Event-Aware Imputation and Prediction of Urban Traffic Using Deep Spatiotemporal Learning Models

Abstract

This dissertation presents a comprehensive investigation into the dual challenges of missing traffic data and the complexities of traffic speed prediction during social events, a topic of growing relevance in urban mobility systems. Urban centers are increasingly experiencing non-recurring disruptions caused by concerts, sports games, festivals, and other social activities, which introduce sharp deviations in regular traffic patterns. At the same time, traffic data, which are foundational for intelligent transportation systems (ITS), often suffer from incompleteness due to sensor failures, transmission errors, and insufficient probe vehicle coverage. This research addressed these intertwined challenges by developing a unified framework combining robust imputation methods with deep learning-based event-aware prediction architectures.

The first contribution is the development of a two-stage imputation pipeline that integrates ensemble-based and generative approaches. Random Forest models are employed to provide fast, robust estimates, while Generative Adversarial Imputation Networks (GAIN) refine the results, capturing complex dependencies and uncertainty. Experiments on Hamilton, Ontario data demonstrate that the framework reduces imputation error (MAPE) by 20–30\% compared to traditional methods, while maintaining scalability under varying missingness levels.

The second major contribution is the development of an Event-Aware LSTM (EA-LSTM) model that explicitly incorporates structured social event features—such as event type, timing, location, and attendance—into a spatiotemporal architecture combining Graph Convolutional Networks, bidirectional LSTMs, and attention mechanisms. The EA-LSTM significantly improves prediction accuracy during disruptions, reducing average error to 3.4\% network-wide and under 9\% near event venues, outperforming conventional deep learning baselines.

The findings demonstrate that integrating contextual event information enhances both traffic imputation and prediction, leading to more robust, interpretable, and scalable models. The research provides practical insights for the deployment of real-time ITS applications, offering tools to support congestion management, dynamic signal control, and event traffic planning in complex urban environments.