Date of Award
Fall 2024
Degree Type
Restricted to Claremont Colleges Dissertation
Degree Name
Mathematics, PhD
Program
Institute of Mathematical Sciences
Advisor/Supervisor/Committee Chair
Yan Li & Allon Percus
Dissertation or Thesis Committee Member
Marina Chugunova
Terms of Use & License Information
This work is licensed under a Creative Commons Attribution 4.0 License.
Rights Information
© 2024 Zhengming Song
Keywords
Deep Learning, Graph Neural Networks, Neural Network, Spatiotemporal Forecasting, Time Series Forecastin, Traffic Speed Prediction
Subject Categories
Mathematics
Abstract
Accurately forecasting complex, multivariate time series varying across space and time requires models that effectively capture spatial and temporal dependencies inherent in the data. This dissertation introduces a novel architecture based on Graph Neural Networks that addresses the challenges of time series prediction, while also providing unique insights through decomposition of model outputs in the spatio-temporal space. We apply our method to the problem of traffic speed prediction, using benchmark datasets that we curate from Caltrans Performance Measurement System data. These datasets include detailed traffic speed and flow information across hundreds of sensors in four geographic districts in California, as well as spatial configurations like road networks, Euclidean distances, and temporal graphs. We present comparisons of our model with existing methods from the literature. We investigate how different graph structures and the inclusion of external factors, such as auxiliary data, influence the model's predictive performance. Our findings reveal that temporal information plays a critical role in improving accuracy, while spatial features yield mixed results depending on regional characteristics. Incorporating external factors such as traffic flow data generally enhances our model's performance, with the degree of improvement varying based on context and model configuration. Finally, we propose guidance to practitioners on interpreting the results from the model output decomposition.
Recommended Citation
Song, Zhengming. (2024). Spatial-Temporal Multivariate Time Series Forecasting Using Graph Neural Networks, with an Application to Traffic Speed Prediction. CGU Theses & Dissertations, 889. https://scholarship.claremont.edu/cgu_etd/889.
