Generative Digital Twin for Predictive Modeling in Dynamic IoT Scenarios

Digital Twins (DTs) have emerged as a powerful tool for simulating, predicting, and optimizing the behavior of complex real-world systems. In many cases, DTs rely on real-world data generated by Internet of Things (IoT) systems and leverage Deep Learning (DL) techniques for predictive future system states. However, DTs often struggle to adapt to dynamic scenarios where additional loT devices and new configurations are introduced, requiring costly data collection and retraining. In this paper, we address this challenge by exploring the use of Generative Deep Learning (GDL) to build robust DTs capable of handling sparse, noisy, and biased sensor data streams. We propose a transformer-based GDL model that encodes variable-length collections of loT devices while supporting predictive capabilities under novel device configurations and the addition of new loT devices. We evaluate our proposed GDL model in two loT scenarios: smart homes and smart hydraulics systems, where physical models are used to generate the ground truth data. Experimental results demonstrate our approach's adaptability without retraining, achieving an <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R^{2} =99.5\%$</tex> when adding a temperature sensor in the smart home scenario (only 0.3% degradation), and <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$R^{2}=83.8 \%$</tex> with 5 valves in the smart hydraulics scenario, comparing favorably to VARX despite VARX being retrained on each dataset.