Gast ::
| Hauptseite > Publikationsdatenbank > Data-Driven Modeling for Digital Representations in Energy Systems |
| Hauptseite > Publikationsdatenbank > Data-Driven Modeling for Digital Representations in Energy Systems |
| Book/Dissertation / PhD Thesis | FZJ-2026-02710 |
2026
Forschungszentrum Jülich GmbH Zentralbibliothek, Verlag
Jülich
ISBN: 978-3-95806-927-5
This record in other databases:
Please use a persistent id in citations: urn:nbn:de:0001-2608041520547.720046874476 doi:10.34734/FZJ-2026-02710
Abstract: The growing complexity arising from new technologies, particularly their interconnections, enabling dedicated planning, monitoring, and control of modern energy systems, raises the need for powerful concepts that provide detailed insights about the particular energy system and its components. In contrast to classical models that provide a generalized system description, this thesis investigates the gap between the physical and digital worlds, yielding a replica of the physical reality of energy systems and their components. We propose a definition and classification scheme for digital representations, including digital models, shadows, and twins. To provide digital representations, we investigate data-driven grey-box models that balance known and unknown system properties. In particular, we encode system characteristics into Gaussian processes, yielding hybrid models between equation-based and probabilistic machine learning models. We provide a system-specific modeling scheme and an algorithmic solution that guarantees numerical stability. We propose and analyze applications in three energy-related domains. First, we propose a digital twin for power system data quality in large-scale monitoring platforms to detect and treat erroneous data. We propose a signal recovery technique that can recover large periods of missing data of power systems signals based on only local power systems topology. We provide information and detailed insights about key characteristics of the proposed digital twin that generate practical benefits for system operators. In particular, a self-aware failing indication allows system operators to directly evaluate the recovered data. Second, we propose and compare four different types of digital shadows for non-invasive parameter estimation to characterize power electronics converters. We investigate and compare a forward-solver-based approach, a physics-informed neural network, and two different classes of physics-informed Gaussian processes. We discuss their differences as well as their advantages and disadvantages regarding power electronics parameter estimation. In particular, we propose a new class of physics-informed Gaussian processes and a dedicated model-building scheme that allows the application to power electronics converters in the field. By defining affine operator kernels, we show that the proposed digital shadow significantly outperforms previous alternatives when applied to power converters in the field. Moreover, we show that converter parameters can be estimated subject to the particular operational regime, yielding detailed operational insights. Third, in the area of building operations, we propose a digital twin for the generation of pseudo-measurements usable by advanced control systems. Continuously updated by reference measurements conducted at convenient locations, the proposed digital twin provides pseudomeasurements at user-relevant locations. We provide detailed information on different types of realizations, such as model-building characteristics regarding the number of reference inputs and long-term applicability. The given algorithms can be trained on a minimal data set, which immediately enables practical applications. We show that the proposed digital twin trained on a couple of days in summer can be operated in winter
; 
|
The record appears in these collections: |
PDF
Fulltext by OpenAccess repository