Data-Driven Research on Traffic in Transportation Networks: Analysis, Modelling, and Forecasting

Abstract

With the emergence of the massive amount of data collection technologies, data-driven computational approaches pose research opportunities for traffic analytics of intelligent transportation systems in the era of big data, which has provided significant resources and strengthened the understanding of the models of analysis, interpretation, optimization, decision-making, and forecasting. The fine traffic analysis and its approaches require at least tailoring the previous approaches to make them suitable for specific problems, and new methods are sometimes required to be developed to address the specific problems. In this thesis, we develop tailored advanced solutions to three emerging traffic analysis problems, which include traffic dynamic evolution analysis and bottleneck identification, modelling of traffic flow and its influencing factors, and traffic flow forecasting. To the best of our knowledge, the methods are more suitable for utilization in the context of traffic analysis problems in this thesis.

Connectivity is a critical performance assessment of transportation network by measuring how well different locations connect to one another via traffic flow, thus it is a good starting point for analyzing traffic by combining transportation network structure and traffic flows. Despite the rich literature on the connectivity analysis of transportation network, very little attention has been paid to passengers' heterogeneous cognition toward congestion and connectivity incorporating subjective judgment. The first part of the thesis develops a data-driven framework based on percolation theory to reveal the dynamic evolution of metro network connectivity and identify recurrent bottlenecks involving individual cognition. The concept of individual tolerance index of congestion and a measure named network friendliness are proposed. By comparing individual tolerance index and friendliness of metro network, metro network connectivity with regard to different passengers can be depicted quantitatively. The evolution of network connectivity can be monitored both as individual tolerance changes and as time goes on. We also demonstrate how global transit breaks down when identified bottlenecks are congested from the perspective of passengers’ cognition. The proposed method is validated using a real-world case of Shenzhen Metro in China. Results show that the proposed method is effective in capturing the dynamic evolution of Shenzhen metro network connectivity and enable effective identification of transit bottlenecks. The network connectivity and friendliness are found to be significantly increased through a small improvement of the bottlenecks pinpointed.

Understanding the traffic and transportation network connectivity dynamics, we may try to make a prediction and forecasting of future traffic, which is helpful to both the long-term and short-term transportation planning. From a macro perspective of traffic prediction, ridership as the station-level traffic of transit systems, its modelling and prediction play a critical role in transit transportation planning. The second part of the thesis focuses on identifying influencing factors on metro ridership through ridership modelling and prediction. In the first subpart, we identified the influencing factors on Taipei metro ridership for different time resolutions (day of the week, week of the month, and month of the year) through modelling with generalized estimating equations (GEE). Different from previous works, this study looks at longitudinal station-level metro ridership at varied time resolutions. The longitudinal ridership data of Taipei Metro and its potential influencing factors data in the urban environment in the year 2015 are used to validate the effectiveness of our proposed method. The results demonstrate that the proposed method performs well in estimating longitudinal metro ridership. In the second subpart, we propose an Adapted Geographically Weighted LASSO (Ada-GWL) framework for modelling metro ridership, which involves regression-coefficient shrinkage and local model selection. It takes metro network layout into account and adopts network-based distance metric instead of Euclidean-based distance metric, making it so-called adapted to the context of metro networks. The real-world case of Shenzhen Metro is used to elaborate our proposed model. The results show that the proposed Ada-GWL model performs the best compared with the global model (Ordinary Least Square (OLS), Geographically Weighted Regression (GWR), GWR calibrated with network-based distance metric and Geographically Weighted LASSO (GWL) in terms of estimation error and goodness-of-fit. Through understanding the variation of each coefficient across space (elasticities) and variables selection of each station, it provides more realistic conclusions based on local analysis.

From a micro perspective of traffic prediction, short-term traffic forecasting is a central problem for transportation authorities and operators. It is the critical step to determine optimal operational control strategies that respond to passengers’ needs. Since traffic flows can be featured by spatiotemporal dynamics, it is expected that various spatiotemporal forecasting problems can significantly benefit from the newly developed methodologies, models, and algorithms to improve the prediction accuracy in transportation applications further. Therefore, the third part of the thesis mainly focuses on two problems related to short-term traffic forecasting: travel demands between Origin-Destination (OD) pairs forecasting, and nationwide passenger flow forecasting. Both two problems are very challenging since they are affected by complex factors. For the first problem, this thesis develops an innovative travel demand forecasting approach based on a deep learning model, spatiotemporal residual network (ST-ResNet), to forecast short-term OD matrix. The complex relevance among OD pairs, temporal dependencies, and external factors are considered simultaneously. The proposed method is applied to the short-term forecasting of metro OD matrix in Shenzhen, China. The experimental results show that our method can well capture the multiple complex dependencies and outperforms state-of-art baselines in terms of forecasting accuracy. For the second problem, this thesis proposes an approach by extending ST-ResNet to simultaneously forecast inflow and outflow of nationwide city-level passengers in irregular-shaped regions. Region segmentation and approximation strategies are adopted to transform the irregular regions to several regular grids, making it possible to use convolution operation to capture spatial dependencies among the irregular-shaped region. The proposed method is validated by the real-world inter-city travel counts data of China, and the experimental results show that the proposed method can well capture the spatiotemporal dependencies and rule out both the spatial and temporal autocorrelation in residuals. Through comparison, the proposed method outperforms state-of-art baselines in terms of forecasting accuracy.

In general, this thesis fills in the gap in the research of traffic analysis, modelling and forecasting by developing several novel data-driven approaches in various application fields such as metro network bottleneck identification, analysis of influencing factors on metro ridership, short-term forecasting of travel demand in a transit system, and nationwide city-level passenger flow forecasting. The analysis and implications in this thesis can help travelers to understand the traffic dynamics and provide suggestions for authority to construct a user-friendly transportation network in a highly efficient, low-cost way, so that can pave the way for new applications and a near real-time understanding of traffic patterns and dynamics of intelligent transportation systems, in particular in the future realization of the “smart city.”

Date of Award	10 Aug 2020
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Kwok Leung TSUI (Supervisor)