Multi-model bus arrival prediction with intelligent handling of uncertainties

Public transport is a widely adopted and well-known way of ensuring connected mobility
for the broad population in both metropolitan cities and urban areas. It makes good
sense even in a more rural setting to at least offer some strategic travel corridors that are
connected and supplemented with more demand-driven services.
Both politically and scientifically, getting more people to use public transport is considered
an important way to reduce the number of trips with passenger cars, and a way to free
space in cities under pressure, where congestion results in a huge amount of wasted time
on a daily basis. In addition, it is a key initiative to achieve the different climate goals
set by the EU and the UN, among others, and thus implement the green transition of the
transport sector. To support this transition, it is necessary that public transport becomes
more attractive to compete against the benefits offered by a passenger car. There are
many criticisms to address, such as travel time, availability, comfort, and reliable and
accurate traveler information. These must each and every one be developed and improved
continuously in order to become, and not least remain, a relevant, competitive, and green
alternative to passenger cars.
This PhD thesis deals in particular with the improvement of traveler information, more
specifically precise and robust bus arrival and departure time predictions. Since buses
traditionally share lanes with other modes of transport, they are naturally difficult to
predict. In addition, related topics are also addressed that can help to address several
of the points listed above, including better assurance of transfers between two public
transport services that can reduce travel time and analyzing travel behavior that can help
fit the public transport system to people needs in better ways. The thesis is divided into
three parts: i) The first part presents innovative methods and models for primarily shortterm
bus arrival and departure predictions, ii) the second part continues these findings by
investigating the advantages and disadvantages of combining several stand-alone models
in an overall forecasting system as well as the industrial application, while iii) the last part
deals with the use of location data and machine learning models to better understand
travel behavior.
More in detail, the first part deals with the development of several novel methods and
models, each of which contributes to improved accuracy for bus travel time and thus
more precise arrival and departure times for buses. All models are based on various innovative
machine learning techniques, and use large amounts of data to learn patterns and
contexts from previously observed bus location data. Finally, models are also developed
to handle the variability that naturally occurs in travel times, in order to better quantify
and constrain the uncertainty of bus arrival. This information is used concretely to show
how an expert system can improve transfers between buses by prioritizing holding time in
an intelligent way.
The second part expands this approach from a more applied and operational perspective
by presenting a concept for a multi-model prediction system where several independent
base models can collaborate and compete to achieve an overall more functional, robust
and accurate system. It is argued a system with these capabilities is more adoptable under
the real-world conditions seen in the industry. Significant design choices are presented
to implement such a system in a scalable, robust, and cost-effective way, and a fully
operational and open-source system is provided as part of the PhD project.
The third and final part deals with two studies, which relate to the previous parts by
combining location data and advanced machine learning techniques but focus on travel
behavior. The first study deals with estimating walking times when transferring between
bus and train. A model is used to separate passengers into two categories; those who make
a direct transfer, and thus accurately represent the actual travel time between bus and
train, and those passengers who make activities during a transfer, e.g. shopping, nearby
visits, etc. Finally, a method is presented to classify a sequence of positions collected from
smartphones in either a stop or movement. This is an important step in collecting automated
travel diaries from smartphones, which potentially constitute a large and insightful
source of travel behavior information, and can support policymakers and advisors in the
planning of public transport and the transport sector in general.
Overall, it is concluded that the advanced and novel methods and models developed for bus
arrival prediction can improve accuracy, but they are often also unrealistic to directly adopt
and put into production in the industry because a number of conditions are assumed that
are rarely true in real-world scenarios. The proposed multi-model approach solves some
of these challenges, however, by partially compromising on accuracy. Another important
conclusion is that modeling the uncertainty can in some cases be added without notable
consequence for the computational complexity, and can be used for more intelligently
handling of several scenarios in public transport, e.g. transfer synchronization, connection
assurance, and transfer walking time. Finally, it is further concluded that the combination
of location data and machine learning techniques can reveal detailed information on travel
behavior that is not considered realistic to obtain using manual alternatives. Thus, they
provide an automatic and cost-effective source for broad and detailed insight into travel
behavior.