Physics Informed Neural Networks for predicting temperature and loss distribution in power assets

Purpose and goal:

The aim of the project is to further advance knowledge for real-time dynamic operation of power transformers and other grid components using innovative AI methods. Physics-informed machine learning can enable the improved operation and provide more knowledge on heat distribution in power components which in return can benefit power system management and future component design.

Main project objectives: - apply physics-informed neural networks on data from operational equipment; - use a trained model to predict the thermal performance of power equipment for real-time operation. Expected results and effects: The technical results the project is planning to achieve are: - 2D and 3D representation of power transformer using PINNs

- Functional and robust prediction model able to predict power assets temperature distribution 24h ahead - Tested and validated versions of 1D, 2D and possibly 3D models as well as prediction model - Earlier stages of development of the combined digital twin model

The expected effect is better manageable and controllable power components and passage to better utilization and component use. Approach and implementation:

KTH will be the project coordinator, Hitachi Energy is a project partner. The project will be carried out in collaboration between researchers at Hitachi Energy namely Dr. Michele Luvisotto and Dr. Tor Laneryd researchers at KTH namely Dr. Kateryna Morozovska, Federica Bragone and newly hired member of staff for this project specifically.

The project is divided into 6 working packages which will be executed from the project start on May 31st 2023 until the completion planned for May 2026.