In recent days, management of road traffic and controlling the congestion has become major problems in Hyderabad city at any busy junction. Hence, short term traffic flow forecasting has gained greater importance in Intelligent Transport System (ITS). Exponential smoothing models have been profitably employed as a popular linear time series forecasting model. Besides, Artificial Neural Networks (ANNs) are being applied to capture the complex relationships with a various pattern as they assist as a flexible and potential computational tool. But, most of these analyses represented mixed results in terms of the efficacy of the ANNs model in comparison with the linear model. In this paper, we propose a Hybrid model, which is distinctive in integrating the advantages
In the direction of road traffic analysis, some studies have been taken place. Yan H et al,
There are many studies aimed at developing a best forecasting model for road traffic in the literature. But from all those studies it is clear that no model would be appropriate for all the regions and under all the circumstances. For every forecasting problem, conditions and characteristics have to be analyzed properly and completely to develop a model. Objective of this study is to develop a model that best forecasts the Hyderabad road traffic at the selected point.
Short term traffic flow forecasting involves predicting the traffic volume in the next time interval usually in the range of 5 minutes to 30 minutes. For this study we have considered 5days traffic data at 6 no. junction, Amberpet, Hyderabad, Telangana state, India. In any junction it is very important to forecast the short term traffic flow to design planning and operations of traffic signals and various traffic strategies. In this paper an attempt was made to develop a shortterm traffic flow forecasting model using Hybrid model  combination of Exponential smoothing model and Artificial Neural Network (ANN). Three indicators including the Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE) and the Mean Absolute Error (MAE) have been used to check the models’ performance.
The data was taken from the office of Commissionerate, traffic, Hyderabad. The data set used in this study was collected from 6 no. junction, Amberpet, Hyderabad, Telangana state, India. The data was collected for the peak hours in the morning. The data set contains records of number of vehicles passing through the junction from 8 AM to 12 Noon in the intervals of 5 minutes from 08022019 to 13022019. The data was given in the form of video captured by cc TV cameras fixed at the junction. Volume of vehicles was obtained by counting manually. Data extraction was done in the intervals of 5 minutes for 4 directions. 80% of the data has been used for fitting the models and the remaining 20% of the data has been used for testing the adequacy of the model.
Seasonal Exponential Smoothing or HoltWinters method or Method of Winters, is an advanced expansion of the exponential smoothing methodology. It generalizes the procedure to deal with level, trend and seasonality by considering three smoothing parameters α, β, γ and ‘m’ denotes the observations count in a cycle.
The HW method consists one forecast equation and three smoothing equations one for the level
Level
Trend
Seasonal
Forecast
where
Neural networks (NN) have broad applicability to real world. Since, neural networks are best for identifying patterns or trends in data, they are well suited for prediction or forecasting needs. Multilayer perceptron (MLP) is most widely used network structure of Artificial Neural Network (ANN)
MultiLayer Perceptron uses the “back propagation rule” which calculates an error function for each input and back propagates the error from one layer to the previous one. The weights for a particular node are adjusted indirect proportion to the error in the units to which it is connected. An activation function is applied to the weighted sum of the inputs of a neuron to produce the output. In this study we used sigmoid function as activation function. A Sigmoid function is defined as
The MLP learning algorithm using the back propagation rule includes initializing weights (to small random values) and transfer function and then adjust weights by starting from output layer and working backwards. The weight function considered in this paper is
where
For output layer units:
For hidden layer units:
A two step procedure is followed to determine a unit in the output layer
Step1: Using the formula
Step2: Calculate the activity
^{ }where
Linear and non linear models are combined to obtain Hybrid model. Forecasting accuracy can be improved by Hybrid models. Many works have taken place in the light of Hybrid models for solving forecasting problems. Zhang
where
1. Obtain the residuals by modeling the linear part
2. Employ a nonlinear model to these residuals for handling with the nonlinear part. This study used Holt Winters model as linear model.
Let
where
where
The basic Statistical characteristics of the data discussed in 2.1 are presented in


Minimum 
487 
Maximum 
930 
Mean 
637 
Standard deviation 
102.424 
Skewness 
0.4377 
Kurtosis 
3.77702 
To evaluate the performance of the proposed model, three models such as Holt Winters, Artificial Neural Networks and Hybrid HWANN models were considered for the data set discussed in 2.1.
The parameters of the model were estimated as in the




HOLTWINTERS’ Additive model 
Level Trend Season 
Alpha(α) Beta(β) Gamma(γ) 
0.9 0.1 0.1 
For this data, level parameter obtained as 0.9, Trend and Seasonal components parameters were estimated as 0.1.
In this study, MLP network has been used for the prediction of short term traffic flow. Different Artificial Neural Network (ANN) models have been developed on the data set. In the present study RMSE, MAE values were used to evaluate the performance of the model and predicted results. The specification of all the models has been presented in






M1 
1 
1 
0.47045 
34.6948 
16.32106 
M2 
1 
2 
0.46346 
34.4347 
15.84144 
M3 
1 
3 
0.46236 
34.3896 
15.77978 
M4 
1 
4 
0.46147 
34.3539 
15.72724 
Firstly, original data is modeled by HoltWinters model and obtained the residuals and then ANN is fitted to these residuals. Finally, the forecasted values of HWANN model can be obtained by summing the forecasted values of HoltWinters and ANN. The inputs for the ANN at this HWANN model for road traffic data are the errors from the HoltWinters model. By trying the number of neurons in hidden layer from 1 to 4, the results show that 4 neurons in hidden layer is the best model for forecasting the road traffic.
The three models performance was evaluated by the performance evaluation metrics RMSE and MAE. The obtained values of RMSE, MAPE and MAE of the three models are presented in




Holtwinters Additive 
33.5388 
1.953 
14.87386 
ANN 
34.3539 
2.4834 
15.72724 
Hybrid HWANN 

1.9062 

From the view point of RMSE, MAPE and MAE the Hybrid HWANN method is superior for forecasting the Short term road traffic followed by the individual Holt Winters method and individual ANN model. In the literature there are studies related to developing linear and nonlinear forecasting models separately but our study shows that combined model performs better than that of linear and nonlinear models individually. The
As mentioned earlier 20% of the data has been used to check the adequacy of the best model. Using the best identified Hybrid model, values are predicted for the 20% of the data points for testing the model adequacy and plot of the actual and estimated values is presented below Plot of Actual and Predicted values of Testing data.
We proposed a Hybrid model comprising Holt Winters model and ANN model for forecasting of Short Term Road Traffic flow in a busy junction (6no. Amberpet) and explored the forecasting capability of the Hybrid model. Generally road traffic has mixed random characteristics which have to be thoroughly analyzed to obtain the valid results. From the fore studies we observed that no model would be ideal for all the forecasting needs. The urban cities like Hyderabad need continuous and perfect updates on traffic. From the analysis we conclude that the Hybrid model performs better than the individual methods, in terms of less error values which imply that Hybrid method can be used as a solution to forecast the Short Term Road Traffic flow at any traffic bottleneck point in Hyderabad.