We perform an analysis of public transport data from The Hague, the Netherlands, combined from three sources: static network information, automatic vehicles location and automated fare collection data. We highlight the effect of bunching swings, and show that this phenomenon can be extracted using unsupervised machine learning techniques, namely clustering. We also show the correlation between bunching rate and passenger load, and bunching probability patterns for working days and weekends. We present the approach for extracting isolated bunching swings formations (BSF) and show different cases of BSFs, some of which can persist for a considerable time. We applied our approach to the tram line 1 of The Hague, and computed and presented four different patterns of BSFs, which we name “high passenger load”, “whole route”, “evening, end of route”, “long duration”. We analyse each bunching swings formation type in detail.

@en

Hierarchy is regarded as a natural phenomenon of public transport networks (PTN). The imbalanced distribution of passenger flow result in a hierarchical structure of PTN and it is also related to the development of technology and the introduction of new modes. However, there is still a lack of knowledge on how to identify the hierarchical structure of the multi-layer PTN. This study proposes a three-step passenger transfer flow based methodology for separating and ranking the PTN: (1) using passenger journey data to derive transfer flow matrix; (2) applying network representation with Louvain method of community detection to separate the PTN layers; (3) performing ranking method, separating inner-transfer and inter-transfer flow. To demonstrate our method, we use one-month smart card data of The Hague, the Netherlands provided by the PTN operator HTM. Our results show that our method is able to, regardless of the geographic location and the mode of transportation, identify the hierarchy of PTN based on the passenger transfer flow pattern. Temporal attributes are also discussed to illustrate how hierarchy is time-dependent, e.g. with respect to the day of the week and the time of the day. Our method supports public transport (PT) operators during design and optimization of PTN and in determining which sets of higher-level service to prioritize during different time periods.@en

Public transport (PT) plays an increasingly important role in solving mobility challenges, especially in densely populated metropolitan areas. Further improving PT systems requires more advanced planning and operations. Fortunately, the considerable amount of data that have become increasingly available for PT systems offer an opportunity to address this challenge. However, how these data can be effectively used to achieve this goal still remains as an unresolved question in the scientific literature. More research is therefore needed to bridge this gap in order to advance PT systems for addressing mobility challenges. To this end, this dissertation is focused on developing methods and models for translating high-volume data from various sources into novel knowledge and insights that can be used to improve PT planning and operations. This dissertation first examines how to obtain onboard occupancy of PT vehicles by integrating all the three different data sources mentioned above. Second, this dissertation deals with the issue of high-dimensionality in large-scale passenger flows. Third, we propose a k-means-based method to cluster PT stops for constructing zone-to-zone OD matrices. Fourth, this dissertation presents a new method for analyzing the accessibility of PT service networks based on a novel network science approach. Last, we investigate whether passenger flow distribution can be estimated solely based on network properties in PT systems.@en

Network science offers powerful concepts and methods for studying complex systems, such as public transport networks. However, many existing studies on complex network analysis of public transport networks were primarily motivated to test network science concepts using real-life networks. Consequently, important properties related to public transport services in these studies have been overlooked. This has led to claims made by transport researchers that these works do not necessarily improve the understanding of system properties and performance. To bridge this gap, this study integrates network science and public transport accessibility analysis for comparative assessment across multiple public transport networks. The primary contribution of this study pertains to developing a method based on network science for computing public transport accessibility measured as the average travel impedance. The travel impedance metric is defined based on the generalized travel cost between stop pairs, containing initial and transfer waiting times, in-vehicle travel times and time-equivalent transfer penalty costs. To perform efficient computation, we propose a new type of weighted graph representation of public transport networks, which explicitly incorporates travel costs that are derived from general transit feed specification (GTFS) data. Besides the methodological contribution, the secondary one of this study consists in performing a comparative assessment of worldwide tram networks' accessibility. The analysis shows insights into how different travel components (e.g., in-vehicle travel times and waiting and transfer times) specifically account for the variance in accessibility across different networks. Additional references for improving the designing and planning of public transport networks can be obtained from such comparative assessments.

@en

We present a pioneering investigation into the relation between passenger flow distribution and network properties in public transport systems. The methodology is designed in a reverse engineering fashion by utilizing passively measured passenger flow dynamics over the entire network. We quantify the properties of public transport networks using a range of centrality indicators in the topological representations of public transport networks with both infrastructure and service layers considered. All the employed indicators, which originate from complex network science, are interpreted in the context of public transport systems. Regression models are further developed to capture the correlative relation between passenger flow distribution and several centrality indicators that are selected based on the correlation analysis. The primary finding from the case study on the tram networks of The Hague and Amsterdam is that the selected network properties can indeed be used to approximate passenger flow distribution in public transport systems to a reasonable extent. Notwithstanding, no causality is implied, as the correlation may also reflect how well the supply allocation caters for the underlying demand distribution. The significance and relevance of this study stems from two aspects: (1) the unraveled relation provides a parsimonious alternative to existing passenger assignment models that require many assumptions on the basis of limited data; (2) the resulting model offers efficient quick-scan decision support capabilities that can help transport planners in tactical planning decisions.@en

Minimizing passenger transfer times through public transport (PT) transfer synchronization is important during tactical planning and real-time control. However, there are computational challenges for solving this Timetable Synchronization Problem (TSP) for large, real-world urban PT networks. Hence, in this study we propose a data-driven, passenger-oriented methodology as a preparatory selection stage to reduce problem dimensionality by (1) determining the significant transfer hubs in the network, and (2) identifying subsets of lines within these hubs that need to be prioritized for transfer synchronization. In the first phase of our methodology we determine the spatial boundaries of transfer locations, using a clustering technique based on the passenger transfer flow matrix inferred from smartcard data. After that, a subset of hubs to be prioritized for synchronization is selected. In the second phase, we characterize the transfer patterns within the hubs based on a topological representation. Based on these topological graphs, the line bundles that need to be prioritized within the hubs are further identified using a modularity-based community detection technique. We apply our methodology to a real-world case study, i.e. the PT network of The Hague, the Netherlands. For this case study, our approach allows for prioritizing 70% of all transfers within identified transfer locations while only requiring 0.9% of these transfer locations, thus reducing the complexity of solving the TSP substantially at a relatively low cost. Our method supports public transport operators during timetable design and real-time control in determining where and which lines to prioritize when devising measures for improving transfer experience and synchronization.@en

Obtaining load profiles of transit vehicles has remained as a difficult task for transit operators because of technical and financial constraints. Although a significant advance in transit demand and supply data collection has been achieved over the past decade, information related to load profiles at the vehicular level is either impossible or very difficult to retrieve from them. It is not even uncommon to see that these data are underutilized by transit operators owing to considerable deficiencies and shortcomings in the data themselves, and/or the processing algorithms needed to process them. This study is therefore dedicated to addressing this challenge that has largely been overlooked by both researchers and practitioners. First, the issues which hinder the construction of load profiles based on three prevailing transit data sources are identified, including automatic fare collection (AFC), automatic vehicle location (AVL), and general transit feed specification (GTFS) data. Second, a methodology is developed for sequentially addressing all the issues and generating desirable vehicle load profiles. The methodology consists of four steps, including (1) data pre-processing, (2) matching trips in GTFS and AVL, (3) matching passenger rides to vehicle trajectories, and (4) improving vehicle trajectories. The resulting spatiotemporal load profiles of transit vehicles enable detailed investigation into vehicle movements and demand patterns over time and space, including service utilization and the propagation of delays and crowding. Data collected from the urban transit network in The Hague, The Netherlands are used to demonstrate the proposed methodology. The visualization of spatiotemporal load profiles through space-time seat occupancy graphs provides operators with a compact and powerful reference for the improvement of their services.

@en

In this extended abstract, we show the supervised learning approach to predicting passenger load of trams, based on historical passenger load patterns. We look at two different cases: predicting long-term passenger load of any given day and time, and predicting short-term passenger load at a particular public transport vehicle.@en

It has been rarely investigated in the field of public transport whether passenger ow can be explained by network topology. Based on the rich data sets from The Hague, The Netherlands, we conduct this study try- ing to shed light upon this question. The relation between passenger ow and topological measures in different public transport network representations are investigated in detail. Our preliminary results show promising evidence that the passenger ow is indeed correlated to topological measures in the case study network. Several linear regression models are also constructed to quan- tify the explanatory power of these topological measures.@en

For large urban networks and hubs, optimizing transfer synchronization becomes computationally challenging. The objective of this paper is therefore to develop a generic, data-driven methodology to determine the key line/direction-combinations to synchronize based on passenger flows. We developed an approach to detect communities of directional lines based on passenger transfer flows, by calculating modularity using a C-space inspired network representation. Our results show intuitive clusters to prioritize for synchronization on a network level for tactical planning, and on the hub level for real-time coordination.@en

We perform analysis of public transport data from March 2015 from The Hague, the Netherlands, combined from three sources: static network information, automatic vehicles location (AVL) and automated fare collection (AFC) data. We highlight the effect of bunching swings, and show that this phenomenon can be extracted using unsupervised machine learning techniques, namely clustering. We show different cases of bunching swings, some of which can persist for a considerable time. We also show the correlation of bunching rate with passenger load, and bunching probability patterns for working days and weekends. We show, how formations of bunching swings can be extracted, and clustered into four different types, which we name "high passenger load", "whole route", "evening late route", "long duration". We analyse each bunching swings formation type in detail.@en

Transit networks are complex systems in which the passenger flow dynamics are difficult to capture and understand. While there is a growing ability to monitor and record travelers' behavior in the past decade, knowledge on network-wide passenger flows, which are essentially high-dimensional multivariate data, is still limited. This paper describes how Principal Component Analysis (PCA) can be leveraged to develop insight into such multivariate time series transformed from raw individual tapping records of smart card data. With a one-month data set of the Shenzhen metro system used in this study, it is shown that a great amount of variance contained in the original data can be effectively retained in lower-dimensional sub-spaces composed of top few Principal Components (PCs). Features of such low dimensionality, PCs and temporal stability of the flow structure are further examined in detail. The results and analysis provided in this paper make a contribution to the understanding of transit flow dynamics and can benefit multiple important applications for transit systems, such as passenger flow modeling and short-term prediction.

@en

So-called tap-in–tap-off smart card data have become increasingly available and popular as a result of the deployment of automatic fare collection systems on transit systems in many cities and areas worldwide. An opportunity to obtain much more accurate transit demand data than before has thus been opened to both researchers and practitioners. However, given that travelers in some cases can choose different origin and destination stations, as well as different transit lines, depending on their personal acceptable walking distances, being able to aggregate the demand of spatially close stations becomes essential when transit demand matrices are constructed. With the aim of investigating such problems using data-driven approaches, this paper proposes a k-means-based station aggregation method that can quantitatively determine the partitioning by considering both flow and spatial distance information. The method was applied to a case study of Haaglanden, Netherlands, with a specified objective of maximizing the ratio of average intra-cluster flow to average inter-cluster flow while maintaining the spatial compactness of all clusters. With a range of clustering of different K performed first using the distance feature, a distance-based metric and a flow-based metric were then computed and ultimately combined to determine the optimal number of clusters. The transit demand matrices constructed by implementing this method lay a foundation for further studies on short-term transit demand prediction and demand assignment. @en