Gradient 4

5 Descent bacroagation algorithm training algorithm was used to train, validate and test the ANN models for dry and rain seasons rediction. The goal of the learning rocess was to find the minimum of the mean square error function. Care was taen during the training rocess to ensure that the networ reach the minimum error function and avoid local minima. The learning rate and momentum coefficient arameters were chosen to otimally control the seed of the networ during learning and weight udates so as to achieve best convergence. The networ was eriodically tested on the testing data set in order to revent overtraining. The networ that roduced the best otimization results was selected as the best model. Neural Networ Model Building Process Consider the three-layer networ of Figure 2 with inut, hidden and outut layers denoted as i, j, and, where i is the inut units index, j is the hidden unit index, and is the outut unit index. The McCulloch-Pitts model or activation function (Senguta, 2009) was adoted as given in Equation (5). A n j1 x w j j (5) Alying bias to the networ, Equation (5) becomes: V A b n j1 x j w j b (6) A general nonlinear sigmoid or logistic transfer function is adoted and alied: Y f V 1 (1 Ex( V )) (7) Where A is the activation function of the outut unit, b is the bias, x j is the inut of the outut unit, w j is the weight between the hidden units and the outut unit, Y is the outut, is the activation of outut unit (including weighted bias inut, ) and β is the sloe of the activation function (assumed to be 1). V b Xj Wj1 Bias, b Xj Xjn Wj2 Wjn Neuro n ΣWjxj Activation Function, A V f(x) Transfer Function Outut, Y Figure 2: Information Processing in ANN models design rocess 5

6 Total networ error is the sum total of all the errors of the outut layer neurons and is comuted as: E Where E is the total networ error, E E 2 1 T Y 2 is the error of the attern for the q th neuron, the target (observed) value of the attern of the q th neuron, and layer for the attern in the q th neuron. Y Y T is is the outut of the outut is determined by state of the outut neuron in the th unit and the weight, wj between the m th neuron in the hidden layer units and the q th neuron of the outut layer unit. Equation (8) was used to comute the total networ error, which is the sum of all the errors of the outut layer neurons. The synatic weights of the hidden layer units on the outut unit are comuted by alying the Delta rule as follows: wj T Y Where Y jm is the outut of the hidden layer units, is the atterns index. The synatic weight on the outut unit was comuted from Equation (9). The weights were thus adjusted in order to minimize the mean square error of the networ. The weights adjustment is done using the method of gradient descent (Senguta, 2009). How much error deends on the change in synatic weights is comuted as: dy dv Y jm (8) (9) de dw j de dy dy dw j T Y Y ( 1Y ) Y jm (10) The weight Δwj is adjusted and udated using Equation (11) given as: j j T Y Y ( 1 Y Yjm de de dy w j ) dw dy dw (11) The increment in weight udate wj for the three layer networs is given as: w j de dw j de dy dy dv dv dy jm dy dv jm jm dv dw jm j T Y. Y(1 Y). w j 1 Y jm Y jm. Y in w j (12) 6

7 Where Y jm in the inut layer; units resectively; is the outut of the m th neuron in the hidden layer, V jm and w j Y in is the outut of the n th neuron V are the activation signals of the hidden layer and outut layer is the momentum term, τ is the momentum coefficient introduced to reduce raid fluctuations in weights adjustments and avert oscillation during networ training rocess. η is the training control arameter called learning rate that seeds u convergence while reducing networ errors. Equations (9), (10) and (11) were used to udate the synatic weights of the ANN models. Equation (12) was used to train the ANN redictive models. The develoed ANN rognostic models (Figures 3 and 5) were trained and genetically otimized on the training data set. Testing data sets were tested on the models to determine their ability to redict from observed data. The training mean square error is shown in Figure 4. Model Verification In order to determine the models ability to accurately redict future values with minimum networ error, the models were trained, validated and tested on the test data. To obtain a good judgment of the model s ability to redict accurately, a new data set was queried on the models. Sensitivity Analysis To achieve the aim of building simle forecasting models for the rediction of seasonal atmosheric ollutants, the best subset of inut redictor variables were selected in the ANN model building rocess. Forward Stewise, bacward stewise elimination and genetic otimization standard inut otimization techniques (Elangasinghe et al., 2014) were alied for the otimization of inut vectors. Year and Month were considered negligible inuts by Forward Stewise, while bacward stewise eliminated only RH to give the best networ. All the inut variables were considered as significant by genetic otimization in maing the best redations. Sensitivity analysis was erformed to determine the resonse of the ANN rognostic models to each inut vector. Each redictor variable was varied while eeing others constant. This analysis was erformed to test the sensitivity of the ANN redicted models to each redictor arameter and to determine the non-linear relationshi between each redictor variable and the modeled ollutant concentrations. RESULTS AND DISCUSSIONS Results and Findings Dry season redictive model was found to erform best with Conjugate Gradient Descent algorithm and architecture of (Figure 3) with learning control arameter η of 0.1 and momentum coefficient of While rainy season redictive model erformed best with Conjugate Gradient Descent algorithm and architecture of (Figure 5) with learning control arameter η of 0.2 and momentum coefficient of 0.8. For most of the time, the redicted values were very close to the observed ollutant concentrations both in trend and attern. It was found that all the develoed ANN rognostic models erform absolutely well with rediction accuracy of more than 95%. 7

8 Observed concentration Figure 3: Dry Season Design Architecture (7-5-1) Figure 4: Training Mean Square Error 8

9 Observed concentration Figure 5: Rainy Season Design Architecture (7-9-1) DISCUSSION Details of ANN design (training/learning) arameters, architectures and networ outut reorts for forecasting of ollutant concentrations in the study area are shown in Table 1. Mean Absolute Errors (MAE), Mean Absolute Related Errors (MARE) and Mean Square Errors (MSE) for each forecasted ollutant are given in Table 1. Time series of redicted versus observed ollutant concentrations for both dry and rainy seasons are shown Figures 6, 7, 9, 10, 12, 13, 15 and 16. The redicted ollutant levels were comared with the measured values to determine the accuracy of the rognostic models in redicting concentrations of air ollutants in the study area. Predicted results comared considerably with observed values with high degree of accuracy. The best fit lines (Figures 8, 11, 14, 17 and 18,) generated between 9

10 CH4 (µg/m3) CH4 (µg/m3) International Journal of Environment and Pollution Research observed and redicted values give correlation coefficients and R 2 very close to 1 for both dry and rainy seasons (Table 1). These coefficients of determinations are accetable as values close to 1.0 are considered to be best fits. Statistically, modeling results further indicated that the relationshis between observed and redicted values are highly significant at 95% of confidence level. Hence, the ANN models roerly trained with ast ollutants data accurately redict air ollutant concentrations in the area. The trend in redicted ollutants indicated that the study area is a major recetor of air ollutants emanating mainly from industrial activities and vehicular exhaust emissions. 400 Observed values Predicted values Number of data set Figure 6: Dry Season Observed versus Predicted CH4 400 Observed values Number of data set Figure 7: Rainy Season Observed versus Predicted CH4 10

11 CO (µg/m 3 ) Predicted CH 4 (µg/m 3 ) International Journal of Environment and Pollution Research Dry season, R² = Rainy season, R² = Dry season Rainy season Observed CH 4 (µg/m 3 ) Figure 8: Dry and Rainy Seasons Best fits of Observed and Predicted CH4 200 Observed values Predicted values Number of data set Figure 9: Dry Season Observed versus Predicted CO 11

12 CO (µg/m Observed values Predicted values Numcer of data set Figure 10: Rainy Season Observed versus Predicted CO Predicted CO (µg/m Dry season, R² = Observed CO (µg/m 3 Rainy season, R² = Dry season Rainy season Figure 11: Dry and Rainy seasons Best fits of Observed and Predicted CO 12

13 Predicted NO 2 (µg/m 3 ) NO 2 (µg/m 3 ) NO 2 (µg/m Observed values Predicted values Number of data set Figure 12: Dry Season Observed versus Predicted NO2 200 Observed values Number of data set Figure 13: Rainy Season Observed versus Predicted NO Dry season, R² = Rainy season, R² = Observed NO 2 (µg/m 3 ) Dry season Rainy season Figure 14: Dry and Rainy Seasons Best fits of Observed and Predicted NO2 13

14 Predicted SO 2 (µg/m 3 ) SO2 (µg/m3) SO 2 (µg/m 3 ) International Journal of Environment and Pollution Research 100 Observed values Predicted values Number of data set Figure 15: Dry Season Observed versus Predicted SO2 100 Observed values Predicted values Number of data set Figure 16: Rainy Season Observed versus Predicted SO Dry season, R² = Rainy season, R² = Observed SO 2 (µg/m 3 Dry season Rainy season Figure 17: Dry and Rainy Seasons Best fits of Observed and Predicted SO2 14

15 Figure 18: Goodness of fits of seasonal ollutants during networ testing a. Table 1: Coefficient of Determination and Correlation Coefficient for observed and redicted ollutant concentrations DRY SEASON ANN Toology Pollutant Correlation Coefficient, R Coefficient of Determination, R 2 MAE (µg/m 3 ) MARE (µg/m 3 ) MSE (µg/m 3 ) CH x10-8. CO x10-10 NO x10-8 SO RAINY SEASON ANN Toology CH x

16 CO x 10-9 NO x10-7 SO x10-7 CONCLUSION Figure 19: Networ error outut resonse during networ testing The relationshis between observed and redicted values are highly significant at 95% of confidence level for all models. The best fit lines generated between observed and redicted values give correlation coefficient of determination very close to 1 (R 2 greater than 0.99) for dry season and rainy season models. Results indicated that the develoed ANN rognostic models accurately catured the non-linear relationshis between ollutant concentrations and meteorological redictor variables that exist in the area. Study indicated that air ollutant concentrations in Eleme region have nonlinear comosite relationshi with industrial, vehicular exhaust emissions, and meteorological factors. All of which were used as inut variables to the neural networs to redict seasonal concentrations of air ollutants in the region. Prediction results also showed that wind seeds and directions are the arameters that mostly influence ollutants concentrations in the ambient air of the study area. Significance of the study to research and ractice (contribution to nowledge) The study which was mainly emirical in nature revealed that ANN model when roerly trained with historical air quality and meteorological data can accurately forecast air ollutant concentrations in an area. In addition, the study exlored the alication of Artificial Neural Networs as a redictive tool to effectively model the dynamic nonlinear relationshis between air ollutants and meteorological variables in an area. The study further rovided effective techniques for the rediction of air ollutants concentrations in the area, thus serves as useful tools in environmental imact assessment (EIA) studies for the rediction of air quality imacts of a new develoment in the study area. 16

17 RECOMMENDATION Further research study is needed to comare ANN model with other modeling aroaches such as with multile regression models for the rediction of air ollutants in the study area. Acnowledgement The authors wish to acnowledge the Management of Environmental and Chemical services Limited for roviding the meteorological and air ollutants data used in this study. REFERENCES Alyuda NeuroIntelligence 2.2, htt:// Amirsasha B. and Farhad N., (2012); Artificial Neural Networs: A Non-Linear Tool for Air Quality Modeling and Monitoring. International Conference on Alied Life Sciences (ICALS2012). Chelani, A.B., Rao, C.V.C., Phade, K.M., Hasan, M.Z., Prediction of sulhur dioxide concentration using artificial neural networs. Environmental Modelling & Software 17, Davis M. L. and Cornwell D. A., (2008): Introduction to Environmental Engineering. McGraw Hill Comanies, Inc; New Yor, 4ed. Dorling S., Gavin C., Kuonen J., and Ari K., (2003); Air Quality Forecasting Using Neural Networ Aroaches. Air Quality Forecasting Meeting, Culham, UEA ENV. Elangasinghe M. A., Singhal N., Dirs K. N., Salmond J. A. (2014); Develoment of an ANN based air ollution forecasting system with exlicit nowledge through sensitivity analysis. Atmosheric Pollution Research 5 (2014) Elamel, A., Abdul-Wahab, S., Bouhamra, W., Aler, E., Measurement and rediction of ozone levels around a heavily industrialized area: a neural networ aroach. Advances in Environmental Research 5, Gaurang P., Amit G., Kosta Y. P. and Devyani P. (2010); Searching Most Efficient Neural Networ Architecture Using Aaie s Information Criterion (AIC). International Journal of Comuter Alications ( ) Volume 1 No. 5. Grivas, G., Chaloulaou, A., Artificial neural networ models for rediction of PM10 hourly concentrations, in the greater area of Athens, Greece. Atmosheric Environment 40, Hooyberghs, J., Mensin, C., Dumont, G., Fierens, F., Brasseur, O., A neural networ forecast for daily average PM10 concentrations in Belgium. Atmosheric Environment 39, Kuonen, J., Partanen, L., Karinen, A., Ruusanen, J., Junninen, H., Kolehmainen, M., Gwynne, M., (2003); Extensive evaluation of neural networ models for the rediction of NO2and PM concentrations, comared with a deterministic modelling system and measurements in central Helsini. Atmosheric Environment 37, Ming C., Yafeng Y. Min X. (2009): Prediction of hourly air ollutant concentrations near urban arterials using artificial neural networ aroach Transortation Research Part D Elsevier. Transortation Research Part D 14 (2009) 32 41, j. Primoz M. and Marija Z. B. (2011). Artificial Neural Networs - a Useful Tool in Air Pollution and Meteorological Modelling, Advanced Air Pollution, Dr. Farhad Nejadoori (Ed.), 17

18 ISBN: , InTech, Available from: htt:// Senguta S., (2009). Lecture Series on Neural Networs and Alications. Deartment of Electronics and Electrical, Communication Engineering, Indian Institute of Technology Kharagur Centre for Educational Technology, I. I. T. Kharagur. htt:// Singh, K.P., Guta, S., Kumar, A., Shula, S.P., Linear and nonlinear modeling aroaches for urban air quality rediction. Science of the Total Environment 426, Sudeshana P, Ala S., Asho.K.S., and Srivastava J. K. (2013); MODELING OF AMBIENT FOR SOx AND NOx POLLUTANTS THROUGH ARTIFICAL NEURAL NETWORK IN SENSITIVE AREA OF UJJAIN CITY. J. International Journal of Chemical Sciences and Alications. Deartment of Chemical Engineering, Ujjain Engineering College, Ujjain, (M.P), India. Sudhir N., Rashmi N. Sangeeta K. (2013); FORECASTING CARBON MONOXIDE CONCENTRATION USING ARTIFICIAL NEURAL NETWORK MODELING. International Journal of Comuter Alications ( ) International Conference on Current Trends in Advanced Comuting ICCTAC