• Citer

From Prediction to Causal Inference: Time Series Modeling for Next-Generation Digital Twins

par Paldino, Gian Marco
Président du jury Defrance, Matthieu
Promoteur Bontempi, Gianluca
Publication Non publié, 2026-03-03

Thèse de doctorat

• Accès en ligne
• Détails
• Contenu
• Statistiques

Résumé :

Despite the promise of Digital Twins, current data-driven models often fail in industrial settings. Such models are often black-box approximations that struggle with data scarcity, non-stationary environments, and the inherent complexity of physical topologies. This doctoral thesis proposes a methodological framework to evolve the Digital Twin from a passive observer into an intelligent, adaptive, and causal agent capable of robust operation in critical infrastructure and manufacturing. The research follows a trajectory from observation to structural understanding, validating contributions across three industrial domains: national power grids, high-voltage transmission assets, and precision manufacturing. First, addressing the challenge of Context-Awareness, we develop a hybrid framework for Dynamic Security Assessment in power grids. By explicitly engineering topological features derived from graph theory and combining supervised learning with unsupervised anomaly detection, we show that incorporating physical connectivity enhances proactive security predictions. To address the prohibitive costs of sensor deployment, the second part of this research focuses on the ’Cold Start’ problem in Dynamic Thermal Rating. We show that traditional data-heavy training is economically unviable for distributed infrastructure. By introducing parameter-based Transfer Learning, we create ´´Virtual Sensors” that generalize physical knowledge from data-rich source domains to data-poor targets. This approach yields accurate thermal ratings with minimal local calibration, outperforming standard IEEE physics-based models. Moving beyond statistical correlation to causal inference is the subsequent third step. We introduce the Temporal Dependency to Causality framework, a supervised learning approach that infers causal directionality from time-series asymmetry. This approach distinguishes invariant mechanisms from spurious correlations, achieving state-of-the-art zero-shot generalization on complex synthetic and biological benchmarks. We synthesize these elements into an Adaptive and Causal Digital Twin, applied to the high-dimensional non-stationarity of the plastic injection molding industrial process. We show that naive domain adaptation in non-stationary processes can lead to ´´destructive adaptation”. By constraining adaptation with invariant causal features, the proposed framework improves robustness in the studied domain. We conclude that true industrial intelligence would benefit from models that are not only predictive but also topologically aware, adaptable to new contexts, and grounded in causal structure.