Big data in railway operations: using artificial neural networks to predict train delay propagation
暂无分享,去创建一个
Advances in big data, data collection and machine learning have made
it possible to apply new machine learning concepts to a wide array of
problems. The aim of this thesis is to explore the possibility of predicting
secondary delays in a railway network using a recurrent neural network.
This can eventually be used to perform risk analysis and alternative
evaluation combined with stochastic delay modelling. Empirical data from
Irish Rail is used to test this method and verify the results.
First a RailML data model is constructed containing infrastructure,
timetable and rolling stock information based on multiple data sources
from Irish Rail. Significant features are identified and extracted from
this data model for around 60.000 different delay combinations. Then a
sequential approach is used incorporating a recurrent neural network to
predict the total knock-on delay. This sequential approach allows for input
data in variable lengths, avoiding information loss due to generalization of
features. The model is trained with mini-batch gradient descent using the
RMSprop algorithm on a large portion of the 60.000 training examples, and
validated using the remainder of the example delay combinations.
A coefficient of determination of R2 = 0; 7029 is achieved, which is
comparable to similar machine learning methods presented in literature.
The resulting accuracy is in the same order of magnitude as similar research
using support vector machines. While results are less accurate then
the results that can be achieved with micro-simulation tools, a series of
improvements of the proposed method are presented which might be able
to elevate the results of this method to a higher level.