Local traffic patterns extraction with network-wide consistency in large urban networks

Abstract

Abstract With the increasing availability of traffic data in large urban areas, there is an opportunity to infer more sophisticated traffic patterns and trends which till now has been difficult to obtain and understand. The inferred traffic patterns can then serve as input into short-term traffic prediction models or to predict traffic demand in the network for deployment of traffic management policies. However, the network-wide consistency, the interaction between variables and the difficulty during forecasting in choosing a pattern at early AM peak hour require robust, but flexible enough pattern extraction method. A wide range of researchers tackle some of these points, but it is difficult to find the one that fulfills all of them. Thus, this paper presents traffic flow pattern extraction methodology to identify daily traffic flow profiles consistent among all the detectors as a patterns based on two-phase methodology consisting of decision tree algorithm, Pathmox, and 2-step iterative clustering. The first phase uses qualitative variables (i.e., holidays, large events, weather, day-of-week) to capture meaningful and robust patterns in a tree-based configuration. The second phase consists of two steps that aim to reduce still high variability in some patterns non-attributable to the qualitative factors already exploited in the first phase. The advantages of this approach are (1) its capacity to break down variability in patterns due to both known and unknown factors; (2) it does not rely on specific network settings and (3) network-wide scale consistent patterns are identified. The traffic pattern extraction method is evaluated with real traffic flow data collected in period of one year on a motorway network M4 and M7 in Sydney, Australia. Results show that the proposed method extracts more identifiable patterns and more efficiently captures trends in the data with almost non-overlapping conditional variability bands compared to spatial-clustering approach based on k-means.