Digital Twin for Urban Transportation

QC 20260511

Abstract

This licentiate thesis develops and prototypically demonstrates Digital Twin (DT) architectures for urban transportation with a focus on road-traffic emissions, Public Transportation (PT) and data-driven path flow estimation to support DT simulation workflows. To advance current transport DTs from isolated sensor dashboards and disconnected simulations, the thesis addresses three research objectives: 1) to formulate a transferable DT architecture for network-level road-traffic emission nowcasting, forecasting, and retrospective analysis, 2) to design an open, automated, and extendable DT development pipeline for PT, and 3) to investigate the robustness and transferability of Partial Least Squares Regression (PLSR) for reconstructing path flows from link flow observations. 

Paper I develops and demonstrates an emission-oriented DT framework that integrates camera-based sensing, demand estimation, hybrid traffic simulation, and interactive 3D visualization. Traffic cameras are processed using a computer-vision pipeline that detects and classifies vehicles and extracts speed and acceleration used to nowcast emissions at camera locations. To further estimate network-level emissions, the camera data is used to estimate dynamic OD demand as input for microscopic simulation. A Unity-based 3D platform integrates sensor and simulation output using MQTT-based data streaming. This platform enables joint exploration of near-real-time emissions at sensor locations and simulated emission estimates in the surrounding network. A case study in Kista, Stockholm, illustrates the framework’s ability to support both emission nowcasting and scenario analysis, for example by assessing changes in network-level emissions under reduced parking availability.

Since DTs integrate multiple complex components, their development is often resource-intensive and time-consuming. To lower entry barriers Paper II proposes an open, automated DT development pipeline for PT that relies on open data, open standards, and open source software. PT operations are represented using GTFS data, where its static component serves as input for microscopic traffic simulation to enable joint simulation of PT vehicles and private traffic interactions. The GTFS real-time feeds enable both monitoring of current PT operations through low-latency visualization and retrospective analysis of events using a database storing historic observations. The link- and vehicle-level traffic data is displayed together with automatically derived OpenStreetMap building models in a Cesium-based web platform. This interactive visualization allows users to switch between nowcasting, scenario-based forecasting, and playback of historical operations within a 3D spatial context facilitating informed decision making. A case study in Kista, Stockholm demonstrates the pipeline’s technical feasibility by showcasing real-time PT operations and simulation-based scenarios visualized in the 3D interactive DT platform.

Paper III formulates and evaluates a PLSR-based path flow estimator as a data-driven alternative to conventional OD matrix estimation (ODME) within DT simulation workflows. Estimating a larger number of OD pairs from a smaller number of link counts, as well as collinearities in the observations render the path flow estimation problem ill-posed. PLSR learns a low-dimensional latent representation that maximizes the covariance between observed link flows and path flows, providing implicit regularization for the ill-posed inverse problem. While the method was used for similar problem settings, it has not yet been used for path flow estimation. Thus, this study evaluates its suitability by assessing its stability and transferability using a synthetic test network and controlled data-generating processes that reflect practically relevant OD and route choice structures. The experiments indicate that PLSR achieves the lowest reconstruction errors when variability in path flows is dominated by OD demand fluctuations. Increased path choice randomness, however, is reducing recoverability at first, but the performance stabilizes once strong path competition regimes are entered. The experiments further indicate that PLSR transfers reliably when fitting and deployment share the same correlation regimes, but performance deteriorates near regime boundaries where OD-driven correlations give way to path choice competition-based correlations. These findings suggest that PLSR can serve as a fast, data-driven path flow estimator in DT contexts. Though, to ensure reliable estimates over time it requires continuous monitoring of the underlying correlation regimes to detect shifts and retrain the model when needed.

Link to DiVA