Spatio-temporal clustering for non-recurrent traffic congestion detection on urban road networks

Abstract

Non-Recurrent Congestion events (NRCs) frustrate commuters, companies and traffic operators because they cause unexpected delays. Most existing studies consider NRCs to be an outcome of incidents on motorways. The differences between motorways and urban road networks, and the fact that incidents are not the only cause of NRCs, limit the usefulness of existing automatic incident detection methods for identifying NRCs on urban road networks. In this paper we propose an NRC detection methodology to support the accurate detection of NRCs on large urban road networks. To achieve this, substantially high Link Journey Time estimates (LJTs) on adjacent links that occur at the same time are clustered. Substantially high LJTs are defined as those LJTs that are greater than a threshold. The threshold is calculated by multiplying the expected LJTs with a congestion factor. To evaluate the effectiveness of the proposed NRC detection method, we propose two novel criteria. The first criterion, high-confidence episodes, assesses to what extent substantially high LJTs that last for a minimum duration are detected. The second criterion, the Localisation Index, assesses to what extent detected NRCs could be associated with incidents. The proposed NRC detection methodology is tested for London’s urban road network. The optimum value of the congestion factor is determined by sensitivity analysis by using a Weighted Product Model (WPM). It is found out those LJTs that are at least 40% higher than their expected values should belong to an NRC; as such NRCs are found to maintain the best balance between the proposed evaluation criteria.