Environ Sci Pollut Res DOI 10.1007/s11356-014-2821-z
RESEARCH ARTICLE
Artificial intelligence modeling to evaluate field performance of photocatalytic asphalt pavement for ambient air purification Somayeh Asadi & Marwa Hassan & Ataallah Nadiri & Heather Dylla
Received: 21 August 2013 / Accepted: 21 March 2014 # Springer-Verlag Berlin Heidelberg 2014
Abstract In recent years, the application of titanium dioxide (TiO2) as a photocatalyst in asphalt pavement has received considerable attention for purifying ambient air from trafficemitted pollutants via photocatalytic processes. In order to control the increasing deterioration of ambient air quality, urgent and proper risk assessment tools are deemed necessary. However, in practice, monitoring all process parameters for various operating conditions is difficult due to the complex and non-linear nature of air pollution-based problems. Therefore, the development of models to predict air pollutant concentrations is very useful because it can provide early warnings to the population and also reduce the number of measuring sites. This study used artificial neural network (ANN) and neuro-fuzzy (NF) models to predict NOx concentration in the air as a function of traffic count (Tr) and climatic conditions including humidity (H), temperature (T), solar
Responsible editor: Michael Matthies S. Asadi Department of Civil and Architectural Engineering, Texas A&M University-Kingsville, MSC 194, 700 University Blvd, Kingsville, TX 7836, USA e-mail: [email protected] M. Hassan (*) Department of Construction Management, Louisiana State University, 3130 A PFT Hall, Baton Rouge, LA 70803, USA e-mail: [email protected] A. Nadiri University of Tabriz, 29 Bahman Boulevard, Tabriz, East Azarbaijan, Iran e-mail: [email protected] H. Dylla Department of Construction Management, Louisiana State University, Baton Rouge, LA 70803, USA e-mail: [email protected]
radiation (S), and wind speed (W) before and after the application of TiO2 on the pavement surface. These models are useful for modeling because of their ability to be trained using historical data and because of their capability for modeling highly non-linear relationships. To build these models, data were collected from a field study where an aqueous nano TiO2 solution was sprayed on a 0.2-mile of asphalt pavement in Baton Rouge, LA. Results of this study showed that the NF model provided a better fitting to NOx measurements than the ANN model in the training, validation, and test steps. Results of a parametric study indicated that traffic level, relative humidity, and solar radiation had the most influence on photocatalytic efficiency. Keywords Artificial neural network . Neuro-fuzzy . Nitrogen oxides . Titanium dioxide . Photocatalytic pavement
Introduction Photocatalytic pavements are currently being investigated as a potential solution to remove air pollutants. These pavements consist of a surface layer of titanium dioxide (TiO2) nanoparticles, which can reduce or oxidize both organic and inorganic particles that are absorbed into active sites when irradiated by UV light (Beeldens 2006). Several laboratory studies have shown the effectiveness of photocatalytic pavements; however, understanding the efficacies in real-world environment remains a challenge (Berdahl and Akbari 2008). Previous researchers have shown that the efficiency of photocatalytic pavement depends on environmental factors including the concentration of pollutants, relative humidity, temperature, UV intensity, and wind speed, which complicate the interpretation of field test results (Sleiman et al. 2009; Venturini and Bacchi 2009; Bengtsson and Castellote 2010; Hassan et al. 2012).
Environ Sci Pollut Res
Two techniques have been used to evaluate the photocatalytic degradation in the field. The first is to measure the reduction of NOx concentration directly by measuring the ambient air pollution concentrations and the second is to measure the reduction indirectly by measuring the byproducts created from the degradation process (Dylla et al. 2012b). Using these two techniques, several approaches have been used in order to study photocatalytic efficiency in the field (Venturini and Bacchi 2009; Beeldens 2008; Maggos et al. 2007, 2008; Li and Qian 2009; Chen and Chu 2011). A review of the literature reveals that several approaches are available for predicting NOx concentrations in the air. One approach is to create a model to predict the concentrations of NOx in the air for a given area without photocatalytic pavement and an area with photocatalytic pavement. There are several roadway microenvironments pollution dispersion models such as Gaussian plume dispersion, computational fluid dynamics (CFD), atmospheric box, and statistical models (Lin and Yu 2008). Both statistical and computational fluid dynamic models have been developed to characterize photocatalytic environments (Dylla et al. 2013; Moussiopoulos et al. 2008). Among these, statistical models are simple and do not require additional laboratory kinetic studies to predict photocatalytic efficiency. Many scientific investigations have been conducted to predict different air pollutants’ concentrations using statistical tools including regression, multi-regression, and artificial neural networks (Karppinen et al. 2000; Lin and Wu 2003; Sharma et al. 2004). Using these tools, the pollution concentrations are estimated by statistical relationship between various factors collected at a particular receptor (Sharma et al. 2004). Regression statistics previously used by authors to evaluate the efficiency of photocatalytic pavements provided fair results. The developed model considered different factors such as solar radiation, relative humidity, wind speed, traffic, and temperature (Dylla et al. 2013). The results of the statistical models showed a good agreement between the predicted and observed data but some points had high errors. Statistical models may be able to obtain a relationship between the input and the output variables without describing the causes and effects in the formation of pollutants. However, they are not able to evaluate the intrinsic non-linear nature of the problem and to predict short-term pollution levels (Agirre-Basurko et al. 2006; Barai et al. 2007; Akkoyunlu et al. 2010; Yetilmezsoy and Abdul-Wahab 2012). To overcome this problem, mathematical models have been used as vital tools in both design and operation especially in cases which deal with high dimensional data such as air quality. A review of literature reveals that several artificial intelligence (AI) models are available for predicting the air pollutant concentrations (He and Ma 2010; Sousa et al. 2007; Abdul-Wahab and Al Alawi 2002; Carnevale et al. 2009; Pires et al. 2012). Over the last four decades, different AI models
including artificial neural network (ANN) and neuro-fuzzy (NF) have been proposed as alternatives to traditional statistical models. AI models can be used in the modeling of different real-life processes in environmental engineering due to their predictive capabilities and non-linear characteristics (Hydrology 2000; Maier and Dandy 2000); (Asghari et al. 2010; Maier et al. 2010; Nadiri et al. 2013a, b). Using AI models, some studies predict the concentrations of different air pollutions such as carbon monoxide (CO) (Gautam et al. 2008; Cai et al. 2009), particles measuring 10 μm or less (PM10) (Hooyberghs et al. 2005; Patricio and Jorge 2006; Paschalidou et al. 2011), ozone (Sousa et al. 2007; AbdulWahab and Al Alawi 2002; Pires et al. 2012; Salazar-Ruiz et al. 2008) and sulfur monoxide (SO) (Brunelli et al. 2007). Sousa et al. (2007) constructed multiple linear regression and artificial neural networks models based on principal components to predict ozone concentrations. A new methodology based on feedforward artificial neural networks using principal components as inputs was used to predict next day hourly ozone concentrations. Then, they compared the results of the developed model with multiple linear regressions, feedforward artificial neural networks based on the original data, and also with principal component regression. They found that the use of principal components as inputs improved both model prediction by reducing their complexity and eliminating data collinearity. Later, He and Ma (He and Ma 2010) used a back-propagation neural network based on principal component analysis to model the internal greenhouse humidity during winter season in North China. The results of this model were compared with the stepwise regression model. The results indicated that the stepwise regression method was less accurate than the back-propagation neural network model. Each AI method has its own advantages. The fuzzy models tend to be robust to parameter changes, and are also tolerant to imprecision and uncertainty (Bardossy and Disse 1993). While the ANN model represents non-linear relationships and learns these relationships directly from the data being modeled (Palani et al. 2008). Obviously, the NF model takes advantage of the fuzzy logic (FL) and ANN in modeling. The NF model combines the advantages of fuzzy systems—deal with explicit knowledge—ANN deals with implicit knowledge. This NF model, by taking advantage of ANN and fuzzy logic methods, have shown high capability to estimate air pollution (Yildirim and Bayramoglu 2006; Carnevale et al. 2009). The objective of this study was the construction of AI models which give the possibility of predicting the NOx concentration in the air for a given area without photocatalytic pavement and an area with photocatalytic pavement as a function of traffic count (Tr) and climatic conditions including humidity (H), temperature (T), solar radiation (S), and wind speed (W). This model relates the NOx concentration to both
Environ Sci Pollut Res
meteorological variables as well as to traffic volume and help to understand the effect of each parameter on the efficiency of photocatalytic asphalt pavement to reduce NOx concentration in the air. The AI models are useful for modeling because of their ability to be trained using historical data and because of their capability for modeling highly non-linear relationships.
Field site description Data used in the models were obtained from a field study located in Baton Rouge, Louisiana, where 0.2 miles of asphalt pavement was sprayed with an aqueous TiO2 solution (Dylla et al. 2013). The photocatalytic spray coat was a mixture of TiO2 anatase nanorods 12 nm wide by 75 nm long suspended in an aqueous liquid at 2 % by volume (Dylla and Hassan 2012a). Before application, the roadway was cleared of any debris. A primer was applied before the photocatalytic coat. A distributor truck was used in the application process. Mounted on the back of the truck, a spray bar fitted with nozzles distributed TiO2 water-based solution at the specified application rate, 16.1 to 21.5 ml/m2. The application rate was adjusted by altering the truck speed. Further, the nozzles had electrostatic precipitators to separate the TiO2 nanoparticles suspended in the aqueous solution and to ensure a more even coverage. Equipment used for field data collection was housed in an air-conditioned trailer. A Thermo 42i NOx analyzer was used for monitoring NO, NO2, and NOx concentrations. The NOx analyzer was calibrated in accordance to EPA calibration procedures using the gas phase titration (GPT) alternative. A Thermo 146i gas calibrator was used for calibration of the NOx analyzer. The NOx analyzer was calibrated at five different spans for NO and four different ozone settings for NO2 to confirm linearity and ozone converter efficiency. In addition, the Department of Environmental Quality (DEQ) air monitoring station is located next to the airconditioned trailer. After calibration, the NOx analyzer was connected to ambient air at the pavement level using a stainless steel pipe placed in the middle lane and was covered with a protective bump.
Table 1 Parameters range of variation Tr (V/5 min) S (W/m2) W (m/s) H (%) T (°C)
To monitor climatic conditions at the site, a Davis 6152 Wireless Vantage Pro weather station was installed in the field which recorded and stored meteorological data including humidity, ambient air temperature, wind speed, and solar radiation continuously each minute. A portable Peek ADR-1000 traffic counter capable of counting vehicles per 5-min intervals was employed. Traffic counter was installed to count the number of vehicles per 5-min per lane. In order to consider the effect of climatic conditions on the performance of photocatalytic activities, data were collected under different climatic conditions.
Parameters selection The goal of this study was to predict NOx concentration in the air as a function of different parameters before and after application of TiO2 on the asphalt pavement. The selected parameters and their range of variation are shown in Table 1. These parameters were selected based on laboratory test results (Hassan et al. 2012). Laboratory results established a relationship between NOx concentration and humidity, UV intensity, temperature, and flow rate.
Methodology Artificial intelligence methods Different non-linear AI models may be used for air pollutants’ prediction. In this study, ANN and NF models were adopted for prediction of NOx concentrations. The main goal of the models is to predict NOx concentration and investigate the sensitivity of NOx concentration to related variables defined as input data. The input data for the AI models for the case study in Baton Rouge, LA, are shown in Table 1. A brief description of the two AI models is provided in the following sections. ANN An ANN model was developed to predict NOx concentration in this study. ANN is a universal approximator to surrogate
Description
Min
Max
Mean
Standard deviation
Traffic count Solar radiation Wind speed Relative humidity Temperature
0 0 0 32 11
36 858 4 94 30
15 123 0.8 77 20.4
13.4 197.2 0.91 14.01 4.11
Environ Sci Pollut Res
complex systems (Hydrology 2000; Maier et al. 2010). Neural networks are composed by simple connected elements (neurons) operating in parallel. Each neuron includes a nonlinear function relating inputs and outputs (activation function). The most widely used neural network is the multi-layer perceptron (MLP) network (Hydrology 2000; Nourani et al. 2008) that consists of an input layer, a hidden layer, and an output layer. This study considers one hidden layer for the MLP network as shown in Fig. 1a. The structure of the ANN model, applied to predict NOx concentration, contains five neurons in the input layer pertain to the input data (Tr, H, S, W, and T), one neuron in the output layer and four neurons in the hidden layer that was determined via trial and error. The normalized input signal spreads through the network in a forward direction via connections between neurons. Incoming signals are linearly combined and converted to outgoing signals. The signal conversion is conducted by assigned activation functions. The mathematical expression for a three-layer feedforward ANN is given as (Hydrology 2000; Nadiri et al. 2013a): ! X Oj ¼ f 1 bj þ W ji I i ð1Þ
Wji and Wkj are weights that control the strength of connections between two layers, and the biases bj and bk are used to adjust the mean value for input layer and hidden layer, respectively. The activation function for the hidden layer is typically a continuous and bounded non-linear transfer function such as sigmoid and log sigmoid functions. The activation function for the output layer is usually a linear function. In this study, hyperbolic tangent sigmoid and linear function was selected for f1 and f2, respectively. The output Ok of the ANN is the NOx concentration. In the ANN training step, a supervised learning algorithm was used to estimate the weights Wji and Wkj (Hydrology 2000). To construct ANN model and to prevent over-fitting, a cross validation method procedure is usually recommended (Hydrology 2000). The cross validation involves partitioning a sample of data into complementary subsets, performing the analysis on one subset (training set), and validating the analysis on the other subsets (called the validation set and testing set). Therefore, the measured data were randomly divided into three groups, training (70 %), validation (15 %), and testing (15 %). NF
i
Ok ¼ N Ox ¼ f 2 bk þ
X
! W kj O j
ð2Þ
j
where f1 and f2 are the activation functions for the hidden and output layers, respectively, Ii is the ith input, Oj is the jth output, Fig. 1 Schematics of a ANN structure and b NF structure
Neuro-fuzzy (NF), as shown in Fig. 1b, combines the advantage of the FL and ANN methods by applying various learning techniques developed in the ANN literature to optimize the FL model (Nayak et al. 2004). Based on the fuzzy-set theory, an element of the world belongs to a set, specifying a feature of the element (linguistic variable), with a value ranging from 0 to 1 according to a function (membership function). Fuzzy
Environ Sci Pollut Res
sets have also ambiguous boundaries and gradual transitions between defined sets, which are appropriate to deal with the nature of uncertainty in the system and human errors (PulidoCalvo and Gutiérrez-Estrada 2009). Following the conventional logic, an inference system based on the rules in the form of “if-then” is formulated for the fuzzy logic (Carnevale et al. 2009). The first step for preparation of the NF model to predict NOx concentration is construction of a FL model. A FL model developed by Takagi and Sugeno (Takagi and Sugeno 1985; Sugeno and Yasukawa 1993) termed as TS-FL model, was adopted. In this model, the output membership functions are either constant (zero order) or linear (first order). Input clusters and output membership functions are extracted by a clustering process. An effective subtractive clustering method (Chen and Wang 1999) for TS-FL modeling was used for the extraction of clusters and fuzzy “if-then” rules. The important parameter in subtractive clustering, which controls number of clusters and fuzzy “if-then” rules, is cluster radius that takes values between 0 and 1. Decreasing the cluster radius increases the number of clusters and leads to smaller clusters. In contrast, a large cluster radius produces large clusters in the data and results in few rules (Aqil et al. 2007). A TS-FL model without sufficient number of rules cannot cover the entire domain and with too many rules, becomes complicated and may lead to low performance. Therefore, the cluster radius had to be optimized via trial and error to have a suitable number of clusters and rules. Searching for the optimal cluster radius can be accomplished by systematically varying cluster radius value from 0 to 1 until a minimal root mean squared error (RMSE) is achieved. To apply the first-order TS-FL model, a generalized Gaussian function was used to develop membership functions for the five input data, see Table 1. Each input was clustered into nine classes. Then, based on membership functions of the five input data, a set of fuzzy if-then rules were developed to linearly aggregate the input data as the output. Since the subtractive clustering method was used, the number of rules was equal to the number of clusters. For NOx prediction in this study, a fuzzy “if-then” rule i can be expressed as: Rule i: if (Tr belongs to MFiTr), (H belongs to MFiH), (S belongs to MFiS), (W belongs to MFiW), and (T belongs to MFiT), then ðN Ox Þi ¼ mi Tr þ ni H þ pi S þ qi W þ zi T þ ci
ð3Þ
where NOx is the output of rule i, MFiTr denotes the membership function of the ith cluster of input Tr, MFiTr represents the membership function of the ith cluster of input H, and so forth. mi, pi, qi, zi, and ci are coefficients to be determined by linear least-squares estimation. The second step of NF construction
is adopting a five-layer MLP network to optimize coefficients of linear membership function of output and also the Gaussian membership function parameters of input. The NF structure is shown in Fig. 1b. In each layer, the operations are as follows: Layer 1: Generate membership function of input data. The output of neuron i is defined by O1i ¼ μji ðX Þ
ð4Þ
where j is the number of input and i denotes the membership function index. X={Tr, H, S, W, T} represents a set of input. μji(X) is a fuzzy set associated with neuron i which is a membership function. Layer 2: Calculate firing strength Wi for the ith rule via multiplication via “and” operator: O2i ¼ wi ¼ μ1i ðX Þμ2i ðX Þμ3i ðX Þμ4i ðX Þμ5i ðX Þð5Þ
Layer 3: Compute the normalized firing strengths for the ith neuron: wi O3i ¼ wi ¼ X wi
ð6Þ
i
Layer 4: Compute the contribution of the ith rule in the model output based on the first-order TS-FL method: O4i ¼ wi ðNOx Þi ¼ wi ðmi Tr þ ni H þ pi S þ qi W þ zi T þ ci Þ
ð7Þ
Layer 5: Calculate the final output as the weighted average of all rule outputs (aggregation): O5i ¼ N Ox
X
w i ðN O x Þi
ð8Þ
i
The NF parameters in Eq. 8 and the membership function parameters were optimized using a hybrid algorithm, which is a combination of the gradient decent and the leastsquares method (Aqil et al. 2007).
Environ Sci Pollut Res
Efficiency criteria
Results and analysis
Two statistical measures were used to evaluate the effectiveness of the ANN and NF models and their ability to make accurate predictions. The root mean square error (RMSE) is calculated by:
Artificial neural network
i¼1
ð9Þ
n
where Y i ; Yb i and n are the measured data, the predicted data, and the number of measurements, respectively. RMSE indicates the discrepancy between measured and predicted values. The R2, coefficient of determination, was also used, which is given by:
i R2 ¼ 1− X n
Y i −Yb
Y i −Y
2 ð10Þ
2
i
where Y is the mean of observed data. The best fit between measured and predicted values will give RMSE and R2 close to 0 and 1, respectively.
70
70
70
60
60
60
50
50
50
Predicted values
Predicted values
a
40 30 20 10
40 30 20 10
0
30 20
0 0 10 20 30 40 50 60 70
Actual values
0 10 20 30 40 50 60 70
Actual values
Actual values
40
40
40
30
30
30
Predicted values
b
40
10
0 0 10 20 30 40 50 60 70
Predicted values
Fig. 2 Comparison between the measured and predicted values of NOx concentrations for ANN in the train, test, and validation steps for a before TiO2 application and b after TiO2 application
Predicted values
n X
20
10
0
Predicted values
RMSE ¼
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi n uX 2 u u Y i −Yb j t
For the construction of both ANN models (i.e., before and after TiO2 application), the presented network in Fig. 1a was applied. The numbers of hidden nodes were determined to be four nodes through the trial and error method. The weights and biases in the ANN models were estimated by minimizing error using the Levenberg-Marquardt (LM) algorithm (Hydrology 2000). The first ANN model was developed to predict NOx concentration before application of TiO2. The R2 and RMSE of this model after 354 epochs of training were 0.96 and 3.1 ppb, respectively. For the validation and test steps, the R2 and RMSE were 0.95 and 3.3 ppb and 0.93 and 3.4 ppb, respectively. The second ANN model was developed with 372 epochs to predict NOx concentration after TiO2 application on the asphalt pavement. The RMSE for the training, validation, and test steps after the application of TiO2 were 3.1, 3.4, and 3.7 ppb, respectively. The R2 for the training, validation, and test were 0.96, 0.94, and 0.92, respectively. The scatter plot of measured and ANN estimated NOx concentration is shown in Fig. 2a and b. From these figures, it is noted that the data are concentrated around the solid line, which indicates a model with acceptable accuracy since the closer the distances of these data to the solid line are, the better the accuracy of the model.
20
10
0 0
10 20 30 40 Actual values
20
10
0 0
10 20 30 40 Actual values
0
10 20 30 40 Actual values
Environ Sci Pollut Res 70
70
60
60
60
50
50
50
40 30 20 10
Predicted values
70
Predicted values
a Predicted values
Fig. 3 Comparison between the measured and predicted values of NOx concentrations for NF in the train, test, and validation steps for a before TiO2 application and b After TiO2 application
40 30 20 10
0
0 0 10 20 30 40 50 60 70
Actual values
Actual values
40
40
30
30
30
20
10
Predicted values
40
Predicted values
Predicted values
20
0 10 20 30 40 50 60 70
Actual values
b
30
10
0 0 10 20 30 40 50 60 70
40
20
10
0
10
0 0
10
20
30
Actual values
Neuro-fuzzy The first step for construction of the two NF models (i.e., before and after TiO2 application) was the determination of optimal TS-FL clusters of five input data and rules. To develop an efficient TS-FL model with a reasonable number of fuzzy “if-then” rules, the cluster radius was optimized. Based on the minimum RMSE criteria for both models, a cluster radius of 0.5 and three rules were determined for the TS-FL model. In each rule, the parameters, mi, pi, qi, zi, and ci in the output membership function were estimated by the leastsquare error method. The second step was using ANN models to estimate parameters in the Gaussian membership functions and the coefficients of the model output (Eq. 8). A hybrid algorithm, which combines the least-squares method and the back-propagation gradient descent method, was employed (Zounemat-Kermani and Teshnehlab 2008). The first NF model was developed to predict the NOx concentration before TiO2 application. The R2 and RMSE of this model after five epochs of training were 0.98 and 1.7 ppb, respectively. The R2 and RMSE of 0.97 and 1.8 ppb were obtained in the validation step. The R2 and RMSE were 0.96 and 2.1 ppb, respectively, in the test step. The second NF model with five epochs was developed to predict NOx concentration after the application of TiO2. The RMSE for the training, validation, and testing steps after application of TiO2 were 2.1, 2.15, and 2.16 ppb, respectively. The coefficients of determination for the train, validation, and test steps were
40
20
0 0
10
20
30
40
0
10
Actual values
20
30
40
Actual values
0.97, 0.95, and 0.94, respectively. The measured and NF estimated NOx concentration scatter plots are shown in Fig. 3a, b. It can be observed that the data obtained from NF model are well correlated with the measured data points, which correspond to an excellent fit. The RMSE and the coefficients of determination for ANN and NF models are presented in Table 2. By taking advantage of FL and ANN models, the NF models provided a slightly Table 2 Fitting test parameters for ANN and NF models
Criterion
Step
Before application R2 Train RMSE R2 Validation RMSE R2 Test RMSE After application R2 Train RMSE R2 Validation RMSE R2 Test RMSE
Model ANN
NF
0.96 3.1
0.98 1.77
0.95 3.3 0.93 3.4
0.97 1.83 0.96 2.11
0.96 3.1 0.94 3.4 0.92 3.7
0.97 2.1 0.95 2.15 0.94 2.16
Environ Sci Pollut Res Fig. 4 Effect of traffic levels on NOx concentration
Concentration (ppb)
Before
After
35 30 25 20 15 10 5 0 Max
Mean
Min
Number of Vehicle ( V/5 min)
better fitting to NOx data than ANN model in the training, validation, and testing steps. Both models outperformed the statistical model initially used by the authors (Dylla et al. 2013). Parametric study The photocatalytic efficiency depends on several environmental factors and operating conditions, which makes it very complex to investigate. A parametric study was conducted using the NF models to study the effects of different parameters on NOx reduction efficiency. The NF model was used due to its considerable accuracy over ANN in the training, validation, and test steps. The objective of the parametric study was to evaluate the effect of each factor on the photocatalytic degradation of NOx. To achieve this objective, a sensitivity analysis was carried out in which each parameter was varied between the minimum and maximum level while the other parameters were kept constant at the mean value. The minimum, maximum, and mean values of each parameter are presented in Table 1.
Effect of traffic flow The first parameter investigated in this study was the traffic flow. Figure 4 shows a direct relationship between the number of vehicles per 5 min and NOx concentrations. It should be noted that in order to evaluate the effect of traffic flow on the NOx reduction based on the NF models, other parameters such as humidity, solar radiation, wind speed, and ambient temperature were kept constant at their mean values. As expected, increasing traffic flow increases NOx concentration as more pollutants are emitted to the environment. The developed model predicted that the TiO2 photocatalytic coating resulted in a significant reduction of NOx concentration at the different traffic levels. The model predicted approximately 60 % NOx reduction efficiency at the maximum number of vehicles per 5 min. Effect of relative humidity Figure 5 shows the effect of relative humidity on NOx concentrations per 5 min before and after the application of TiO2
Before
Fig. 5 Effect of humidity on NOx Concentration
After
Concentration (ppb)
25
20 15 10 5 0
Max
Mean Humidity
Min
Environ Sci Pollut Res
Before
Fig. 6 Effects of solar radiation on NOx concentration
After
Concentration (ppb)
25 20 15 10 5 0 Max
Mean
Min
Solar Radiation (W/m^2)
based on the NF model. As shown in this figure, increasing the relative humidity increases NOx concentrations in both models. Nevertheless, NOx concentrations before application are higher than after application at the different humidity levels indicating that the TiO2 photocatalytic layer is reducing NOx concentrations in ambient air. Figure 5 shows a clear trend, as the humidity increases, NOx concentration increases. This is mainly due to the excessive water vapor building on the TiO2 catalyst surface. Excessive water inhibits the photocatalytic reaction as it competes with pollutants for adsorption sites on the surface. This is in agreement with laboratory test results (Hassan et al. 2012).
Effect of solar radiation Photocatalytic oxidation necessitates UV light, part of the solar radiation spectrum, for the reaction to take place (Fujishima et al. 2000). The wavelength is the main factor influencing photocatalytic activity; in contrast, the strength of photon flux affects the rate of degradation (Fujishima et al.
2000; Zhao and Yang 2003). Therefore, an increase in UV irradiance is expected to increase photocatalytic NOx reduction efficiency. Increasing solar radiation from minimum (0 W/m2) to maximum level (858 W/m2) significantly increased the photocatalytic reaction rate, which resulted in higher NOx reduction efficiency (Fig. 6). Based on the NF models, at a solar radiation of 858 W/m2, NOx reduction efficiency per 5 min was approximately 77 %.
Effect of wind speed Figure 7 presents the effect of wind speed on NOx concentration per 5-min interval before and after application of TiO2. As shown in this figure, increasing wind speed decreases NOx concentration. This can be attributed to the fact that higher wind speed will help disperse the pollutants at the pavement level at a faster rate. Based on the model, at the maximum wind speed (4 m/s), NOx reduction efficiency per 5 min was approximately 42 % which is in agreement with laboratory results. It should be noted that higher wind speed would
Before
Fig. 7 Effects of a solar radiation and b wind speed on NOx Concentration
After
Concentration (ppb)
20 15 10
5 0 Max
Mean Wind Speed (m/s)
Min
Environ Sci Pollut Res
reduce the time available for the photocatalytic reaction to take place (Sleiman et al. 2009; Hassan et al. 2012).
Summary and conclusions This study improves the previous regression models developed by authors to assess and predict photocatalytic pavement’s field efficiency by employing artificial intelligence models. Both artificial neural network (ANN) and neurofuzzy (NF) models were used to predict NOx concentrations in the air as a function of traffic count (Tr) and climatic conditions including humidity (H), solar radiation (S), and wind speed (W) before and after the application of TiO2 on the pavement surface. Data were collected from a full-scale field study of a photocatalytic asphalt pavement located in Baton Rouge, LA. Results showed that the NF model provided better fitting to NOx measurements than ANN model in the training, validation, and testing steps. The R 2 and RMSE of NF model for the train, validation, and test steps were higher than ANN model. The R2 of the NF model was 0.98, 0.97, and 0.96, respectively for train, validation, and test steps before TiO2 application. The RMSE of the NF model was 1.77, 1.83, and 2.11 ppb, respectively, for train, validation, and test steps which were considerably lower than the RMSE obtained by ANN model. Results of a parametric study showed that the traffic level, relative humidity, ambient air temperature, solar radiation, and wind speed have significant effects on the performance of TiO2 in NOx degradation. Traffic level, relative humidity, and solar radiation had the most significant effects on photocatalytic NOx reduction. The increase in wind speed and relative humidity negatively affected the effectiveness of NOx reduction efficiency. However, the increase in UV light intensity improved NOx removal efficiency of the surface coating. The model predicted approximately 60 % NOx reduction efficiency at the maximum number of vehicles per 5 min, 77 % efficiency at solar radiation of 858 W/m2, 42 % efficiency at the maximum wind speed (4 m/s). Based on the results obtained in this study, further research is needed to assess the effects of additional factors including the impact of vehicle classification and vehicle activity. In addition, the durability of the photocatalytic coating in the field should be assessed. Acknowledgments This work was funded through a grant from the Gulf Coast Research Center for Evacuation and Transportation Resiliency. The authors would like to acknowledge PURETI for donating the materials needed for the construction of the field study and the support of Louisiana Transportation Research Center (LTRC) for granting access to their laboratory.
References Abdul-Wahab SA, Al Alawi SM (2002) Assessment and prediction of tropospheric ozone concentration levels using artificial neural networks. Environ Model Softw 17(3):219–228 Agirre-Basurko E, Ibarra-Berastegi G, Madariaga I (2006) Regression and multilayer perceptron-based models to forecast hourly O3 and NO2 levels in the Bilbao area. Environ Model Softw 21(4):430–446 Akkoyunlu A, Yetilmezsoy K, Erturk F, Oztemel E (2010) A neural network-based approach for the prediction of urban SO2 concentrations in the Istanbul metropolitan area. Int J Environ Pollut 40(4): 301–321 Aqil M, Kita I, Yano A, Nishiyama S (2007) Analysis and prediction of flow from local source in a river basin using a neuro-fuzzy modelling tool. J Environ Manag 85(1):215–223 ASCE Task Committee on Application of Artificial Neural Networks in Hydrology (2000) Artificial neural network in hydrology, part I and II. J Hydraul Eng 5(2):115–137 Asghari MA, Nadiri AA, Fijani E (2010) Spatial prediction of fluorideconcentration using artificial neural networks and geostatic models. J Water Soil Sci 19(1):129–145 Barai SV, Dikshit AK, Sharma S (2007) Neural network models for air quality prediction: a comparative study. Soft Comput Ind Appl 39: 290–305 Bardossy A, Disse M (1993) Fuzzy rule-based models for infiltration. Water Resour Res 29(2):373–382 Beeldens A (2006) An environmental friendly solution for air purification and self-cleaning effect: the application of TiO2 as photocatalyst in concrete. In: Proceedings of Transport Research Arena, Göteborg, Belgian Road Research Centre, Sweden Beeldens A (2008) Air purification by pavement blocks: final results of the research at the BRRC. In: Transport Research Arena Europe, Ljubljana Bengtsson N, Castellote M (2010) Photocatalytic activity for NO degradation by construction materials: parametric study and multivariable correlations. J Adv Oxid Technol 13(3):341–349 Berdahl P, Akbari H (2008) Evaluation of titanium dioxide as a photocatalyst for removing air pollutants. California Energy Commission. PIER Energy-Related Environmental Research Program Brunelli U, Piazza V, Pignato L, Sorbello F, Vitabile S (2007) Two-days ahead prediction of daily maximum concentrations of SO2, O3, PM10, NO2, CO in the urban area of Palermo, Italy. Atmos Environ 41:2967–2995 Cai M, Yin Y, Xie M (2009) Prediction of hourly air pollutant concentrations near urban arterials using artificial neural network approach. Transp Res D 14:32–41 Carnevale C, Finzi G, Pisoni E, Volta M (2009) Neuro-fuzzy and neural network systems for air quality control. Atmos Environ 43:4811– 4821 Chen M, Chu J-W (2011) NOx photocatalytic degradation on active concrete road surface—from experiment to real-scale application. J Clean Prod 19(11):1266–1272 Chen MS, Wang SW (1999) Fuzzy clustering analysis for optimizing fuzzy membership functions. Fuzzy Sets Syst 103(2):239–254 Dylla H, Hassan MM (2012a) Characterization of nanoparticle release during construction of photocatalytic pavements using engineered nanoparticles. J Nanoparticle Res 14:4 Dylla H, Hassan MM, Osborn D (2012b) Field evaluation of photocatalytic concrete pavements’ ability to remove nitrogen oxides. J Transp Res Rec 2290:154–160 Dylla H, Asadi S, Hassan M (2013) Evaluating photocatalytic asphalt pavement effectiveness in real world environments through developing models: a statistical and kinetic study. In: 88th American Asphalt Pavement Technology Annual Meeting, Denver, CO
Environ Sci Pollut Res Fujishima A, Rao TN, Tryk DA (2000) Titanium dioxide photocatalysis. J Photochem Photobiol C: Photochem Rev 1(1):1–21 Gautam AK, Chelani AB, Jain VK, Devotta S (2008) A new scheme to predict chaotic time series of air pollutant concentrations using artificial neural network and nearest neighbor searching. Atmos Environ 42:4409–4417 Hassan M, Mohammad L, Asadi S, Dylla H, Cooper S (2012) Sustainable photocatalytic asphalt pavements for mitigation of nitrogen oxide and sulfur dioxide vehicle emissions. J Mater Civ Eng 25(3):365–371 He F, Ma C (2010) Modeling greenhouse air humidity by means of artificial neural network and principal component analysis. Comput Electron Agric 71:S19–S23. doi:10.1016/j.compag.2009. 07.011 Hooyberghs J, Mensink C, Dumont G, Fierens F, Brasseur O (2005) A neural network forecast for daily average PM10 concentrations in Belgium. Atmos Environ 39:3279–3289 Karppinen A, Kukkonen J, Elolähde T, Konttinen M, Koskentalo T, Rantakrans E (2000) A modelling system for predicting urban air pollution: model description and applications in the Helsinki metropolitan area. Atmos Environ 34:3723–3733 Li L, Qian C (2009) A lab study of photo-catalytic oxidation and removal of nitrogen oxides in vehicular emissions and its fieldwork on Nanjin no.3 bridge of Yangtze River. J Pavement Resour Technol 2(5):218–222 Lin CH, Wu YL (2003) Semi-statistical model for evaluating the effects of source emissions and meteorological effects on daily average NOx concentrations in south Taiwan. Atmos Environ 37:2051–2059 Lin J, Yu D (2008) Traffic-related air quality assessment for open road tolling highway facility. J Environ Manag 88:962–969 Maggos T, Bartiz J, Liakou M, Gobin C (2007) Photocatalytic degradation of NOx gases using TiO2-containing paint: A real scale study. J Hazard Mater 146:668–673 Maggos T, Plassais A, Bartzis JG, Vasilakos C, Moussiopoulos A, Bonafous L (2008) Photocatalytic degradation of NOx in a pilot street canyon configuration using TiO2-mortar panels. Environ Monit Assess 136:35–44 Maier HR, Dandy GC (2000) Neural network for the prediction and forecasting water resources variables: a review of modeling issues and applications. Environ Model Softw 15(1):101–124 Maier HR, Jain A, Dandy GC, Sudheer KP (2010) Methods used for the development of neural networks for the prediction of water resource variables in river systems: current status and future directions. Environ Model Softw 25(8):891–909 Moussiopoulos N, Barmpas Ph, Ossanlis I, Bartiz J (2008) Comparison of numerical and experimental results for the evaluation of the depollution effectiveness of photocatalytic coverings in street canyons. 13 (3):357–368 Nadiri AA, Chitsazan N, Tsai F, Moghaddam A (2013a) Bayesian artificial intelligence model averaging for hydraulic conductivity estimation. J Hydrol Eng 19(3):520–532. doi:10.1061/(ASCE)HE.19435584.0000824 Nadiri AA, Fijani E, Tsai F, Moghaddam AA (2013b) Supervised committee machine with artificial intelligence for prediction of fluoride concentration. J Hydroinformatics 15(4):1474–1490 Nayak PC, Sudheer KP, Rangan DM, Ramasastri KS (2004) A neurofuzzy computing technique for modeling hydrological time series. J Hydrol 291(1–2):52–66
Nourani V, Asgharimogaddam A, Nadiri AA (2008) An ANN-based model for spatiotemporal groundwater level forecasting. Hydrol Process 22:5054–5066 Palani S, Liong SY, Tkalich P (2008) An ANN application for water quality forecasting. Mar Pollut Bull 56(9):1586–1597 Paschalidou AK, Karakitsios S, Kleanthous S, Kassomenos PA (2011) Forecasting hourly PM10 concentration in Cyprus through artificial neural networks and multiple regression models: implications to local environmental management. Environ Sci Pollut Res 18(2): 316–327 Patricio P, Jorge R (2006) An integrated neural network model for PM10 forecasting. Atmos Environ 430:2845–2851 Pires JCM, Gonçalves B, Azevedo FG, Carneiro AP, Rego N, Assembleia AJB, Lima JFB, Silva PA, Alves C, Martins FG (2012) Optimization of artificial neural network models through genetic algorithms for surface ozone concentration forecasting. Environ Sci Pollut Res 19:3228–3234 Pulido-Calvo I, Gutiérrez-Estrada JC (2009) Improved irrigation water demand forecasting using a soft-computing hybrid model. Biosyst Eng 102(2):202–218 Salazar-Ruiz E, Ordieres JB, Vergara EP, Capuz-Rizo SF (2008) Development and comparative analysis of tropospheric ozone prediction models using linear and artificial intelligence-based models in Mexicali, Baja California (Mexico) and Calexico, California (US). Environ Model Softw 23(8):1056–1069. doi:10.1016/j. envsoft.2007.11.009 Sharma N, Chaudhry KK, Chalapati Rao CV (2004) Vehicular pollution prediction modeling: a review of highway dispersion models. Transp Rev 24(4):409–435 Sleiman M, Conchon P, Ferronato C, Chovelon JM (2009) Photocatalytic oxidation of toluene at indoor air levels (ppbv): towards a better assessment of conversion, reaction intermediates and mineralization. Appl Catal B Environ 86(3–4):159–165 Sousa S, Martins F, Alvimferraz M, Pereira M (2007) Multiple linear regression and artificial neural networks based on principal components to predict ozone concentrations. Environ Model Softw 22(1): 97–103. doi:10.1016/j.envsoft.2005.12.002 Sugeno M, Yasukawa T (1993) A fuzzy logic-based approach to qualitative modelling. IEEE Trans Fuzzy Syst 1(1):7–31 Takagi T, Sugeno M (1985) Fuzzy identification of systems and its application to modeling and control. IEEE Trans Syst Man Cybern 15(1):116–132 Venturini L, Bacchi M (2009) Research, design, and development of a photocatalytic asphalt pavement. Proceedings of 2nd International Conference on Environmentally Friendly Roads. ENVIROAD, Warsaw, Poland Yetilmezsoy K, Abdul-Wahab S (2012) A prognostic approach based on fuzzy-logic methodology to forecast PM10 levels in Khaldiya residential area, Kuwait. Aerosol Air Qual Res 12:1217–1236 Yildirim Y, Bayramoglu M (2006) Adaptive neuro-fuzzy based modelling for prediction of air pollution daily levels in city of Zonguldak. Chemosphere 63:1575–1582 Zhao J, Yang X (2003) Photocatalytic oxidation of indoor air purification: a literature review. Build Environ 38:645–654 Zounemat-Kermani M, Teshnehlab M (2008) Using adaptive neurofuzzy inference system for hydrological time series prediction. Appl Soft Comput 8(2):928–936
RESEARCH ARTICLE
Artificial intelligence modeling to evaluate field performance of photocatalytic asphalt pavement for ambient air purification Somayeh Asadi & Marwa Hassan & Ataallah Nadiri & Heather Dylla
Received: 21 August 2013 / Accepted: 21 March 2014 # Springer-Verlag Berlin Heidelberg 2014
Abstract In recent years, the application of titanium dioxide (TiO2) as a photocatalyst in asphalt pavement has received considerable attention for purifying ambient air from trafficemitted pollutants via photocatalytic processes. In order to control the increasing deterioration of ambient air quality, urgent and proper risk assessment tools are deemed necessary. However, in practice, monitoring all process parameters for various operating conditions is difficult due to the complex and non-linear nature of air pollution-based problems. Therefore, the development of models to predict air pollutant concentrations is very useful because it can provide early warnings to the population and also reduce the number of measuring sites. This study used artificial neural network (ANN) and neuro-fuzzy (NF) models to predict NOx concentration in the air as a function of traffic count (Tr) and climatic conditions including humidity (H), temperature (T), solar
Responsible editor: Michael Matthies S. Asadi Department of Civil and Architectural Engineering, Texas A&M University-Kingsville, MSC 194, 700 University Blvd, Kingsville, TX 7836, USA e-mail: [email protected] M. Hassan (*) Department of Construction Management, Louisiana State University, 3130 A PFT Hall, Baton Rouge, LA 70803, USA e-mail: [email protected] A. Nadiri University of Tabriz, 29 Bahman Boulevard, Tabriz, East Azarbaijan, Iran e-mail: [email protected] H. Dylla Department of Construction Management, Louisiana State University, Baton Rouge, LA 70803, USA e-mail: [email protected]
radiation (S), and wind speed (W) before and after the application of TiO2 on the pavement surface. These models are useful for modeling because of their ability to be trained using historical data and because of their capability for modeling highly non-linear relationships. To build these models, data were collected from a field study where an aqueous nano TiO2 solution was sprayed on a 0.2-mile of asphalt pavement in Baton Rouge, LA. Results of this study showed that the NF model provided a better fitting to NOx measurements than the ANN model in the training, validation, and test steps. Results of a parametric study indicated that traffic level, relative humidity, and solar radiation had the most influence on photocatalytic efficiency. Keywords Artificial neural network . Neuro-fuzzy . Nitrogen oxides . Titanium dioxide . Photocatalytic pavement
Introduction Photocatalytic pavements are currently being investigated as a potential solution to remove air pollutants. These pavements consist of a surface layer of titanium dioxide (TiO2) nanoparticles, which can reduce or oxidize both organic and inorganic particles that are absorbed into active sites when irradiated by UV light (Beeldens 2006). Several laboratory studies have shown the effectiveness of photocatalytic pavements; however, understanding the efficacies in real-world environment remains a challenge (Berdahl and Akbari 2008). Previous researchers have shown that the efficiency of photocatalytic pavement depends on environmental factors including the concentration of pollutants, relative humidity, temperature, UV intensity, and wind speed, which complicate the interpretation of field test results (Sleiman et al. 2009; Venturini and Bacchi 2009; Bengtsson and Castellote 2010; Hassan et al. 2012).
Environ Sci Pollut Res
Two techniques have been used to evaluate the photocatalytic degradation in the field. The first is to measure the reduction of NOx concentration directly by measuring the ambient air pollution concentrations and the second is to measure the reduction indirectly by measuring the byproducts created from the degradation process (Dylla et al. 2012b). Using these two techniques, several approaches have been used in order to study photocatalytic efficiency in the field (Venturini and Bacchi 2009; Beeldens 2008; Maggos et al. 2007, 2008; Li and Qian 2009; Chen and Chu 2011). A review of the literature reveals that several approaches are available for predicting NOx concentrations in the air. One approach is to create a model to predict the concentrations of NOx in the air for a given area without photocatalytic pavement and an area with photocatalytic pavement. There are several roadway microenvironments pollution dispersion models such as Gaussian plume dispersion, computational fluid dynamics (CFD), atmospheric box, and statistical models (Lin and Yu 2008). Both statistical and computational fluid dynamic models have been developed to characterize photocatalytic environments (Dylla et al. 2013; Moussiopoulos et al. 2008). Among these, statistical models are simple and do not require additional laboratory kinetic studies to predict photocatalytic efficiency. Many scientific investigations have been conducted to predict different air pollutants’ concentrations using statistical tools including regression, multi-regression, and artificial neural networks (Karppinen et al. 2000; Lin and Wu 2003; Sharma et al. 2004). Using these tools, the pollution concentrations are estimated by statistical relationship between various factors collected at a particular receptor (Sharma et al. 2004). Regression statistics previously used by authors to evaluate the efficiency of photocatalytic pavements provided fair results. The developed model considered different factors such as solar radiation, relative humidity, wind speed, traffic, and temperature (Dylla et al. 2013). The results of the statistical models showed a good agreement between the predicted and observed data but some points had high errors. Statistical models may be able to obtain a relationship between the input and the output variables without describing the causes and effects in the formation of pollutants. However, they are not able to evaluate the intrinsic non-linear nature of the problem and to predict short-term pollution levels (Agirre-Basurko et al. 2006; Barai et al. 2007; Akkoyunlu et al. 2010; Yetilmezsoy and Abdul-Wahab 2012). To overcome this problem, mathematical models have been used as vital tools in both design and operation especially in cases which deal with high dimensional data such as air quality. A review of literature reveals that several artificial intelligence (AI) models are available for predicting the air pollutant concentrations (He and Ma 2010; Sousa et al. 2007; Abdul-Wahab and Al Alawi 2002; Carnevale et al. 2009; Pires et al. 2012). Over the last four decades, different AI models
including artificial neural network (ANN) and neuro-fuzzy (NF) have been proposed as alternatives to traditional statistical models. AI models can be used in the modeling of different real-life processes in environmental engineering due to their predictive capabilities and non-linear characteristics (Hydrology 2000; Maier and Dandy 2000); (Asghari et al. 2010; Maier et al. 2010; Nadiri et al. 2013a, b). Using AI models, some studies predict the concentrations of different air pollutions such as carbon monoxide (CO) (Gautam et al. 2008; Cai et al. 2009), particles measuring 10 μm or less (PM10) (Hooyberghs et al. 2005; Patricio and Jorge 2006; Paschalidou et al. 2011), ozone (Sousa et al. 2007; AbdulWahab and Al Alawi 2002; Pires et al. 2012; Salazar-Ruiz et al. 2008) and sulfur monoxide (SO) (Brunelli et al. 2007). Sousa et al. (2007) constructed multiple linear regression and artificial neural networks models based on principal components to predict ozone concentrations. A new methodology based on feedforward artificial neural networks using principal components as inputs was used to predict next day hourly ozone concentrations. Then, they compared the results of the developed model with multiple linear regressions, feedforward artificial neural networks based on the original data, and also with principal component regression. They found that the use of principal components as inputs improved both model prediction by reducing their complexity and eliminating data collinearity. Later, He and Ma (He and Ma 2010) used a back-propagation neural network based on principal component analysis to model the internal greenhouse humidity during winter season in North China. The results of this model were compared with the stepwise regression model. The results indicated that the stepwise regression method was less accurate than the back-propagation neural network model. Each AI method has its own advantages. The fuzzy models tend to be robust to parameter changes, and are also tolerant to imprecision and uncertainty (Bardossy and Disse 1993). While the ANN model represents non-linear relationships and learns these relationships directly from the data being modeled (Palani et al. 2008). Obviously, the NF model takes advantage of the fuzzy logic (FL) and ANN in modeling. The NF model combines the advantages of fuzzy systems—deal with explicit knowledge—ANN deals with implicit knowledge. This NF model, by taking advantage of ANN and fuzzy logic methods, have shown high capability to estimate air pollution (Yildirim and Bayramoglu 2006; Carnevale et al. 2009). The objective of this study was the construction of AI models which give the possibility of predicting the NOx concentration in the air for a given area without photocatalytic pavement and an area with photocatalytic pavement as a function of traffic count (Tr) and climatic conditions including humidity (H), temperature (T), solar radiation (S), and wind speed (W). This model relates the NOx concentration to both
Environ Sci Pollut Res
meteorological variables as well as to traffic volume and help to understand the effect of each parameter on the efficiency of photocatalytic asphalt pavement to reduce NOx concentration in the air. The AI models are useful for modeling because of their ability to be trained using historical data and because of their capability for modeling highly non-linear relationships.
Field site description Data used in the models were obtained from a field study located in Baton Rouge, Louisiana, where 0.2 miles of asphalt pavement was sprayed with an aqueous TiO2 solution (Dylla et al. 2013). The photocatalytic spray coat was a mixture of TiO2 anatase nanorods 12 nm wide by 75 nm long suspended in an aqueous liquid at 2 % by volume (Dylla and Hassan 2012a). Before application, the roadway was cleared of any debris. A primer was applied before the photocatalytic coat. A distributor truck was used in the application process. Mounted on the back of the truck, a spray bar fitted with nozzles distributed TiO2 water-based solution at the specified application rate, 16.1 to 21.5 ml/m2. The application rate was adjusted by altering the truck speed. Further, the nozzles had electrostatic precipitators to separate the TiO2 nanoparticles suspended in the aqueous solution and to ensure a more even coverage. Equipment used for field data collection was housed in an air-conditioned trailer. A Thermo 42i NOx analyzer was used for monitoring NO, NO2, and NOx concentrations. The NOx analyzer was calibrated in accordance to EPA calibration procedures using the gas phase titration (GPT) alternative. A Thermo 146i gas calibrator was used for calibration of the NOx analyzer. The NOx analyzer was calibrated at five different spans for NO and four different ozone settings for NO2 to confirm linearity and ozone converter efficiency. In addition, the Department of Environmental Quality (DEQ) air monitoring station is located next to the airconditioned trailer. After calibration, the NOx analyzer was connected to ambient air at the pavement level using a stainless steel pipe placed in the middle lane and was covered with a protective bump.
Table 1 Parameters range of variation Tr (V/5 min) S (W/m2) W (m/s) H (%) T (°C)
To monitor climatic conditions at the site, a Davis 6152 Wireless Vantage Pro weather station was installed in the field which recorded and stored meteorological data including humidity, ambient air temperature, wind speed, and solar radiation continuously each minute. A portable Peek ADR-1000 traffic counter capable of counting vehicles per 5-min intervals was employed. Traffic counter was installed to count the number of vehicles per 5-min per lane. In order to consider the effect of climatic conditions on the performance of photocatalytic activities, data were collected under different climatic conditions.
Parameters selection The goal of this study was to predict NOx concentration in the air as a function of different parameters before and after application of TiO2 on the asphalt pavement. The selected parameters and their range of variation are shown in Table 1. These parameters were selected based on laboratory test results (Hassan et al. 2012). Laboratory results established a relationship between NOx concentration and humidity, UV intensity, temperature, and flow rate.
Methodology Artificial intelligence methods Different non-linear AI models may be used for air pollutants’ prediction. In this study, ANN and NF models were adopted for prediction of NOx concentrations. The main goal of the models is to predict NOx concentration and investigate the sensitivity of NOx concentration to related variables defined as input data. The input data for the AI models for the case study in Baton Rouge, LA, are shown in Table 1. A brief description of the two AI models is provided in the following sections. ANN An ANN model was developed to predict NOx concentration in this study. ANN is a universal approximator to surrogate
Description
Min
Max
Mean
Standard deviation
Traffic count Solar radiation Wind speed Relative humidity Temperature
0 0 0 32 11
36 858 4 94 30
15 123 0.8 77 20.4
13.4 197.2 0.91 14.01 4.11
Environ Sci Pollut Res
complex systems (Hydrology 2000; Maier et al. 2010). Neural networks are composed by simple connected elements (neurons) operating in parallel. Each neuron includes a nonlinear function relating inputs and outputs (activation function). The most widely used neural network is the multi-layer perceptron (MLP) network (Hydrology 2000; Nourani et al. 2008) that consists of an input layer, a hidden layer, and an output layer. This study considers one hidden layer for the MLP network as shown in Fig. 1a. The structure of the ANN model, applied to predict NOx concentration, contains five neurons in the input layer pertain to the input data (Tr, H, S, W, and T), one neuron in the output layer and four neurons in the hidden layer that was determined via trial and error. The normalized input signal spreads through the network in a forward direction via connections between neurons. Incoming signals are linearly combined and converted to outgoing signals. The signal conversion is conducted by assigned activation functions. The mathematical expression for a three-layer feedforward ANN is given as (Hydrology 2000; Nadiri et al. 2013a): ! X Oj ¼ f 1 bj þ W ji I i ð1Þ
Wji and Wkj are weights that control the strength of connections between two layers, and the biases bj and bk are used to adjust the mean value for input layer and hidden layer, respectively. The activation function for the hidden layer is typically a continuous and bounded non-linear transfer function such as sigmoid and log sigmoid functions. The activation function for the output layer is usually a linear function. In this study, hyperbolic tangent sigmoid and linear function was selected for f1 and f2, respectively. The output Ok of the ANN is the NOx concentration. In the ANN training step, a supervised learning algorithm was used to estimate the weights Wji and Wkj (Hydrology 2000). To construct ANN model and to prevent over-fitting, a cross validation method procedure is usually recommended (Hydrology 2000). The cross validation involves partitioning a sample of data into complementary subsets, performing the analysis on one subset (training set), and validating the analysis on the other subsets (called the validation set and testing set). Therefore, the measured data were randomly divided into three groups, training (70 %), validation (15 %), and testing (15 %). NF
i
Ok ¼ N Ox ¼ f 2 bk þ
X
! W kj O j
ð2Þ
j
where f1 and f2 are the activation functions for the hidden and output layers, respectively, Ii is the ith input, Oj is the jth output, Fig. 1 Schematics of a ANN structure and b NF structure
Neuro-fuzzy (NF), as shown in Fig. 1b, combines the advantage of the FL and ANN methods by applying various learning techniques developed in the ANN literature to optimize the FL model (Nayak et al. 2004). Based on the fuzzy-set theory, an element of the world belongs to a set, specifying a feature of the element (linguistic variable), with a value ranging from 0 to 1 according to a function (membership function). Fuzzy
Environ Sci Pollut Res
sets have also ambiguous boundaries and gradual transitions between defined sets, which are appropriate to deal with the nature of uncertainty in the system and human errors (PulidoCalvo and Gutiérrez-Estrada 2009). Following the conventional logic, an inference system based on the rules in the form of “if-then” is formulated for the fuzzy logic (Carnevale et al. 2009). The first step for preparation of the NF model to predict NOx concentration is construction of a FL model. A FL model developed by Takagi and Sugeno (Takagi and Sugeno 1985; Sugeno and Yasukawa 1993) termed as TS-FL model, was adopted. In this model, the output membership functions are either constant (zero order) or linear (first order). Input clusters and output membership functions are extracted by a clustering process. An effective subtractive clustering method (Chen and Wang 1999) for TS-FL modeling was used for the extraction of clusters and fuzzy “if-then” rules. The important parameter in subtractive clustering, which controls number of clusters and fuzzy “if-then” rules, is cluster radius that takes values between 0 and 1. Decreasing the cluster radius increases the number of clusters and leads to smaller clusters. In contrast, a large cluster radius produces large clusters in the data and results in few rules (Aqil et al. 2007). A TS-FL model without sufficient number of rules cannot cover the entire domain and with too many rules, becomes complicated and may lead to low performance. Therefore, the cluster radius had to be optimized via trial and error to have a suitable number of clusters and rules. Searching for the optimal cluster radius can be accomplished by systematically varying cluster radius value from 0 to 1 until a minimal root mean squared error (RMSE) is achieved. To apply the first-order TS-FL model, a generalized Gaussian function was used to develop membership functions for the five input data, see Table 1. Each input was clustered into nine classes. Then, based on membership functions of the five input data, a set of fuzzy if-then rules were developed to linearly aggregate the input data as the output. Since the subtractive clustering method was used, the number of rules was equal to the number of clusters. For NOx prediction in this study, a fuzzy “if-then” rule i can be expressed as: Rule i: if (Tr belongs to MFiTr), (H belongs to MFiH), (S belongs to MFiS), (W belongs to MFiW), and (T belongs to MFiT), then ðN Ox Þi ¼ mi Tr þ ni H þ pi S þ qi W þ zi T þ ci
ð3Þ
where NOx is the output of rule i, MFiTr denotes the membership function of the ith cluster of input Tr, MFiTr represents the membership function of the ith cluster of input H, and so forth. mi, pi, qi, zi, and ci are coefficients to be determined by linear least-squares estimation. The second step of NF construction
is adopting a five-layer MLP network to optimize coefficients of linear membership function of output and also the Gaussian membership function parameters of input. The NF structure is shown in Fig. 1b. In each layer, the operations are as follows: Layer 1: Generate membership function of input data. The output of neuron i is defined by O1i ¼ μji ðX Þ
ð4Þ
where j is the number of input and i denotes the membership function index. X={Tr, H, S, W, T} represents a set of input. μji(X) is a fuzzy set associated with neuron i which is a membership function. Layer 2: Calculate firing strength Wi for the ith rule via multiplication via “and” operator: O2i ¼ wi ¼ μ1i ðX Þμ2i ðX Þμ3i ðX Þμ4i ðX Þμ5i ðX Þð5Þ
Layer 3: Compute the normalized firing strengths for the ith neuron: wi O3i ¼ wi ¼ X wi
ð6Þ
i
Layer 4: Compute the contribution of the ith rule in the model output based on the first-order TS-FL method: O4i ¼ wi ðNOx Þi ¼ wi ðmi Tr þ ni H þ pi S þ qi W þ zi T þ ci Þ
ð7Þ
Layer 5: Calculate the final output as the weighted average of all rule outputs (aggregation): O5i ¼ N Ox
X
w i ðN O x Þi
ð8Þ
i
The NF parameters in Eq. 8 and the membership function parameters were optimized using a hybrid algorithm, which is a combination of the gradient decent and the leastsquares method (Aqil et al. 2007).
Environ Sci Pollut Res
Efficiency criteria
Results and analysis
Two statistical measures were used to evaluate the effectiveness of the ANN and NF models and their ability to make accurate predictions. The root mean square error (RMSE) is calculated by:
Artificial neural network
i¼1
ð9Þ
n
where Y i ; Yb i and n are the measured data, the predicted data, and the number of measurements, respectively. RMSE indicates the discrepancy between measured and predicted values. The R2, coefficient of determination, was also used, which is given by:
i R2 ¼ 1− X n
Y i −Yb
Y i −Y
2 ð10Þ
2
i
where Y is the mean of observed data. The best fit between measured and predicted values will give RMSE and R2 close to 0 and 1, respectively.
70
70
70
60
60
60
50
50
50
Predicted values
Predicted values
a
40 30 20 10
40 30 20 10
0
30 20
0 0 10 20 30 40 50 60 70
Actual values
0 10 20 30 40 50 60 70
Actual values
Actual values
40
40
40
30
30
30
Predicted values
b
40
10
0 0 10 20 30 40 50 60 70
Predicted values
Fig. 2 Comparison between the measured and predicted values of NOx concentrations for ANN in the train, test, and validation steps for a before TiO2 application and b after TiO2 application
Predicted values
n X
20
10
0
Predicted values
RMSE ¼
vffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi n uX 2 u u Y i −Yb j t
For the construction of both ANN models (i.e., before and after TiO2 application), the presented network in Fig. 1a was applied. The numbers of hidden nodes were determined to be four nodes through the trial and error method. The weights and biases in the ANN models were estimated by minimizing error using the Levenberg-Marquardt (LM) algorithm (Hydrology 2000). The first ANN model was developed to predict NOx concentration before application of TiO2. The R2 and RMSE of this model after 354 epochs of training were 0.96 and 3.1 ppb, respectively. For the validation and test steps, the R2 and RMSE were 0.95 and 3.3 ppb and 0.93 and 3.4 ppb, respectively. The second ANN model was developed with 372 epochs to predict NOx concentration after TiO2 application on the asphalt pavement. The RMSE for the training, validation, and test steps after the application of TiO2 were 3.1, 3.4, and 3.7 ppb, respectively. The R2 for the training, validation, and test were 0.96, 0.94, and 0.92, respectively. The scatter plot of measured and ANN estimated NOx concentration is shown in Fig. 2a and b. From these figures, it is noted that the data are concentrated around the solid line, which indicates a model with acceptable accuracy since the closer the distances of these data to the solid line are, the better the accuracy of the model.
20
10
0 0
10 20 30 40 Actual values
20
10
0 0
10 20 30 40 Actual values
0
10 20 30 40 Actual values
Environ Sci Pollut Res 70
70
60
60
60
50
50
50
40 30 20 10
Predicted values
70
Predicted values
a Predicted values
Fig. 3 Comparison between the measured and predicted values of NOx concentrations for NF in the train, test, and validation steps for a before TiO2 application and b After TiO2 application
40 30 20 10
0
0 0 10 20 30 40 50 60 70
Actual values
Actual values
40
40
30
30
30
20
10
Predicted values
40
Predicted values
Predicted values
20
0 10 20 30 40 50 60 70
Actual values
b
30
10
0 0 10 20 30 40 50 60 70
40
20
10
0
10
0 0
10
20
30
Actual values
Neuro-fuzzy The first step for construction of the two NF models (i.e., before and after TiO2 application) was the determination of optimal TS-FL clusters of five input data and rules. To develop an efficient TS-FL model with a reasonable number of fuzzy “if-then” rules, the cluster radius was optimized. Based on the minimum RMSE criteria for both models, a cluster radius of 0.5 and three rules were determined for the TS-FL model. In each rule, the parameters, mi, pi, qi, zi, and ci in the output membership function were estimated by the leastsquare error method. The second step was using ANN models to estimate parameters in the Gaussian membership functions and the coefficients of the model output (Eq. 8). A hybrid algorithm, which combines the least-squares method and the back-propagation gradient descent method, was employed (Zounemat-Kermani and Teshnehlab 2008). The first NF model was developed to predict the NOx concentration before TiO2 application. The R2 and RMSE of this model after five epochs of training were 0.98 and 1.7 ppb, respectively. The R2 and RMSE of 0.97 and 1.8 ppb were obtained in the validation step. The R2 and RMSE were 0.96 and 2.1 ppb, respectively, in the test step. The second NF model with five epochs was developed to predict NOx concentration after the application of TiO2. The RMSE for the training, validation, and testing steps after application of TiO2 were 2.1, 2.15, and 2.16 ppb, respectively. The coefficients of determination for the train, validation, and test steps were
40
20
0 0
10
20
30
40
0
10
Actual values
20
30
40
Actual values
0.97, 0.95, and 0.94, respectively. The measured and NF estimated NOx concentration scatter plots are shown in Fig. 3a, b. It can be observed that the data obtained from NF model are well correlated with the measured data points, which correspond to an excellent fit. The RMSE and the coefficients of determination for ANN and NF models are presented in Table 2. By taking advantage of FL and ANN models, the NF models provided a slightly Table 2 Fitting test parameters for ANN and NF models
Criterion
Step
Before application R2 Train RMSE R2 Validation RMSE R2 Test RMSE After application R2 Train RMSE R2 Validation RMSE R2 Test RMSE
Model ANN
NF
0.96 3.1
0.98 1.77
0.95 3.3 0.93 3.4
0.97 1.83 0.96 2.11
0.96 3.1 0.94 3.4 0.92 3.7
0.97 2.1 0.95 2.15 0.94 2.16
Environ Sci Pollut Res Fig. 4 Effect of traffic levels on NOx concentration
Concentration (ppb)
Before
After
35 30 25 20 15 10 5 0 Max
Mean
Min
Number of Vehicle ( V/5 min)
better fitting to NOx data than ANN model in the training, validation, and testing steps. Both models outperformed the statistical model initially used by the authors (Dylla et al. 2013). Parametric study The photocatalytic efficiency depends on several environmental factors and operating conditions, which makes it very complex to investigate. A parametric study was conducted using the NF models to study the effects of different parameters on NOx reduction efficiency. The NF model was used due to its considerable accuracy over ANN in the training, validation, and test steps. The objective of the parametric study was to evaluate the effect of each factor on the photocatalytic degradation of NOx. To achieve this objective, a sensitivity analysis was carried out in which each parameter was varied between the minimum and maximum level while the other parameters were kept constant at the mean value. The minimum, maximum, and mean values of each parameter are presented in Table 1.
Effect of traffic flow The first parameter investigated in this study was the traffic flow. Figure 4 shows a direct relationship between the number of vehicles per 5 min and NOx concentrations. It should be noted that in order to evaluate the effect of traffic flow on the NOx reduction based on the NF models, other parameters such as humidity, solar radiation, wind speed, and ambient temperature were kept constant at their mean values. As expected, increasing traffic flow increases NOx concentration as more pollutants are emitted to the environment. The developed model predicted that the TiO2 photocatalytic coating resulted in a significant reduction of NOx concentration at the different traffic levels. The model predicted approximately 60 % NOx reduction efficiency at the maximum number of vehicles per 5 min. Effect of relative humidity Figure 5 shows the effect of relative humidity on NOx concentrations per 5 min before and after the application of TiO2
Before
Fig. 5 Effect of humidity on NOx Concentration
After
Concentration (ppb)
25
20 15 10 5 0
Max
Mean Humidity
Min
Environ Sci Pollut Res
Before
Fig. 6 Effects of solar radiation on NOx concentration
After
Concentration (ppb)
25 20 15 10 5 0 Max
Mean
Min
Solar Radiation (W/m^2)
based on the NF model. As shown in this figure, increasing the relative humidity increases NOx concentrations in both models. Nevertheless, NOx concentrations before application are higher than after application at the different humidity levels indicating that the TiO2 photocatalytic layer is reducing NOx concentrations in ambient air. Figure 5 shows a clear trend, as the humidity increases, NOx concentration increases. This is mainly due to the excessive water vapor building on the TiO2 catalyst surface. Excessive water inhibits the photocatalytic reaction as it competes with pollutants for adsorption sites on the surface. This is in agreement with laboratory test results (Hassan et al. 2012).
Effect of solar radiation Photocatalytic oxidation necessitates UV light, part of the solar radiation spectrum, for the reaction to take place (Fujishima et al. 2000). The wavelength is the main factor influencing photocatalytic activity; in contrast, the strength of photon flux affects the rate of degradation (Fujishima et al.
2000; Zhao and Yang 2003). Therefore, an increase in UV irradiance is expected to increase photocatalytic NOx reduction efficiency. Increasing solar radiation from minimum (0 W/m2) to maximum level (858 W/m2) significantly increased the photocatalytic reaction rate, which resulted in higher NOx reduction efficiency (Fig. 6). Based on the NF models, at a solar radiation of 858 W/m2, NOx reduction efficiency per 5 min was approximately 77 %.
Effect of wind speed Figure 7 presents the effect of wind speed on NOx concentration per 5-min interval before and after application of TiO2. As shown in this figure, increasing wind speed decreases NOx concentration. This can be attributed to the fact that higher wind speed will help disperse the pollutants at the pavement level at a faster rate. Based on the model, at the maximum wind speed (4 m/s), NOx reduction efficiency per 5 min was approximately 42 % which is in agreement with laboratory results. It should be noted that higher wind speed would
Before
Fig. 7 Effects of a solar radiation and b wind speed on NOx Concentration
After
Concentration (ppb)
20 15 10
5 0 Max
Mean Wind Speed (m/s)
Min
Environ Sci Pollut Res
reduce the time available for the photocatalytic reaction to take place (Sleiman et al. 2009; Hassan et al. 2012).
Summary and conclusions This study improves the previous regression models developed by authors to assess and predict photocatalytic pavement’s field efficiency by employing artificial intelligence models. Both artificial neural network (ANN) and neurofuzzy (NF) models were used to predict NOx concentrations in the air as a function of traffic count (Tr) and climatic conditions including humidity (H), solar radiation (S), and wind speed (W) before and after the application of TiO2 on the pavement surface. Data were collected from a full-scale field study of a photocatalytic asphalt pavement located in Baton Rouge, LA. Results showed that the NF model provided better fitting to NOx measurements than ANN model in the training, validation, and testing steps. The R 2 and RMSE of NF model for the train, validation, and test steps were higher than ANN model. The R2 of the NF model was 0.98, 0.97, and 0.96, respectively for train, validation, and test steps before TiO2 application. The RMSE of the NF model was 1.77, 1.83, and 2.11 ppb, respectively, for train, validation, and test steps which were considerably lower than the RMSE obtained by ANN model. Results of a parametric study showed that the traffic level, relative humidity, ambient air temperature, solar radiation, and wind speed have significant effects on the performance of TiO2 in NOx degradation. Traffic level, relative humidity, and solar radiation had the most significant effects on photocatalytic NOx reduction. The increase in wind speed and relative humidity negatively affected the effectiveness of NOx reduction efficiency. However, the increase in UV light intensity improved NOx removal efficiency of the surface coating. The model predicted approximately 60 % NOx reduction efficiency at the maximum number of vehicles per 5 min, 77 % efficiency at solar radiation of 858 W/m2, 42 % efficiency at the maximum wind speed (4 m/s). Based on the results obtained in this study, further research is needed to assess the effects of additional factors including the impact of vehicle classification and vehicle activity. In addition, the durability of the photocatalytic coating in the field should be assessed. Acknowledgments This work was funded through a grant from the Gulf Coast Research Center for Evacuation and Transportation Resiliency. The authors would like to acknowledge PURETI for donating the materials needed for the construction of the field study and the support of Louisiana Transportation Research Center (LTRC) for granting access to their laboratory.
References Abdul-Wahab SA, Al Alawi SM (2002) Assessment and prediction of tropospheric ozone concentration levels using artificial neural networks. Environ Model Softw 17(3):219–228 Agirre-Basurko E, Ibarra-Berastegi G, Madariaga I (2006) Regression and multilayer perceptron-based models to forecast hourly O3 and NO2 levels in the Bilbao area. Environ Model Softw 21(4):430–446 Akkoyunlu A, Yetilmezsoy K, Erturk F, Oztemel E (2010) A neural network-based approach for the prediction of urban SO2 concentrations in the Istanbul metropolitan area. Int J Environ Pollut 40(4): 301–321 Aqil M, Kita I, Yano A, Nishiyama S (2007) Analysis and prediction of flow from local source in a river basin using a neuro-fuzzy modelling tool. J Environ Manag 85(1):215–223 ASCE Task Committee on Application of Artificial Neural Networks in Hydrology (2000) Artificial neural network in hydrology, part I and II. J Hydraul Eng 5(2):115–137 Asghari MA, Nadiri AA, Fijani E (2010) Spatial prediction of fluorideconcentration using artificial neural networks and geostatic models. J Water Soil Sci 19(1):129–145 Barai SV, Dikshit AK, Sharma S (2007) Neural network models for air quality prediction: a comparative study. Soft Comput Ind Appl 39: 290–305 Bardossy A, Disse M (1993) Fuzzy rule-based models for infiltration. Water Resour Res 29(2):373–382 Beeldens A (2006) An environmental friendly solution for air purification and self-cleaning effect: the application of TiO2 as photocatalyst in concrete. In: Proceedings of Transport Research Arena, Göteborg, Belgian Road Research Centre, Sweden Beeldens A (2008) Air purification by pavement blocks: final results of the research at the BRRC. In: Transport Research Arena Europe, Ljubljana Bengtsson N, Castellote M (2010) Photocatalytic activity for NO degradation by construction materials: parametric study and multivariable correlations. J Adv Oxid Technol 13(3):341–349 Berdahl P, Akbari H (2008) Evaluation of titanium dioxide as a photocatalyst for removing air pollutants. California Energy Commission. PIER Energy-Related Environmental Research Program Brunelli U, Piazza V, Pignato L, Sorbello F, Vitabile S (2007) Two-days ahead prediction of daily maximum concentrations of SO2, O3, PM10, NO2, CO in the urban area of Palermo, Italy. Atmos Environ 41:2967–2995 Cai M, Yin Y, Xie M (2009) Prediction of hourly air pollutant concentrations near urban arterials using artificial neural network approach. Transp Res D 14:32–41 Carnevale C, Finzi G, Pisoni E, Volta M (2009) Neuro-fuzzy and neural network systems for air quality control. Atmos Environ 43:4811– 4821 Chen M, Chu J-W (2011) NOx photocatalytic degradation on active concrete road surface—from experiment to real-scale application. J Clean Prod 19(11):1266–1272 Chen MS, Wang SW (1999) Fuzzy clustering analysis for optimizing fuzzy membership functions. Fuzzy Sets Syst 103(2):239–254 Dylla H, Hassan MM (2012a) Characterization of nanoparticle release during construction of photocatalytic pavements using engineered nanoparticles. J Nanoparticle Res 14:4 Dylla H, Hassan MM, Osborn D (2012b) Field evaluation of photocatalytic concrete pavements’ ability to remove nitrogen oxides. J Transp Res Rec 2290:154–160 Dylla H, Asadi S, Hassan M (2013) Evaluating photocatalytic asphalt pavement effectiveness in real world environments through developing models: a statistical and kinetic study. In: 88th American Asphalt Pavement Technology Annual Meeting, Denver, CO
Environ Sci Pollut Res Fujishima A, Rao TN, Tryk DA (2000) Titanium dioxide photocatalysis. J Photochem Photobiol C: Photochem Rev 1(1):1–21 Gautam AK, Chelani AB, Jain VK, Devotta S (2008) A new scheme to predict chaotic time series of air pollutant concentrations using artificial neural network and nearest neighbor searching. Atmos Environ 42:4409–4417 Hassan M, Mohammad L, Asadi S, Dylla H, Cooper S (2012) Sustainable photocatalytic asphalt pavements for mitigation of nitrogen oxide and sulfur dioxide vehicle emissions. J Mater Civ Eng 25(3):365–371 He F, Ma C (2010) Modeling greenhouse air humidity by means of artificial neural network and principal component analysis. Comput Electron Agric 71:S19–S23. doi:10.1016/j.compag.2009. 07.011 Hooyberghs J, Mensink C, Dumont G, Fierens F, Brasseur O (2005) A neural network forecast for daily average PM10 concentrations in Belgium. Atmos Environ 39:3279–3289 Karppinen A, Kukkonen J, Elolähde T, Konttinen M, Koskentalo T, Rantakrans E (2000) A modelling system for predicting urban air pollution: model description and applications in the Helsinki metropolitan area. Atmos Environ 34:3723–3733 Li L, Qian C (2009) A lab study of photo-catalytic oxidation and removal of nitrogen oxides in vehicular emissions and its fieldwork on Nanjin no.3 bridge of Yangtze River. J Pavement Resour Technol 2(5):218–222 Lin CH, Wu YL (2003) Semi-statistical model for evaluating the effects of source emissions and meteorological effects on daily average NOx concentrations in south Taiwan. Atmos Environ 37:2051–2059 Lin J, Yu D (2008) Traffic-related air quality assessment for open road tolling highway facility. J Environ Manag 88:962–969 Maggos T, Bartiz J, Liakou M, Gobin C (2007) Photocatalytic degradation of NOx gases using TiO2-containing paint: A real scale study. J Hazard Mater 146:668–673 Maggos T, Plassais A, Bartzis JG, Vasilakos C, Moussiopoulos A, Bonafous L (2008) Photocatalytic degradation of NOx in a pilot street canyon configuration using TiO2-mortar panels. Environ Monit Assess 136:35–44 Maier HR, Dandy GC (2000) Neural network for the prediction and forecasting water resources variables: a review of modeling issues and applications. Environ Model Softw 15(1):101–124 Maier HR, Jain A, Dandy GC, Sudheer KP (2010) Methods used for the development of neural networks for the prediction of water resource variables in river systems: current status and future directions. Environ Model Softw 25(8):891–909 Moussiopoulos N, Barmpas Ph, Ossanlis I, Bartiz J (2008) Comparison of numerical and experimental results for the evaluation of the depollution effectiveness of photocatalytic coverings in street canyons. 13 (3):357–368 Nadiri AA, Chitsazan N, Tsai F, Moghaddam A (2013a) Bayesian artificial intelligence model averaging for hydraulic conductivity estimation. J Hydrol Eng 19(3):520–532. doi:10.1061/(ASCE)HE.19435584.0000824 Nadiri AA, Fijani E, Tsai F, Moghaddam AA (2013b) Supervised committee machine with artificial intelligence for prediction of fluoride concentration. J Hydroinformatics 15(4):1474–1490 Nayak PC, Sudheer KP, Rangan DM, Ramasastri KS (2004) A neurofuzzy computing technique for modeling hydrological time series. J Hydrol 291(1–2):52–66
Nourani V, Asgharimogaddam A, Nadiri AA (2008) An ANN-based model for spatiotemporal groundwater level forecasting. Hydrol Process 22:5054–5066 Palani S, Liong SY, Tkalich P (2008) An ANN application for water quality forecasting. Mar Pollut Bull 56(9):1586–1597 Paschalidou AK, Karakitsios S, Kleanthous S, Kassomenos PA (2011) Forecasting hourly PM10 concentration in Cyprus through artificial neural networks and multiple regression models: implications to local environmental management. Environ Sci Pollut Res 18(2): 316–327 Patricio P, Jorge R (2006) An integrated neural network model for PM10 forecasting. Atmos Environ 430:2845–2851 Pires JCM, Gonçalves B, Azevedo FG, Carneiro AP, Rego N, Assembleia AJB, Lima JFB, Silva PA, Alves C, Martins FG (2012) Optimization of artificial neural network models through genetic algorithms for surface ozone concentration forecasting. Environ Sci Pollut Res 19:3228–3234 Pulido-Calvo I, Gutiérrez-Estrada JC (2009) Improved irrigation water demand forecasting using a soft-computing hybrid model. Biosyst Eng 102(2):202–218 Salazar-Ruiz E, Ordieres JB, Vergara EP, Capuz-Rizo SF (2008) Development and comparative analysis of tropospheric ozone prediction models using linear and artificial intelligence-based models in Mexicali, Baja California (Mexico) and Calexico, California (US). Environ Model Softw 23(8):1056–1069. doi:10.1016/j. envsoft.2007.11.009 Sharma N, Chaudhry KK, Chalapati Rao CV (2004) Vehicular pollution prediction modeling: a review of highway dispersion models. Transp Rev 24(4):409–435 Sleiman M, Conchon P, Ferronato C, Chovelon JM (2009) Photocatalytic oxidation of toluene at indoor air levels (ppbv): towards a better assessment of conversion, reaction intermediates and mineralization. Appl Catal B Environ 86(3–4):159–165 Sousa S, Martins F, Alvimferraz M, Pereira M (2007) Multiple linear regression and artificial neural networks based on principal components to predict ozone concentrations. Environ Model Softw 22(1): 97–103. doi:10.1016/j.envsoft.2005.12.002 Sugeno M, Yasukawa T (1993) A fuzzy logic-based approach to qualitative modelling. IEEE Trans Fuzzy Syst 1(1):7–31 Takagi T, Sugeno M (1985) Fuzzy identification of systems and its application to modeling and control. IEEE Trans Syst Man Cybern 15(1):116–132 Venturini L, Bacchi M (2009) Research, design, and development of a photocatalytic asphalt pavement. Proceedings of 2nd International Conference on Environmentally Friendly Roads. ENVIROAD, Warsaw, Poland Yetilmezsoy K, Abdul-Wahab S (2012) A prognostic approach based on fuzzy-logic methodology to forecast PM10 levels in Khaldiya residential area, Kuwait. Aerosol Air Qual Res 12:1217–1236 Yildirim Y, Bayramoglu M (2006) Adaptive neuro-fuzzy based modelling for prediction of air pollution daily levels in city of Zonguldak. Chemosphere 63:1575–1582 Zhao J, Yang X (2003) Photocatalytic oxidation of indoor air purification: a literature review. Build Environ 38:645–654 Zounemat-Kermani M, Teshnehlab M (2008) Using adaptive neurofuzzy inference system for hydrological time series prediction. Appl Soft Comput 8(2):928–936
