Abstract
The availability of large tranches of data and its influence on traffic flow, make the problem of short-term traffic speed prediction very complex in nature. For more than 40 years, various statistical time series forecasting methods have been applied for traffic speed prediction, and in the last 20 years, machine learning-based methods have gained prevalence. However, more recently, recurrent neural network (RNN) based methods have emerged to show better results for traffic speed prediction\cite{tian2015pred, zhao2017lstm,fu2017usin,chen2016long,dai2017deep,dai2019deep, kanestrom2017traf,shao2016traf,jia2017traf}. As the interest in applying RNN models to the traffic speed predictions started to grow, we found some critical and important unanswered questions with respect to the application of such methods. From these open questions, as part of this research study, we studied the following three questions for multi-step-ahead short-term traffic speed predictions: - What is the impact of using the direct and recursive strategies on the accuracy of RNN models as compared to using the multi-input-multi-output (MIMO) strategy? - What is the impact of different aggregation intervals for the input data? - What is the impact of including additional variables such as volume, occupancy, time of day, day of week, and spatial location as represented by station or sensor id? Our study resulted in the following observations, conclusions, and recommendations: - We observed that GRU architecture based RNN models had better accuracy as compared to other architectures. Thus we recommend that modeling efforts start with GRU architecture. - We observed that combining direct strategy with RNN gives the same accuracy as MIMO (i.e. many-to-many RNN architecture), however, the time taken for building many-to-many RNN architecture is much less as compared to direct strategy. Thus we recommend avoiding the use of recursive or direct strategies and advise the use of many-to-many RNN architecture without combining any other strategies. - We observed that there was no significant difference in the accuracy of 5,10,15 minute aggregations and that adding additional variables does not necessarily always result in higher accuracy. In both cases, we observed that using autotuning with a Bayesian algorithm was able to reduce the error metrics to a smaller range for almost all the combinations of aggregations and multivariate features. Among different data aggregations and feature sets, we suggest converting these choices as hyper-parameters and let the Bayesian algorithm based hyper-parameter tuner select the best combinations for your dataset. With the above contributions, this dissertation seeks to give traffic engineers a better understanding of how to apply modern methods for multi-step-ahead short-term traffic speed predictions.
Notes
If this is your thesis or dissertation, and want to learn how to access it or for more information about readership statistics, contact us at STARS@ucf.edu
Graduation Date
2020
Semester
Fall
Advisor
Wiegand, Rudolf
Degree
Doctor of Philosophy (Ph.D.)
College
College of Engineering and Computer Science
Degree Program
Modeling and Simulation
Format
application/pdf
Identifier
CFE0008323; DP0023760
URL
https://purls.library.ucf.edu/go/DP0023760
Language
English
Release Date
December 2020
Length of Campus-only Access
None
Access Status
Doctoral Dissertation (Open Access)
STARS Citation
Fandango, Armando, "Investigation of Recurrent Neural Network architectures based Deep Learning for Short-Term Traffic Speed Forecasting" (2020). Electronic Theses and Dissertations, 2020-. 352.
https://stars.library.ucf.edu/etd2020/352