Data-driven and machine learning approaches for transport planning and management

Road safety and transportation demand forecasting remain critical challenges in urban mobility and planning. This research addresses two key problems: the accurate prediction of accident severity and transport demand forecasting using machine learning approaches. The severity of traffic accidents is influenced by multiple interacting factors, including road conditions, traffic flow, and environmental variables, yet existing models struggle with imbalanced data, lack of interpretability, and predictive uncertainty. Similarly, demand forecasting models must capture complex spatio-temporal patterns in urban mobility, which traditional methods fail to adequately address. For road safety and accident severity prediction, this study employs an XGBoost-based machine learning framework that integrates accident records and network-derived traffic flow data. The dataset consists of accident records from Rome spanning 2006 to 2022, enriched with traffic assignment-derived flow variables. To address the class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) is applied, while SHAP analysis is used to interpret feature contributions. Conformal Prediction is incorporated to quantify predictive uncertainty, ensuring reliability in severity classification. The results show that integrating traffic flow data enhances prediction accuracy, particularly for severe crash cases. The analysis reveals that factors such as speed, congestion, and road design play pivotal roles in determining crash severity. For transport demand forecasting, deep learning architectures, including Bi-Directional Long Short-Term Memory (Bi-LSTM), 3D-Convolutional Neural Networks (3D-CNNs), and Temporal Graph Networks (TGNet), are implemented to predict urban travel demand. The study utilizes the NYC Yellow Taxi dataset, leveraging high-resolution spatio-temporal data to train models that capture both spatial and temporal dependencies. The results indicate that deep learning models outperform traditional methods, but their effectiveness is highly dependent on data granularity and representation. This thesis contributes to the fields of road safety modelling and transport demand prediction by introducing machine learning frameworks, integrating traffic assignment techniques, and enhancing model interpretability. The research has significant implications for engineering design and transport policy, enabling proactive road safety interventions, efficient resource allocation, and improved urban mobility planning.