|M.Sc Student||Barsky Yana|
|Subject||Feature Engineering Based Methodology for Congestion|
|Department||Department of Civil and Environmental Engineering||Supervisors||Dr. Ayelet Galtzur|
|Professor Shlomo Bekhor|
|Full Thesis text - in Hebrew|
Traffic Management (TM) aims to maximize the use of existing infrastructure, improve service level and ensure the safety of all road users. Congestion on the approaches to signalized intersections is a common phenomenon for which TM actions are intended to provide a fast and efficient solution.
Early detection of congestion formation on approaches to signalized intersections enables the activation of an alternative Traffic Signal Plan, increasing the capacity of the specific approaches by prolonging its green duration at the expense of other approaches. Hence, a high degree of prediction accuracy is a necessary condition for the effective implementation of such measure.
Urban signalized intersections typically exhibit rapid-intensive fluctuations of traffic flow in short periods of time, which poses a challenge of reliable traffic state prediction. The most suitable approach for dealing with that challenge is the use of data driven algorithms, offering a self-learning pattern recognition techniques, in which a prediction model is created by training the algorithm with historical traffic data measured from traffic sensors.
A crucial factor affecting the prediction quality of a model is the way that historical data is used as input to the algorithm in the training phase. Despite the recognition of the importance of the process of building input variables incorporating domain expert knowledge, a process called Feature Engineering, there is no structured methodology for implementing this process for short-term traffic forecasting.
This paper presents a Feature Engineering based methodology for real-time congestion forecasting on approaches to urban signalized intersections. The proposed methodology is an iterative process aiming to find at each iteration a set of features, based on incorporation of data mining techniques and traffic knowledge in building and selecting the most promising features, that result in an improved congestion forecasting accuracy and higher stability of prediction model, i.e. robustness to changes in dataset. For the purpose of this paper, four types of Performance Indicators (PI's) were developed to assess the quality of prediction at each iteration.
The developed methodology was tested on signalized sub-network, comprised of the congested approach on a major arterial in Tel Aviv-Yafo and five surrounding road sections having a potential impact on the quality of the prediction in the examined approach. Historic traffic data include volume and occupancy values for each loop detector from the road sections within the sub-network.
Examining the PI's over three iterations of a developed methodology results in reduction of input variables from 66 to 5, along with significant (statistically tested) improvement in all PI values, yielding a higher classification accuracy and stability of prediction model.
The results of the test case indicate that the application of the feature engineering is expected to improve the quality of the congestion prediction model. The method has a great potential to be extended for other sub-networks in the urban space. Although the test case was demonstrated on the basis of data based on loop detectors, the method is expected to be applicable for additional data types, e.g. traffic data obtained from Bluetooth sensors.