Modelling urban air quality using artificial neural network

Abstract

This paper describes the development of artificial neural network-based vehicular exhaust emission models for predicting 8-h average carbon monoxide concentrations at two air quality control regions (AQCRs) in the city of Delhi, India, viz. a typical traffic intersection (AQCR1) and a typical arterial road (AQCR2). Maximum of ten meteorological and six traffic characteristic variables have been used in the models’ formulation. Three scenarios were considered—considering both meteorological and traffic characteristics input parameters; only meteorological inputs; and only traffic characteristics input data. The performance of all the developed models was evaluated on the basis of index of agreement (d) and other statistical parameters, viz. the mean and the deviations of the observed and predicted concentrations, mean bias error, mean square error, systematic and unsystematic root mean square error, coefficient of determination and linear best fit constant and gradient (Willmott in B Am Meteorol Soc 63:1309, 1982). The forecast performance of the developed models, with meteorological and traffic characteristics (d=0.78 for AQCR1 and d=0.69 for AQCR2) and with only meteorological inputs (d=0.77 for AQCR1 and d=0.67 for AQCR2), were comparable with the measured data.

Introduction

In the recent past, vehicular air pollution has emerged as a major factor contributing to the quality of living in densely populated urban areas. Episodes of air pollution in urban centers are associated with sudden occurrence of high concentration of vehicular exhaust emissions (VEEs), which are generally governed by the local meteorology and dispersion mechanism (Nagendra and Khare 2002a). To assess the seriousness of the VEE threat and to take preventive measures, short-term forecasting of air quality is needed (Kolehmainen et al. 2001; Zannetti 1990). Line source models play an important role in local air quality management by providing the forecast of current and future VEE concentrations at the AQCRs. The traffic characteristic data may be used to forecast future changes based upon a range of ‘what if’ scenarios (Nagendra and Khare 2002a; Beaton et al. 1972). Deterministic and statistical-based approaches are frequently used in line source modelling. Since the relationship of VEE with the meteorology and traffic characteristic data is highly nonlinear, both deterministic and statistical models under perform in predicting the air quality (Nagendra and Khare 2002a). In the recent past, the artificial neural network (ANN) technique has drawn attention of researchers in its application to model VEEs (Nagendra and Khare 2002b). The neural network modelling approach offers several advantages over traditional ‘phenomenological’ or ‘semi-empirical’ models. It exhibits rapid information processing and is able to develop the mapping of input and output variables (Bose and Liang 1998). The most convincing advantage is that the accuracy of the neural network prediction is generally higher than that of the other kind of models (Viotti et al. 2002). In addition, neural networks deal with the non-linearity, handles noisy or missing data, create their own relationship amongst information, work with large number of variables (parameters) and provide general solutions with good predictive accuracy (Gardner and Dorling 1998).

Neural networks have found extensive applications in recent years for information processing e.g. voice recognition, hand-written character recognition and image processing (Rumelhart and McClelland 1996; Schalkoff 1992). Recently, a number of researchers have applied the neural network modelling techniques to forecast VEEs in the ambient air (Nagendra and Khare 2002a, b). Moseholm et al. (1996) studied the usefulness of neural network in understanding the relationships between traffic parameters and CO concentrations measured near an intersection, which was sheltered from wind by multi-storied buildings. In another work, Dorzdowicz et al. (1997) developed a line source neural network model for estimating hourly mean concentrations of CO in the urban area of Rosario city. The results indicated that neural network was accurately predicting the CO concentration comparable with observed data. Gardner and Dorling (1999) developed multilayer perceptron neural network models using hourly NO_x and NO₂ and meteorological data for the Central London. The results showed better performance of multilayer perceptron as compared to the regression models for the same location developed by Shi and Harrison (1997). Perez and Trier (2001) developed a multilayer neural network-based model to predict NO and NO₂ concentrations at a traffic junction in Santiago, Chile. The model performed much better than their previously developed persistence and regression models. Viotti et al. (2002) developed ANN-based short- and long-term air quality models for forecasting vehicular air pollutant concentrations in the city of Perugia, Italy. The models predicted the concentrations with reasonable accuracy. Lin (2002) has applied neural network technique to predict particulate matter (PM₁₀) concentrations and atmospheric visibility in Kaohsiung and Ping-Dong city, Taiwan. The inputs to the neural network model are PM₁₀ concentration; atmospheric pressure, surface radiation, relative humidity, atmospheric temperature and cloud cover conditions. The model results have shown high correlation with measured PM₁₀ concentration and reasonable correlations with atmospheric visibility. Kukkonen et al. (2003) have evaluated five neural network models, a linear stochastic model and a deterministic modelling system for predicting hourly NO₂ and PM₁₀ concentrations at two stations in central Helsinki, Finland. The results have indicated better performance of neural network models than other two models. In another study, Dorling et al. (2003) have developed neural network-based air quality forecasting system for the management of air quality during episodic conditions at selected cities in Finland, Germany, Czech Republic and UK. Lu et al. (2003) have developed the neural network model using principal component analysis and radial basis function to predict the respirable particulate matter, NO_x and NO₂ concentrations near roadside in Hong Kong. The comparison between the predicted and measured pollutant concentrations has indicated better performance of the neural network models. Recently, Nagendra and Khare (2004) developed ANN-based line source models for predicting hourly CO concentrations on an urban roadway. The results have shown that the neural network models are able to capture traffic ‘wake’ effects on the CO dispersion in the near field regions of the roadway.

The present paper discusses the development and performance evaluation of ANN-based VEE models for predicting 8-h average CO concentrations at two AQCRs, one representing a traffic intersection (AQCR1) and other, an arterial road (AQCR2), in the city of Delhi, India.

Study area

Figure 1 shows the study area along with air quality sampling stations being monitored by Central Pollution Control Board (CPCB), New Delhi. AQCR1 is located adjacent to kerb side of highly trafficked road—Bahadur Shah Zafar marg, which is one of the roads crossing Income Tax Office (ITO) intersection. This region comes under the central business district housing a number of government office buildings along with reputed educational institutes. The region is considered to be an area of most intense human activity and also tagged for the ‘worst air quality’ in the city. The AQCR2—Sirifort monitoring station, is located in the south of central Delhi. The surrounding area is comprised of dense residential localities, commercial establishments, institutional areas and famous sport complex. This station is situated near moderately busy traffic road, Khelgaon marg.

Fig. 1

Details of AQCRs in Delhi city

Data

The eight hourly CO concentration data have been collected from CPCB, New Delhi for a period of 3 years from January 1997 to December 1999, for both the AQCRs. The meteorological data including 8-h average observations of cloud cover, pressure, mixing height, sunshine hours, visibility, temperature, wind speed, wind direction and humidity have been collected from Indian Meteorological Department, New Delhi. Pasquill-Gifford stability scheme has been used to determine hourly stability categories (Schnell and Dey 2000). The eight hourly average traffic characteristics data have been collected from Central Road Research Institute, New Delhi for the respective AQCRs. The vehicles have been classified into four groups, viz. two wheelers, three wheelers, four wheelers gasoline powered and four wheeler diesel powered, for which emission factors, developed by Indian Institute of Petroleum (Pundir et al. 1994), have been used for estimating CO and NO₂ source strengths.

Development of ANN-based CO models

Table 1 provides the model input data and ANN architecture used for the development of various ANN-based VEE models, named as ANNCO for predicting 8-h average CO concentrations. The inputs to model are directly connected to the quantity of information given to the neural network and is generally constituted from meteorological and traffic characteristic data. Theoretical fundamentals on VEE dispersion processes help to find the right neural network inputs. Three sets of models have been formulated—first, considering both meteorological and traffic characteristics input data (ANNCO_A); the second, considering only meteorological input data (ANNCO_B); and the third, considering only traffic input data (ANNCO_C). The output corresponding to these inputs is 8-h average CO concentrations. The number of hidden layers and its neuron, learning rate (η), momentum term (μ), learning algorithm and activation function, depend on the problem complexity viz. the number of training patterns and the amount of noise in the data (Gardner and Dorling 1998; Bose and Liang 1998; Wasserman 1989). The formulation of ANN-based VEE models has been discussed elsewhere (Nagendra and Khare 2003).

Table 1 Neural network architecture and Input variables for various ANN-based VEE models

During ANN-based VEE model training, it is advisable to divide the data into three partitions, namely, ‘training data set’, ‘validation data set’ and ‘test data set’ (to avoid over training, which results poor model performance on validation data set). The ‘training data set’ forms the bulk of the data used for training purpose; the ‘validation data set’ is used during training in order to check the generalization performance of the neural network model. Training can be stopped when the performance of the model on the ‘validation data set’ gives minimum error. Finally, the ‘test data set’ is used to test the final neural network model (Gardner and Dorling 1998). This suggests that the ‘training data set’ must fully represent all the cases for which the ANN-based VEE model is required to be generalized. When a representative ‘training data set’ is not available, accurate prediction accuracy impedes (Boznar et al. 1993). In the present work, the ANNCO models have been trained on two-year data (1997 and 1998); and the data pertaining to the year, 1999, has been used as both the validation and test data sets. The number of patterns used for training, validation and testing of ANN models is shown in Table 2. The random selections of the data for training, generalization and for the test purpose are based on seasonal variations in the meteorology and CO concentration in the AQCRs. Since, the ANN-based VEE models have been trained on data of a selected AQCR, it can therefore only be used for such regions whose characteristics are similar to the present selected AQCRs (Comrie 1997).

Table 2 Training, validation and test data sets for ANN-based VEE models

The performance of all the models has been evaluated using the statistical parameters viz. mean of the observed and predicted concentrations ( $\bar O$ and $\bar P$, respectively) and their standard deviations (σ _O and σ _P , respectively); mean bias error (MBE); mean square error (MSE); systematic and unsystematic root mean square error (RMSE_S and RMSE_U); coefficient of agreement (r²); linear best fit constant (a) and gradient (b); and index of agreement (d) (Willmott 1982). The ‘d’ is a descriptive statistics that reflects the degree to which the observed variate is accurately estimated by the simulated variate. The ‘d’ is not a measure of correlation or association in the formal sense, but rather a measure of the degree to which model predictions are error free. At the same time, ‘d’ is a standardized measure, in order that it may be easily interpreted and cross-comparisons of its magnitudes for a variety of models (regardless of units) can readily be made. It varies between 0 and 1. A computed value of 1 indicates perfect agreement between the observed and predicted observations, while 0 connotes complete disagreement (Willmott 1982). The value of ‘d’ is expressed as:

$$ d = 1 - \frac{{\sum\nolimits_{i = 1}^N {\left( {P_i - O_i } \right)^2 } }} {{\sum\nolimits_{i = 1}^N {\left[ {\left| {P_i - \bar O} \right| - \left| {O_i - \bar O} \right|} \right]^2 } }} $$

(1)

where, N is number of the data points; O _i observation data points; P _i predicted data points; $\bar O$ mean of the observed data points.

It is the most commonly used statistical indicator in the air quality model performance studies (Khare and Sharma 1999; Gardner and Dorling 1999, 2000; Comrie 1997)

AQCR1

Meteorological and traffic characteristic variables as model input

The total data set of 8-h average CO values at AQCR1 includes, 3,285 values. The total of 106 missing values represented 3% of the whole data set. About 65% of the total data have been used for model training, 16% for validation, and 16% for final test of the 8-h average ANN-based CO model. Several hundreds of experiments have been performed to determine the best combination of the η, μ, number of the hidden layers, H, learning algorithm and activation function. A fully connected feed-forward neural network, with 17 neurons in the input layer, 3 neurons in the single hidden layer and 1 neuron in the output layer, shows best prediction on ‘validation data set’. The best RMSE and ‘d’ values have been found for the model parameters, ‘η’=0.001 and ‘μ’=0.3 with 400 epoch. Table 3 describes performance of the ANNCO_A1 model during generalization on the ‘validation data set’.

Table 3 Estimates of the statistics during generalization of the ANNCO_A1 model

Meteorological variables as model input

The purpose of formulating this model is twofold. First, to develop ANN-based VEE model (ANNCO_B1) to forecast 8-h average CO concentration using routinely monitored meteorological variables. Second, to study the sensitivity of the traffic characteristic variables. The number of training and validation patterns remains same as that of ANNCO_A1 model. Neural network architecture, 10:3:1 has been used for the development of the ANNCO_B1 model. After repeated experiments, the best model prediction on validation data set has been achieved at 1,000 epoch with ‘η’=0.001 and ‘μ’=0.1 (Table 4).

Table 4 Estimates of the statistics during generalization of the ANNCO_B1 model

Traffic characteristic variables as model input

This model has been developed with five traffic characteristic variables as input to the ANN-based CO model (ANNCO_C1), i.e. two-wheeler, three-wheeler, gasoline-powered four wheeler, diesel-powered four-wheeler and source strength of CO. Neural network architecture 5:3:1 has been used for the development of the ANNCO_C1 model. Table 5 presents the generalization statistics of the ANNCO_C1 model on validation data set. The best model prediction has been obtained at ‘η’=0.001 and ‘μ’=0.1 with 20 epoch. Table 6 presents summary of the ANNCO_A1, ANNCO_B1 and ANNCO_C1 models parameter and their performance statistics on validation data set at AQCR1.

Table 5 Estimates of the statistics during generalization of the ANNCO_C1 model

Table 6 Summary of the parameters for the models for the site AQCR1 and the model performance statistics on validation data set

AQCR2

Meteorological and traffic characteristic variables as model input

At AQCR2, a total of 859 missing values represented 26% of the whole data (3,285). Out of the total data, 56% has been used for training purpose, 8% for validation and another 10% for final model test. Network architecture of 17:3:1 has been used for the development of ANNCO_A2 model. This model is similar to ANNCO_A1 model of AQCR1, consisting of 17 input variables. After repeated experiments, the best prediction has been obtained at 1,400 epoch with ‘η’=0.001 and ‘μ’=0.5 (Table 7).

Table 7 Estimates of the statistics during generalization of the ANNCO_A2 model

Meteorological variables as model input

A 10:3:1 structure has been used for the development of the ANNCO_B2 model. Ten meteorological variables have been used in the model formulation. With η=0.001 and μ=0.7, the ANNCO_B2 model gives best prediction on ‘validation data set’ at 4,000 epoch. Table 8 describes performance of the ANNCO_B2 model during generalization on the ‘validation data set’.

Table 8 Estimates of the statistics during generalization of the ANNCO_B2 model

Traffic characteristic variables as model input

Five traffic characteristics variables have been used in formulating the ANNCO_C2 model. A 5:3:1 neural network architecture has been used for the development of the ANNCO_C2 model. Table 9 presents the generalization statistics of the ANNCO_C2 model on ‘validation data set’. The best model prediction has been obtained at η=0.001 and μ=0.7 with 20 epoch.

Table 9 Estimates of the statistics during generalization of the ANNCO_C2 model

The summary of the ANNCO_A2, ANNCO_B2 and ANNCO_C2 models parameter and their performance statistics on validation data set at AQCR2 are presented in Table 10.

Table 10 Summary of the parameters for the models for the site AQCR2 and the model performance statistics on validation data set

Results and discussion

Tables 11 and 12 gives the performance statistics of the trained ANN-based VEE model predictions on test data sets at AQCR1 and AQCR2, respectively. The prediction performance of ANNCO_A1 model at AQCR1 and ANNCO_A2 model at AQCR2 shows that the mean of predicted CO concentration at AQCR1 (P=5.56 ppm) is much higher than that of the observed mean (3.78 ppm); while at AQCR2, it is lower (P=3.39 ppm) than the observed mean value (4.18 ppm). At AQCR1, the MBE value is positive (1.78 ppm), indicating a tendency of the model to over predict; and at AQCR2, it is negative (−0.79 ppm), indicating a tendency of the model to under predict. The standard deviation (σ _P ) of the ANNCO_A1 and ANNCO_A2 model predictions has been found to be 2.35 and 2.1 ppm, respectively. At AQCR1, σ _P is matching with the standard deviation of the observed data. At AQCR2, the difference between the standard deviations of observed and predicted data is marginal (1.25 ppm). This explains that ANNCO_A1 model is reproducing the variations in the test data set with reasonable accuracy; where as, the ANNCO_A2 model is showing moderate variations in the test data set. A low RMSE_S value at AQCR1 indicates that the ANNCO_A1 model predictions are closely matching with the actual observations when compared to the ANNCO_A2 model predictions at AQCR2. Further, the ‘d’ values for the ANNCO_A1 and ANNCO_A2 models are 0.78 and 0.69, respectively. This explains that, at AQCR1, 78% of the model predictions are error free; and at AQCR2, only 69% of model predictions are error free. It shows that ANNCO_A1 model predictions are more accurate than the ANNCO_A2. Figures 2 and 3 show observed versus predicted CO concentrations at AQCR1 and AQCR2, respectively, indicating that both models under-predict CO concentrations, when observed concentrations are towards the higher side.

Table 11 Performance statistics of the models for the site AQCR1

Table 12 Performance statistics of the models for the site AQCR2

Fig. 2

Comparison of 8-h average CO observations vs the ANNCO_A1 model predictions for the test dataset at AQCR1

Fig. 3

Comparison of 8-h average CO observations vs the ANNCO_A2 model predictions for the test dataset at AQCR2

The mean of the ANNCO_B1 model predicted CO concentration at AQCR1 is higher than the observed mean; and at AQCR2, the mean of the ANNCO_B2 model predictions is slightly lower than that of the observed mean. The MBE value at AQCR1 is positive (1.45 ppm), indicating a tendency of the model to over-predict; and at AQCR2, it is negative (−0.71 ppm), indicating a tendency of the model to under predict. The standard deviation of the ANNCO_B1 model predictions is matching with the standard deviation of the observed data set at AQCR1. At AQCR2, the difference between the standard deviation of observed and predicted data is quite high. It explains that ANNCO_B1 model is able to reproduce the variations in the test data set at AQCR1; where as, at AQCR2, ANNCO_B2 model fails to reproduce the variations in the test data set. A low RMSE_S value at AQCR1 indicates that ANNCO_B1 model predictions are closely matching with the actual observations when compared with the ANNCO_B2 model predictions at AQCR2. Further, the ‘d’ values for the ANNCO_B1 and ANNCO_B2 models explain that at AQCR1, 77% of the model predictions are error free and at AQCR2, 67% of the model predictions are error free. It shows that ANNCO_B1 model is more accurate in its prediction performance than the ANNCO_B2 model. Figures 4 and 5 show observed versus predicted CO concentrations at AQCR1 and AQCR2, respectively, indicating that both models under-predict CO concentrations when observed concentrations are towards the higher side.

Fig. 4

Comparison of 8-h average CO observations vs the ANNCO_B1 model predictions for the test dataset at AQCR1

Fig. 5

Comparison of 8-h average CO observations vs the ANNCO_B2 model predictions for the test dataset at AQCR2

Comparative performance of ANNCO_B1 vs ANNCO_A1 model and ANNCO_B2 vs ANNCO_A2 model

The RMSE_S values for ANNCO_B1 and ANNCO_B2 models increase marginally when compared with the ANNCO_A1 and ANNCO_A2 models, respectively. Further, the ‘d’ values for AQCR1 indicate that the ANNCO_B1 model (d=0.77) performance decreases marginally, when compared with the ANNCO_A1 model (d=0.78). Similarly, at AQCR2, the ANNCO_B2 (d=0.67) performance decreases marginally when compared with the ANNCO_A2 model (d=0.69). Thus with increase in time average (8-h average), the relationship between CO dispersion with the meteorological and the traffic characteristic data does not represent the real non-linear relationship, i.e. in 8-h average data, the non-linear relationship between CO concentration with meteorological and traffic characteristic variables is less apparent. As a result, the elimination of traffic characteristic variables from the model inputs causes negligible effect on model performance. Comrie (1997) and Gardner and Dorling (1998) have also reported that increase in time average from 1 h to 24 h decreases the non-linearity among the variables.

The mean of the ANNCO_C1 model predicted CO concentration (5.03 ppm) at AQCR1 is higher than the observed mean (3.78 ppm); while at AQCR2, the mean of the ANNCO_C2 model prediction is lower (3.49 ppm) than the observed mean value (4.18 ppm). The MBE values at AQCR1 are positive, indicating a tendency of the model to over-predict; and at AQCR2, it is negative, indicating a tendency of the model to under-predict. The standard deviations (σ _P ) of the ANNCO_C1 and the ANNCO_C2 model predictions are very low when compared to the observed standard deviations (σ _O ). It explains that both models are inadequate to reproduce the variations in the test data set. The high RMSE_S values indicate that both models perform poorly on the test data set. The ‘d’ values for the ANNCO_C1 and ANNCO_C2 models are 0.4 and 0.26, respectively. This explains that at AQCR1, 40% of the model predictions are error free and at AQCR2, 26% of the model predictions are error free. Figures 6 and 7 show the observed versus predicted CO concentrations at AQCR1 and AQCR2, respectively, indicating that both models under-predict CO concentrations when observed concentrations are towards the higher side; and over-predicts CO concentrations when observed concentrations are towards the lower side.

Fig. 6

Comparison of 8-h average CO observations vs the ANNCO_C1 model predictions for the test dataset at AQCR1

Fig. 7

Comparison of 8-h average CO observations vs the ANNCO_C2 model predictions for the test dataset at AQCR2

Comparative performance of ANNCO_C1 vs ANNCO_B1 and ANNCO_A1 models and ANNCO_C2 vs ANNCO_B2 and ANNCO_A2 models

The RMSE_S value for ANNCO_C1 model increases by 0.83 and 0.73 ppm when compared with ANNCO_A1 and ANNCO_B1 models, respectively. However, for ANNCO_C2 model, the RMSE_S value increases by 1 and 0.9 ppm when compared with ANNCO_A2 and ANNCO_B2 models, respectively. Further, the ’d’ values indicate that ANNCO_C1 model (d=0.4) performs poorly at AQCR1, when compared with ANNCO_A1 (d=0.78) and ANNCO_B1 models (d=0.77). At AQCR2, the ANNCO_C2 model (d=0.26) also shows poor performance when compared with the ANNCO_A2 (d=0.69) and ANNCO_B2 model (d=0.67). The reason for the poor performance of ANNCO_C1 and ANNCO_C2 models may be explained by the following facts. First, these models are developed considering only traffic characteristic variables as their inputs. As a result, the models explain the CO dispersion due to traffic wake effect only. Second, due to the absence of meteorological input variables, these models fail to take into account the ‘lag effect’. “Lag effect” is a phenomenon that results in the accumulation of CO in the atmosphere during inversion conditions, increasing its ambient concentration even when traffic is just a trickle. This phenomenon frequently occurs during critical winter periods (November–March), when inversion conditions prevail during nighttime, particularly 4–6 h after 6 p.m. (Khare and Sharma 1999).

Limitations of the neural network models

ANNs may be viewed as a supplement to the conventional computing techniques (Flood and Kartam 1994). The problem during the ANN model training is selection of network architecture i.e., number of hidden layers, nodes in those layers and their interconnection. Apart from these, no thumb rules exist in the selection of data set for training, testing and validation of neural network model (Nagendra and Khare 2002a; Gardner and Dorling 1998). Nagendra and Khare (2003) have discussed, in detail, the step-by-step ANN-based VEE modelling procedure to overcome the above limitations in application of the neural network technique in air quality predictions.

Conclusions

ANN-based VEE models have been developed to predict air quality in terms of CO near a traffic intersection (AQCR1) and an arterial road (AQCR2), in the Delhi city. The results show that ANN-based CO models, with both meteorological and traffic characteristic variables and with only meteorological variables show best performance on the test data set at both the AQCRs. The study also shows that elimination of traffic characteristic variables from the model inputs causes negligible effect on model performance. However, the models developed with only traffic characteristics inputs show poor performance on the test data set at both the AQCRs and reflect their inability to take into account the ‘lag-effect’.