Problem description
A digital twin is a representation of a physical system that uses incoming measurements to update its models and support predictions and scenario analysis [1]. For power transformers, a digital twin could estimate internal temperatures, anticipate thermal stress, and help operators assess alternative loading decisions. Developing such a twin requires both reliable thermal models and software that connects sensor streams, maintains the transformer’s estimated state, and makes predictions available for monitoring and decision support.
This thesis will investigate digital twin architectures and Scientific Machine Learning (SciML) for transformer thermal monitoring. Lumped physical models may oversimplify behaviour, detailed physical models can be computationally expensive, and purely data-driven models may require extensive data. SciML combines physical knowledge with learning from data to investigate a balance between accuracy, data efficiency, physical consistency, and computational cost [2].
Example
Suppose an operator wants to assess a temporary increase in transformer loading. The twin uses recent sensor measurements to update its thermal state, then simulates temperature trajectories under alternative loading profiles. How reliable are these predictions if measurements arrive late or are missing? How should the model and data pipeline work together to keep the twin useful as conditions change?
Methodology
Develop a digital twin prototype using historical time-series data provided by Statnett. The architecture will connect data ingestion and validation, transformer state management, interchangeable thermal models, and a monitoring and scenario interface, while supporting future integration with live sensor data. Historical measurements will be processed chronologically to compare update strategies in terms of latency, state freshness, and robustness to simulated missing or delayed measurements. Modularity will be evaluated by replacing a model or data source.
Develop a physics-informed thermal model and compare it with physical and data-driven baselines using chronological training and test periods. Evaluate prediction error, training data requirements, physical consistency, and computational cost. Scenario experiments will explore alternative loading profiles under explicit assumptions about future loading and ambient conditions.
What you will learn
You will gain experience in digital twin representation and software architecture, sensor-data pipelines, and model synchronisation. You will also learn to combine physical knowledge with machine learning, evaluate dynamical models, and design experiments using industrial data.
Background
Python or Julia experience and basic knowledge of neural networks, statistics, and differential equations are helpful; prior power-systems knowledge is not required. The project is intended to use data from multiple Statnett transformers and may involve Statnett and SINTEF.
[1] X. Liu and I. David. Introduction to Digital Twins for the Smart Grid. arXiv:2602.14256v1, 2026.
[2] G. E. Karniadakis et al. Physics-informed machine learning. Nature Reviews Physics, 2021.