Year: 2026
Type: dissertation
Abstract: Deep learning (DL)-based traffic prediction models require complete sensor data, yet real-world datasets are incomplete due to sensor faults, coverage gaps, and com¬munication failures. This thesis develops a two-stage imputation and long-term prediction pipeline using extended PeMS03 data. First, it evaluates imputation under realistic missingness by varying mechanism, rate, and per-sensor distribution across uniform and heterogeneous sensor-specific settings. An Automated Machine Learning (AutoML) framework searches statistical, Machine Learning (ML), and DL methods to select an imputer per scenario. Results show no single method is universally op¬timal, supporting scenario-aware selection. Finally, it examines how spatiotemporal graph models generalize over one year. The Automated Dilated Spatiotemporal Synchronous Graph Modeling (Auto-DSTSGN) network is most robust, while forecasting accuracy degrades non-monotonically due to non-stationarity and seasonal shifts rather than temporal staleness alone. Overall, this work provides guidance for reliable traffic prediction under realistic missing-data conditions and long-term deployment.
Author: Khatiriolyaee, Mahshid
Institution: Ontario Tech University
Language: English
Subfield: Transportation
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7213989471
