Reliability-based Urban Traffic State Prediction with Multi-source Data via Bayesian Learning

Abstract

This thesis focuses on reliability-based urban traffic state prediction with the use of multiple data sources. Precise and reliable predictions of urban traffic state and associated variability are essential to the development of reliability-based intelligent transportation systems. With the advancements in information and communications technologies, proactive short-term prediction of traffic state variables is becoming increasingly feasible for mitigating traffic congestion and improving network efficiency. This offers opportunities for innovations in real-time traffic state and uncertainty prediction with multiple data sources via advanced data analytics and modeling techniques. Therefore, this thesis aims to develop models that are able to place the level of significance on the corresponding traffic state prediction.

We first present a generalized Bayesian data fusion algorithm for making full use of multiple traffic data sources to estimate journey time variability considering the prevailing traffic states. Feeding data collected from multiple data sources are classified based on the associated traffic conditions via a mixture distribution model, and the corresponding estimation biases of the individual data sources are determined by different statistical distributions. The proposed framework is implemented and tested on a Hong Kong corridor with actual data collected from the field, and different statistical distributions of prior and likelihood knowledge are applied and compared. The findings of the case study show significant improvement in the journey time estimations of the proposed method compared with the individual measurements. The results also highlight the benefit of incorporating the traffic state classifier and prior knowledge in the fusion framework.

The second work develops a distribution-free reliability-based prediction approach for estimating journey time intervals with multi-source data using a two-stream deep learning framework. The prediction framework consists of a long short-term memory module for extracting temporal features and a convolutional neural network module for extracting spatial-temporal features from the heterogeneous data. The precision and reliability of the prediction are assessed respectively by the balance between the Mean Prediction Interval Width and Prediction Interval Coverage Probability metrics. For computational effectiveness, a Gaussian approximation is adopted to formulate a smooth and differentiable loss function for training the prediction framework. The computational experiments are conducted based on a real-world Hong Kong corridor, where multi-source data including traffic and weather conditions are collected. The proposed framework shows significant improvements over existing methods in terms of both precision and reliability over a range of traffic and weather conditions.

The final work presents an integrated model-based and data-driven Bayesian framework for predicting the evolution of both traffic states and the associated variability. The proposed framework ensures the interpretability and stability of the predictions with an underlying dynamic linear state space model, and captures sophisticated dynamics of traffic variability via a data-driven recurrent neural network component. The framework is trained with a multivariate Gaussian negative log-likelihood loss function, and the Monte Carlo dropout is introduced during the training and testing processes to quantify both model and stochastic uncertainties. The proposed framework is implemented and tested with actual traffic data collected from a Hong Kong strategic route. The case study shows that the proposed prediction framework can simultaneously retain the interpretability of the results and capture the complex dynamics of the evolution of traffic variability.

In conclusion, this thesis contributes to the development of reliability-based intelligent transportation systems for providing uncertainty quantifications of traffic predictions with advanced data analytics and deep learning techniques. The findings from the practical implications indicate the potential to apply the proposed approaches to online intelligent traffic applications.

Date of Award	5 Aug 2024
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Andy CHOW (Supervisor)