Public transport origin-destination matrices: pattern recognition and short-term prediction

Sammanfattning: Origin-Destination (OD) matrices are an essential tool in transport planning and management to model user travel patterns. An OD matrix is a picture of the public transport passengers demand in a specific temporal window The use of the metropolitan transportation system as an alternative to private cars enables a decrease of CO2 emissions, air pollution and traffic noise. Public opinion and governments are giving more and more attention to these environmental issues. The overtaking of the public transports on private means of transportation is one of our best weapons to fight global warming. Efficiency is particularly important when we analyse the passenger behaviour in the transport networks of large cities. There we deal with large and growing populations, and it is required a methodical response to the transport system critical events,like congestion of the network during the peak hours. Recently, modern technologies and data driven initiatives enabled large-scale data collection of travel patterns. In particular, smart card validation data can now be stored and used for evaluation and planning purposes. From these data, anonymous spatio-temporal travel patterns can be revealed and studied, and OD matrices with various aggregation levels can be easily generated. With this work, we gather different techniques used in the practice, and propose a methodological framework for the study of large-scale OD matrices. We focus our attention on the use of dimensionality reduction techniques (Principal Component Analysis (PCA), Singular Value Decomposition (SVD)), day clustering algorithms (K-Means, Affinity Propagation), and short-term flow prediction models (Vector Autoregression (VAR), Autoregressive Integrated Moving Average (ARIMA)) on the public transport usage data. Our case study considers the OD matrices for the whole metro and commuter train network in the region of Stockholm, Sweden. With this thesis, we want to reveal and discuss the potential of day clustering methods to detect trends of the passenger flows, and their combination with dimensionality reduction techniques to perform short-term prediction of the flows.