Canzaniello, Marzia (2026) From Prediction to Simulation: Graph-Driven Generative AI for Urban Digital Twins. [Tesi di dottorato]

Documento PDF PhD_Thesis_Canzaniello.pdf Visibile a [TBR] Amministratori dell'archivio Download (31MB) | Richiedi una copia

Abstract

This thesis investigates the integration of graph-conditioned Generative Artificial Intelligence techniques for predictive modeling and the simulation of complex scenarios within Urban Digital Twins. The primary objective is to transcend the limitations of traditional deterministic forecasting by introducing generative frameworks capable of producing counterfactual simulations and ``what-if'' scenarios that rigorously adhere to spatial and temporal constraints. By fusing heterogeneous data sources, including mobility flows, urban infrastructure layout, resource usage, and environmental conditions, the research proposes a unified paradigm that bridges the gap between static prediction and dynamic simulation. The methodological core leverages Graph Neural Networks (GNNs) to encode the relational structure of the urban fabric, coupled with Generative Adversarial Networks (GANs) and Conditional Variational Autoencoders (CVAEs) to model the stochastic nature of urban dynamics. The proposed framework is validated across three distinct real-world domains: municipal waste management (CAPTURE), traffic flow forecasting (GRAPHITE), and urban parking simulation (DTMOB). The results are evaluated not only through standard accuracy metrics but also in terms of decision-making impact, demonstrating significant improvements in service efficiency, sustainability, and operational resilience. Ultimately, this work positions Graph-Driven Generative AI as a pivotal technology for the governance and optimization of next-generation smart cities.

Downloads

Actions (login required)

Modifica documento