A NEURAL NETWORK NOISE PREDICTION MODEL FOR TEHRAN URBAN ROADS

. Over the last decades, the number of motor vehicles has increased dramatically in Iran, where different traffic characteristics and urban structures are notable. In the present study, a multilayer perceptron neural network model trained with the Levenberg-Marquardt algorithm was used for predicting the equivalent sound level (L Aeq ) originating from traffic. Fifty-one samples were collected from different areas of Tehran. Input parameters consisted of total traffic volume per hour, average speed of vehicles, percentage of each category of vehicles, road gradient, density of buildings around the road section and a new parameter named “Building Reflection Factor”. These data were randomly used with 80, 10 and 10 percentiles respectively for training, validation and testing of the Artificial Neural Network (ANN). Results yielded by the ANN model were compared with field measurement data, a proposed regression model and some classical well-known models. Our study indicated that the prediction error of the neural network model was much less than that of the regression model and other classical models. Moreover, a statistical t-test was applied for evaluating the goodness-of-fit of the proposed model and proved that the neural network model is highly efficient in estimating road traffic noise levels.


Introduction
In the last decades, the growth in population and vehicles per capita that has led to an increase in urban trips has made our world noisier than ever before. According to WHO reports, traffic noise alone is harmful to the health of almost every third person in the WHO European Region (Euro WHO 2015). Living in a noise-polluted area can cause many short and long-term health problems such as sleep disturbance, as reported by the WHO. Cardiovascular diseases like hypertension and other mental and physical problems are the outcomes of being exposed to excessive noise levels (Euro WHO 2015) so that a vast number of research papers are directed to delineate this issue (Babisch et al. 2013;Brink 2011;Caciari et al. 2013;Fyhri, Klboe 2009;Pirrera et al. 2010). Therefore, a lot of research was conducted to investigate the impact of traffic noise pollution on the environment and the methods of predicting, reducing or controlling this phenomenon (Johnson, Saunders 1968;Delany et al. 1976;Pamanikabud, Vivitjinda 2002;Paulauskas, Klimas 2011;Dintrans, Prendez 2013;Bastián-Monarca et al. 2016) and in many countries, some The ability of neural networks in solving nonlinear and complex problems has been proven and has made it a suitable substitute for linear regression analysis for traffic noise modeling in recent research (Cammarata et al. 1995;Parabat, Nagarnaik 2008;Genaro et al. 2009;Kumar et al. 2014).
Despite tremendous efforts by numerous experts worldwide to develop various prediction models, these models are not reliable for Iran with different traffic characteristics and contribution of older and noisier vehicles. In many areas of Tehran, the capital city of Iran, highways are passing through the residential regions adjacent to the buildings which are considered as a health threat for residents. Also, the proximity of buildings to the highways causes traffic noise to be reflected by the buildings' facades and, as a consequence, noise levels increase. This key point should be considered in developing noise prediction models for this city.
Developing road traffic noise prediction models has attracted several investigators in Iran as well. In a study conducted by Givargis and Karimi (2010), application of neural networks for prediction of traffic noise led to satisfactory results for the city of Tehran. A preliminary neural network using the parameters of UK Calculation of Road Traffic Noise (CoRTN) was utilized in their model without considering the reflection effect of buildings adjacent to the roads. Ignoring the reflection effect of facades on the noise levels in the previously proposed models for Iran was the justification of the present study to develop a more comprehensive model which takes into account this phenomenon.
In this paper, an artificial neural network consisting of 9 input variables, including total traffic volume per hour, the average speed of vehicles, the percentage of each category of vehicles, road gradient, the density of buildings adjacent to the roads and the Building Reflection Factor, is presented. The learning process of the network is based on the random division of gathered data for training, validation and testing. At last, the results of the proposed model were compared with those of a regression model and some well-known classical models. It was found that the results of the ANN model were satisfactory.

Methodology
Tehran, having the largest number of streets and highways and the heaviest traffic in Iran, is one of the most appropriate places for collecting data associated with traffic noise pollution in the country. In this study, after assessing several sites in the city regarding continuous traffic, the existence of buildings adjacent to the roads and absence of disturbing factors such as intersections and traffic lights, 51 samples from 34 points were obtained ( Figure 1).
The data were collected from 7 a.m. till 8 p.m. for a one-month period in early summer. The instrument used in this study was (Lutron SL-4023SD) capable of recording the noise level in one-second intervals located at the height of 1.2 meters above the road surface (According to the ISO 362:1998) (Figures 2, 3). The noise measurements were conducted in dB(A) for 15 minutes in the pilot stage and, by observing a very slight difference between the results of 15 and 5 minutes in the first samples, the measurement duration of 5 minutes was chosen for the remaining points. Results of Pearson correlation between L Aeq in 15 and 5-minute intervals are shown in Table 1 which indicates a high correlation between them (P = 0.98). All the experimental data have been collected in absence of rain, with a wind speed below 5 m/s and relative humidity below 80%. Also, in all measurement sites, the ground type was hard and the sight angle was between 150-180 degrees. Simultaneously, noise recording was accompanied by video recording of traffic flow for 5 minutes using a camera placed on a nearby pedestrian bridge at each point ( Figure 4).

Equivalent continuous (A-weighted) sound level, L Aeq
Equivalent continuous (A-weighted) sound level is defined as the steady level of sound which, in a specific period of time contains the same acoustic energy as the actual timevarying sound level. The equivalent continuous sound level (L Aeq ) in the time period t 1 to t 2 is expressed as Eq. (1): where p(t) is the A-weighted instantaneous acoustic pressure and p 0 is the reference acoustic pressure equal to

Traffic volume per hour, Q
Traffic volume is defined as the total number of passing vehicles through a section. The number of each category of vehicles passing through a defined section was counted for one hour at each station.

Average speed, V
Measuring the speed of vehicles in both directions was done using video analysis by considering a specific distance on the videos and dividing the travel distance by the travel time.

Motorcycles
All powered two-wheelers PM Afterward, the average speed of vehicles was determined for each point and employed in the model ( Figure 5).

Vehicle classification
Each type of vehicle, based on its weight and emitted noise, contributes to the increase of the traffic noise level. Therefore, in this research, the vehicles are divided into four categories which consist of cars, vans and pickups, heavy vehicles and motorcycles. The percentage of each category in the total volume is calculated as well. Categories of vehicles and their descriptions are presented in Table 2. Types of heavy vehicles involved in the model2.

Gradient, G
By using an automatic level (NIVO NAK2), the road gradient in each point was measured. The procedure for measuring the road gradient and the corresponding formula is depicted in Figure 6 and Eq. (2) respectively.
where the values for the parameters a, b and L are obtained as demonstrated in the Figure 6.

Density of buildings facing the observer (D) and Building Reflection Factor (BRF)
The density of buildings (D) at reception point was calculated using Eq. (3): where i θ are the angles subtended by each facade on the opposite side of the road and t θ is the total sight angle. These parameters are shown graphically in Figure 7 and the required data were obtained from Google satellite images. In this study, the level of contribution of buildings in reflecting traffic noise was calculated by means of a novel method named Building Reflection Factor (BRF). For this purpose, to measure the height of the buildings in specified points, panoramic photography at each station was performed ( Figure 8) and the height of all buildings in front of the sound level meter and limited to the angle of view were obtained. Furthermore, distance from each building to the receiver was measured using Google satellite images and finally, the building reflection factor was calculated using Eq. (4).
where R i are the distances from each façade on the opposite side of the road to the reception point as depicted in Figure 7. i L and i H are the roadside width and height of those facades, respectively. i θ and t θ are the same as Eq. (3).
Finally, the collected data were imported into the ANN code for training and testing the network. Statistical descriptions of the data are given in Table 3.

Evaluation of noise pollution in the study area
Evaluation of noise levels at the measurement points indicated the violation of the maximum permissible noise level for commercial-residential areas legislated by the Department of Environment of Iran (60 dBA) in all 51 samples and decibel levels exceeding 75 dB(A) in 14 samples as presented in Figure 9 which could be harmful to human health. Therefore, Tehran's Municipality should consider the noise abatement programs seriously to mitigate the negative impacts of traffic noise pollution in the city. Noise mitigation measures such as the implementation of noise barriers and the insulation of buildings against noise should be considered as well as the scientific arrangement of roads and traffic flow (İlgürel et al. 2016).
Fortunately, the Municipality of Tehran has begun to install noise barriers in these areas in order to reduce the harmful effect of noise pollution on the public health of citizens. In some points which were measured in our study, such barriers were installed after a few months ( Figure 10).

Developing an artificial neural network with collected data
An artificial neural network is a machine learning method inspired by the biological neural networks. It consists of interconnected neurons. The numeric weight corresponding to each connection can be tuned by information in data which makes the network adaptive to inputs and capable of learning. This network is comprised of three layers of neurons; input layer, hidden layer and output layer, all of them having interactions with each other. Data  (5) and (6): where x i are the inputs, i w are weights, k b is the bias, ϕ is the activation function and k y is the output of the network. Selecting the type of activator function depends on the application of the network. In this study, the sigmoid function was utilized, which is defined as Eq. (7) (Haykin 1999;Demuth, Beale 1998 In this research, a multilayer feed forward neural network (Fausett 1994) was developed using MATLAB (R2014b). The dataset was split into 3 subsets of 80%, 10% and 10% for training, validation and testing, respectively. To train the network, the Levenberg-Marquardt optimization technique was used. This technique is a combination of the Gradient Descent Method (GDM) and Gauss-Newton's Method (GNM) with a blending factor, which makes the convergence of weights to the optimal values faster and is defined by the Eq. (8): ( ) where 1 p W + is the weight in the ( 1) p + th iteration, p W the weight in the pth iteration, H is the Hessian matrix, τ is a blending factor, diag[H] is the diagonal of the Hessian matrix and ∇ E is the gradient of error (Levenberg 1944;Marquardt 1963).  S To develop a model with minimum error, six different scenarios were defined. In each scenario, a number of parameters were included. Prediction accuracy in a neural network relies on its architecture, which consists of the number of hidden layers and the number of neurons in each layer.
In order to find the optimal number of neurons, the network is trained for each scenario with a different number of neurons (from the number of input parameters in that scenario to 25). To achieve the best architecture for the neural network, out of 100 iterations of the training process for each number of neurons, the best performance -based on the least Mean Square Error (MSE) and the best correlation coefficient -is selected and compared with the results of different number of neurons (the procedure is shown in Table 4. Comparison of different ANN architecture results in 100 iterations for (LAeq) in the 4th scenario4 for the 4 th scenario). The Mean Square Error (MSE) is calculated using Eq. (9): where N is the number of samples and ei is the difference between predicted and measured values for each sample.
Comparing the results of different scenarios in Table 4 indicates that among all investigated neural networks, the 4 th scenario yielded the highest correlation coefficient with measured data and the 6 th offered the least average of MSE in 100 iterations. Regarding the number of inputs in these two scenarios, the 4 th scenario was selected due to a lower number of input parameters which needs less data collection. As shown in Table 5, incorporation of the BRF parameter in the model lowered the average of MSE and increased the correlation coefficient to the measured values in comparison with the scenarios not containing this parameter.
Therefore, the optimal neural network structure is 6-10-1 and its characteristics are presented in Table 6. Optimal Architecture of Neural Network6 and Figure 12. As depicted in Figure 12, parameters which are involved in the model are traffic volume, average speed, heavy vehicles gradient, building density and building reflection factor.

Regression model
After developing the neural network, a multiple linear regression analysis was carried out to predict L Aeq using the same parameters. A summary of the regression model properties is given in Table 7. Summary of regression model properties a7.
Eq. (10) resulted from the regression analysis: Comparing the results of the regression model and the measurement data showed a prediction error between -4.63 to +3.61 dB(A) (Figure 13).

Neural network model results
The proposed model in the 4 th scenario resulted in a correlation coefficient of R = 0.9914 as shown in Figure 14. The prediction error for L Aeq using the ANN model in comparison with field measurement data was between -1.41 to 1.34 dB(A) (Figure 15).

Goodness of fit
In order to evaluate the performance of the developed model, a statistical paired t-test was applied at 5% significance level and 51 degrees of freedom. If the value of the tstatistic for output data is smaller than the critical t value, then, by accepting the null hypothesis (H0), it can be concluded that the averages of measured and predicted values do not differ significantly (Montgomery, Runger 2004).
The results of the regression and neural network models were compared with field measurements, shown in Table 8. Statistical paired t-test results for neural network and regression models8. The t-value for the neural network model was -0.130 which is much less than the critical t-value ±2.009 indicating a proper fit of predicted results to the measured values.
where Q is the traffic volume per hour, V is the average speed of traffic, P is the percentage of heavy vehicles, G is the gradient of road, Q L is the number of light vehicles per hour, Q P is the number of heavy vehicles per hour and d is the distance from observation point to the center of the traffic lane. The L m. E(25) is the average sound level at a distance of 25 meters from the center of the road lane. The results which consisted of standard deviation, correlation coefficient, MSE and R-Squared are summarized in Table  9. Comparison of proposed neural network with other well-known models 9. Better prediction of the neural network model is concluded based on lowest MSE (0.23463) and highest coefficient of determination (R 2 = 0.983). The comparison of these models to the measurement data is also shown in Figure 16. The better performance of the neural network is due to its greater capability in estimating non-linear relationships between the sound level and the factors affecting it.

Conclusions
The noise pollution produced by road vehicles is really a matter of huge concern in big cities, including Tehran. By selecting 51 stations for noise measurement in different areas of the city, it was shown that the noise levels were higher than the Iran environmental noise guidelines for residential-commercial areas and therefore, special attention from the municipality is required for mitigation or abatement of noise pollution in the city. As an intelligent noise prediction model, our proposed model can serve to assess the impact of the government's noise mitigation strategies or development plans before the implementation stage, such as examination of the environmental impact of highway design alternatives or the prediction of future noise levels.
By considering traffic parameters such as hourly traffic volume, average speed, the percentage of each category of vehicles and environmental factors including gradient, building density and Building Reflection Factor (BRF), six scenarios with different architectures of the multilayer neural network were investigated to estimate the equivalent continuous (A-weighted) sound level (L Aeq ). Among them, a multilayer neural network with a 6-10-1 structure with six input parameters including the BRF novel parameter was selected as the best model. It's high coefficient of determination (R 2 = 0.983) and low amount of prediction error in comparison with regression analysis and other classical models are in favor of the superiority of this model which was confirmed by a statistical paired t-test at 5% significance level.
Since the neural networks are capable of resolving complex problems with a great number of variables, researchers have the opportunity to include more related parameters in the process of noise prediction modeling compared to conventional models. Therefore, developing more precise and comprehensive models by incorporation of more valid and operational variables such as road surface, building facade material, the effect of green areas, etc. would be attainable. It is noteworthy to mention that the different characteristics of vehicles in terms of the modernity and level of noise production makes the results of this study more applicable in Asia region.

Disclosure Statement
There are not any competing financial, professional or personal interests from other parties.