Physics-Informed Neural Networks for modelling temperature and loss distribution in power transformers swegrids-logo

SweGRIDS research area Controllable Power Components
SweGRIDS project code CPC19
Project type post-MSc
Status running
Researcher Federica Bragone   (webpage)
University KTH (EME)
Project period 2021-07-01 to 2021-12-31   
Project supervisor Kateryna Morozovska   (webpage)
Industrial sponsors Hitachi Energy

back to project index

Project abstract

Physics-Informed Neural Networks (PINNs) is a new method in machine learning addressed to solve and identify nonlinear partial differential equations (PDEs). PINNs rely on neural networks (NN) capability to approximate any function and combine this property with the physics outlined in nonlinear PDEs.

Dynamic thermal modelling for power transformers plays a fundamental role in studying their thermal behaviour and ensuring their efficiency and longevity. Conventional dynamic thermal models include the IEC and the IEEE standards however, one of the main disadvantages of these models is that they cannot provide the thermal distribution.

This project focuses on power transformers thermal modelling using PINNs to solve the one-dimensional heat diffusion equation. The aim is to predict the top-oil temperature while also estimating the thermal distribution inside the transformer. The Finite Volume Method (FVM) is utilised to calculate the PDE solution and to benchmark the PINNs predictions. Field measurements taken from a real transformer, including the top-oil temperature, the ambient temperature and the load factor, will guarantee a model very close to the real world.

The figure below shows a schematic representation of the PINNs model for the problem considered,

Summary of work

The Master thesis work carried out on this topic showed promising results that could be improved by changing various aspects of the model. Some of the changes include:
- Construct a different NN architecture to ensure a proper prediction of the top-oil temperature. 
- The boundary condition on the top, which includes the measurements of the top-oil temperature, is considered a Dirichlet boundary condition. The following step is to make it a Neumann boundary condition. 
- Use the dimensionless form of the PDE and normalise the data.
- Use the ambient temperature and the load factor as inputs of the NN (they are currently used as outputs).

During the project time, the aim is to address some of these changes to guarantee a better accuracy and efficiency of the model compared to the Master thesis results.  Moreover, another goal is to write and publish two papers regarding the results obtained. 

Event log


Project reference-group

Project-specific working group (frequent cooperation)
Tor Laneryd,  Hitachi Energy (supervisor)
Michele Luvisotto,  Hitachi Energy (supervisor)
Claes Ahlrot,  E.ON
Ola Ivarsson,  E.ON

General QED reference group (twice per year)
Tommie Lindquist,  Svenska kraftnät
Milan Radosavljević,  Svenska kraftnät
Susanne Olausson,  Energiforsk AB
Ying He,  Vattenfall AB
Susann Persson,  Jämtkraft AB
Matz Tapper,  Svensk Energi
Johan Öckerman,  Vattenfall
Erik Jenelius,  Trafik och logistik KTH
Fredrik Carlsson,  Vattenfall
Thomas Welte,  Sintef Energi
Robert Saers,  Hitachi Energy
Erik Lejerskog,  Ellevio AB
Andrew Kitimbo,  Vattenfall

Publications by this researcher

See alternatively the researcher's full DiVA list of publications, with options for sorting.
Publications in journals and conferences usually will not show until a while after they are published.

Publication list last updated from DiVA on 2022-04-16 02:35.

back to project index

Page started: 2021-07-01
Last generated: 2022-04-16