Wang, YinhaiTsai, Meng-Ju2023-08-142023-08-142023Tsai_washington_0250E_25402.pdfhttp://hdl.handle.net/1773/50283Thesis (Ph.D.)--University of Washington, 2023The rapid advancement of intelligent traffic sensing and communication technologies has introduced a new era of transportation data, offering unprecedented opportunities to predict and manage urban traffic. However, traditional statistical and basic machine learning models are often inadequate in forecasting network-wide traffic states, hampered by the time-varying nature of traffic patterns and the complex geographical relationships on road networks. To compound this challenge, unexpected events like the COVID-19 pandemic have drastically altered traffic patterns, making it harder for transportation agencies to learn representative patterns from historical data. The above underscores the critical need for more advanced models that can adapt to changing conditions and deliver reliable predictions. Building on this pressing issue, the goal of this dissertation is to develop advanced deep learning models in both methodological and practical ways to improve traffic forecasting accuracy under non-stationary circumstances. This dissertation aims to accomplish the goal in six parallel perspectives. Firstly, a model with the capability of capturing patterns from both short- and long-term traffic states should be developed to accommodate unexpected interventions. Secondly, a workflow with customized data processing and analysis components should be designed for extracting other meaningful auxiliary information that could improve the robustness of relative long-term traffic forecasting, such as social media features. These features can then be integrated with traffic data and fed into a model with long-term prediction capability to enhance the robustness and accuracy of network-wide prolonged traffic forecasting. Thirdly, a model able to learn new traffic patterns without forgetting previous knowledge under continuously changing traffic conditions should be created to demonstrate how to tackle the Plasticity-stability dilemma, especially under non-stationary traffic conditions. Fourthly, a novel unified framework with multi-contrastive learning should be developed to improve the robustness of spatial-temporal traffic forecasting, which has a great potential to effectively handle complex and noisy data and learn fine-grained representations suitable for traffic forecasting. Fifthly, a real-time interactive application should be implemented to evaluate live traffic updates and predict future traffic states, enabling drivers to plan their routes more efficiently and reduce congestion on the roads. Lastly, a benchmark should be provided for researchers to expedite researchers to uncover more informative patterns from non-stationary data and evaluate the resilience of models in the transportation industry. This dissertation conducts in-depth research and applications on several key technologies and steps required for building more adaptive and robust architectures. They will address several critical transportation necessities and provide tangible benefits for traffic management and optimization. Specifically, the contributions can be divided into six perspectives: 1) proposing a Multivariate Dual Long Short-term Memory model. It considers short- and long-term traffic patterns and spatial and temporal features for network-wide traffic forecasting under interference. 2) Learning social media features in a Natural Language Processing (NLP)-joined social-aware framework to overcome the ignorance of cultural impacts and boost robustness under unexpected interventions in prediction tasks. 3) Designing an incremental learning framework to solve catastrophic forgetting issues to build a more robust architecture given continuously changing traffic patterns. 4) Developing an innovative unified model with multi-contrastive learning and traffic representation learning to mitigate the challenges of handling complex and noisy traffic data, enabling improved spatial-temporal traffic forecasting capabilities. From a practical standpoint, the contributions are 5) implementing a real-time traffic performance measuring platform to assess current traffic conditions and forecast future network-wide traffic states. 6) Releasing benchmarks and a non-stationary traffic dataset to encourage further research into developing powerful algorithms that can adapt to fluctuating traffic conditions. In conclusion, our research highlights the need for advanced deep-learning models to improve the accuracy and adaptability of traffic forecasting under non-stationary circumstances. The proposed techniques provide promising solutions to overcome traditional modeling challenges and offer practical applications for real-time traffic management. With continued research and development in this field, we can pave the way for smarter, more efficient, and sustainable urban transportation systems.application/pdfen-USCC BYTransportationCivil engineeringThesis