Abstract
Digital twins have a wide range of applications, and the interest in digital twins from academia and industry has increased substantially in recent years, which has been confirmed by annual market growth and the number of publications per year. The initially loose interpretation of the term digital twin is converging towards clearer concepts and classification schemes. However, despite the projected growth, many industries have been struggling with the implementation and utilization of digital twins, which leaves their full potential underutilized in several areas, including wind energy.
The thesis is built on 11 articles, which are contextualized here to collectively identify and alleviate the challenges in the industrial implementation of digital twins and thereby work towards value creation through the implementation of digital twins. Most topics are discussed in the context of wind energy, but many of the findings are transferable to other industries. The lack of accessible data for calibration and validation, the computational cost of models, and the limited awareness about the value of digital twins are only three of the numerous challenges in the implementation of digital twins that are identified and addressed in this work.
An integral component of addressing all challenges related to the modeling of the physical asset and analysis of measurements is the hybrid analysis and modeling paradigm, which combines physics-based models with data-driven methods to harvest the best properties of both approaches. Although all approaches deserve consideration, two methods have been selected here for closer investigation and demonstration, namely physics-inspired losses and physics-guided neural networks.
Despite the many advantages of hybrid analysis and modeling approaches, they do require data to be trained on. Limited data access, another central challenge faced in the integration of digital twins, can therefore restrict the usage of hybrid analysis and modeling. The combination of federated learning with hybrid analysis and modeling is introduced in this work to enable privacy-preserving model training across decentralized data.
To raise awareness about the value of digital twins, three digital twins have been integrated, demonstrated, and disseminated for a floating wind turbine, an onshore wind farm, and a laboratory-scale setup for thermography anomaly detection. The implementations include real-time data streaming, 3D models of the assets and their environment, anomaly detection algorithms, predictive capabilities, and data visualization through virtual reality interfaces. The digital twins have been demonstrated at internal and external events and have also been used for educational purposes, among other applications.