Modeling and Analysis of Data-Driven and Knowledge-Enhanced Traffic Prediction
Year: 2026
Type: dissertation
Abstract: The rapid growth of urban populations has put unprecedented pressure on transportation systems, causing widespread problems such as congestion, accidents, and environmental pollution. Accurate traffic prediction is therefore essential for building smart cities and intelligent transportation systems, as it enables evidence-based policymaking, risk assessment, and efficient traffic management. Early research on traffic prediction relied mainly on statistical models and machine learning techniques, which struggled to capture complex nonlinear spatio-temporal dependencies in large-scale networks. The advent of traffic big data has prompted a paradigm shift towards deep learning approaches. Recent advances in traffic prediction have increasingly integrated advanced temporal models with Graph Neural Networks to model complex temporal dynamics and localized spatial patterns, which yield incremental yet meaningful performance gains. However, these existing approaches still face four key limitations: (1) Traffic patterns exhibit cross-scale temporal evolution (ranging from minutes to months). (2) Spatial correlations are dynamic and multi-scale among different time scales. (3) Traffic data involves not only pairwise spatial relationships but also high-order functional dependencies among multiple regions and short or long-term temporal dependencies across different time periods. (4) Current data-driven models often lack the constraint of domain knowledge, which may lead the models to learn certain false correlations arising from data noise or randomness, thereby limiting the predictive performance. In response to these challenges, this dissertation investigates four core research questions: (1) How to learn cross-scale temporal dependencies? (2) How to capture multi-scale spatial dependencies? (3) How to model hierarchical spatio-temporal semantic associations? (4) How to integrate domain knowledge into data-driven traffic prediction models? Specifically, four traffic prediction models are developed: (1) Cross-Scale Temporal Relationship Representation Method. Dynamic temporal cross-scale representation model (DSTMR) adopts a hierarchical residual-like layer with gating mechanisms to capture multi-scale temporal dependencies. A dynamic graph generation module is designed to adaptively learn dynamic spatial correlations and generate a dynamic spatial graph structure. It further integrates a graph neural network to model spatial correlations, and is applied to bus station passenger flow forecasting. (2) Multi-Scale Spatial Relationship Representation Method. A multi-scale graph structure learning model (DMGL) is developed. This model optimizes graph structure learning strategies and constructs multi-scale graphs to capture spatial information at different time scales. Specifically, a dynamic multi-scale graph generation module is first designed to construct multi-scale dynamic graph structures through time decomposition and aggregation. Experimental results on real taxi regional passenger flow datasets demonstrate that DMGL achieves significant improvements in prediction performance compared to baseline methods. (3) Hierarchical Spatio-Temporal Semantic Associations Representation Method. A spatio-temporal mixed-order relationship representation model (MixSTRL) is designed. It introduces a spatial mixed-order representation that combines a graph neural network for pairwise spatial dependencies and hypergraph neural networks for high-order spatial relationships, integrated through a squeeze-and-excitation mechanism. Furthermore, it constructs a temporal mixed-order representation module that fuses a temporal convolutional network for short-term dynamics with temporal hypergraph neural networks for long-range periodic dependencies, employing a gated fusion strategy. Experimental evaluations on three taxi regional passenger flow datasets verify the effectiveness. (4) Knowledge-Enhanced and Hierarchical Spatial Pattern Learning Method (KA-HSPL). This method first designs a spatio-temporal embedding module that integrates knowledge-augmented spatial embedding and period-aware temporal embedding. Then, it develops a hierarchical spatial pattern learning module, which combines an adaptive graph convolutional network, a customized hypergraph convolutional network, and a spatio-temporal fusion module to model both pairwise and high-order spatial relationships. Finally, it employs a time-aware graph convolutional network to capture dynamic temporal correlations and devises a spatio-temporal joint graph convolutional network for integrated spatio-temporal dependency learning. The experiments on six highway section speed datasets demonstrate the model's effectiveness. (5) Intelligent Traffic Early Warning Demo System. To integrate and implement the designed four urban traffic prediction models, this dissertation develops an intelligent traffic early warning demo system. Focusing on three core tasks: bus station passenger flow prediction, taxi regional passenger flow prediction, and highway section speed prediction, the system constructs three representative application scenarios, including bus scheduling optimization warning, taxi resource dispatching warning, and road congestion state warning. This system effectively supports the transformation of urban traffic management from passive operation to proactive early warning and from experience-driven to data-driven decision-making, and provides a solid foundation for further development and extended application of intelligent transportation systems.
Author: Lilan Peng
Language: English
Subfield: Transportation
Field: Social Sciences
Domain: Social Sciences
Open Access status: green
OpenAlex ID: W7222948398