Spatiotemporal Deep Learning Models for Non-Recurring Congestion Prediction with Multiple Data Sources

Unexpected events, such as car crashes, result in non-recurring congestion (NRC), which increases the challenge of traffic management. Current NRC prediction methods are predominantly qualitative, leading to information loss, or quantitatively limited to single-step, short-term forecasts. This research introduces advanced quantitative, interpretability-enhanced, and spatio-temporal prediction methods with high spatial resolution, specifically tailored for NRC in highway and urban networks. For highway NRC, two novel deep learning models are introduced: the Dual-Stream Autoencoder Sequence-to-Sequence approach and the Two-Encoder-Decoder model with Attention mechanism (Att-2ED). Both models utilize separate encoders to independently process time-series speed data and static crash-related features. Additionally, the Att-2ED model incorporates an attention layer to prioritize input sequences. The models are developed and evaluated using real-world freeway data, demonstrating superior performance compared to existing benchmarks. Its effectiveness is further underscored through a comparative analysis of prediction accuracy across various congestion levels and crash severity. Moreover, the attention mechanisms incorporated in Att-2ED provide interpretability-enhanced predictions, highlighting the significance of the last 10-minute input in the prediction. For urban network NRC, this thesis introduces a hybrid predictive framework combining an attention-enhanced Graph Convolutional Network with LSTM (Att-GCN-LSTM) to predict long-term, network-level NRC, whose outputs are subsequently utilized by a linear regression model to estimate the resulting excessive CO_2 emissions. The Att-GCN-LSTM integrates historical traffic data, road network topology, and non-recurring (NR) event characteristics to predict traffic conditions. The framework is evaluated using simulation data representing road networks in Hamilton, Canada. Comparative evaluations against baseline models confirm the proposed framework's superior predictive performance across different conditions. Two additional case studies further demonstrate the model’s efficacy. Furthermore, the emission analysis indicates that NR events on higher-level roadways or of longer duration increase emissions, and this impact gradually diminishes over time. This research provides transportation agencies and traffic management professionals with powerful, interpretable predictive tools for effectively managing and mitigating congestion caused by NR traffic events.

Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.