Graph Neural Network for Traffic Demand Prediction
Publicerad
Författare
Typ
Examensarbete för masterexamen
Program
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
The urban transportation systems are under more pressure than ever before, with more people living
in urban areas, higher travel demand, congestion and increased emissions. It is an urgent and
indispensable trend to develop sustainable transportation to reduce traffic-related issues. One of the
emerging sustainable transportation systems is shared mobility such as ride hailing, shared e-scooter
and bike/e-bikes. Shared mobility enables the sharing of different types of vehicles and increase the
usage efficiency of vehicles. Meanwhile, active mobility sharing such as bike, e-bike and e-scooter
sharing are environmentally alternatives as compared to private cars and are very prevalent in current
European countries. One of the fundamental components of operating shared mobility is accurate
demand prediction in spatial and temporal dimensions, which are key inputs for vehicle dispatching
and rebalancing. Thanks to the digitalization of transportation systems, big data regarding the usage
demand in different periods and areas in a city are collected. Meanwhile, advancements in artificial
intelligence and machine learning have paved the way for new techniques to leverage the great power
of data-driven methods and big data for precise travel demand prediction of shared mobility system.
The aim of this study is to design a graph neural network algorithm to quantify short-term demand of
shared mobility systems. The algorithm specially takes both spatial- and temporal aspects into
consideration when predicting the demand in a data-driven way. We implement the proposed
algorithm for a case study in bike-sharing demand prediction in the city of Nanjing, China. The data in
the case study cover three months including October, November and December 2016 where each
month has more than 3 million transaction data. The given data contains among other things the ID
for each station, rental- and return time and longitudinal- and latitudinal coordinates. Before initiating
the coding process, a literature review was conducted, where artificial intelligence and most
importantly deep learning and neural networks are analyzed and explained in detail in order to lay the
theoretical foundation and select the best deep learning method for our research purpose. The
algorithm is structured by first defining the structures and layers, and then proceeding with building
the model and training- and testing the model. There are over 800 bike stations in the datasets.
However, not all stations are used for training and testing the graph neural network on account of
computation efficiency and time. The analysis uses 50-, 100-, and 200 stations separately on account
of the potential impacts of data amount on results. In order to measure the accuracy of the proposed
algorithm, two commonly used error metrics in deep learning will be employed, mean absolute error
(MAE) and root mean square error (RMSE). It should be noted that We utilize the available bike sharing
data for case study, but the algorithm is general and can also be used for other mobility systems such
as regular car traffic or train systems as well.
The results show that the MAE for the validation set and the testing set are around 6.2 and 6.27,
respectively, while the RMSE for the validation set and the testing set are approximately 8.6 and 9.1,
respectively. These numbers were for the base architectural layout of the graph neural network and
for the hourly demand. This study also conducts additional experiments for hyperparameter tuning
and also changing the block layout for the model to find the best model, for instance the most notable
improvements were for the 100-station analysis which showed approximately 12-19 % decrease in the
error metrics. Most notably, the lowest MAE and RMSE values are obtained for the 200-station
analysis. In additional to the two error metrics, there will be graphs added that highlights the demand
errors for every station analysis. This showed, from a monthly perspective, that the graph neural
network actually predicted the demand relatively close to the real demand.
The proposed graph neural network has the ability to predict short-term bike demand with relatively
high accuracy. However, there is much room for improvement where meteorological condition could
have also been implemented in the model, which would with a high probability increase the accuracy
of the predictions.
Beskrivning
Ämne/nyckelord
Artificial intelligence, deep learning, neural network, graph neural network