Welcome to transformer-thermal-model Discussions!

[GroenteLepel]

👋 Welcome!

We’re using Discussions as a place to connect with other members of our community. We hope that you:

• Ask questions you’re wondering about.
• Share ideas.
• Engage with other community members.
• Welcome others and are open-minded. Remember that this is a community we
  build together 💪.

To get started, comment below with an introduction of yourself and tell us about what you do with this community.

Here are some tips 💡 for getting started with Discussions:

📢 Announce to your community that Discussions is available! Go ahead and send that tweet, post, or link it from the website to drive traffic here.

🔗 If you use issue templates, link any relevant issue templates such as questions and community conversations to Discussions. Declutter your issues by driving community content to where they belong in Discussions. If you need help, here's a link to the documentation.

➡️ You can convert issues to discussions either individually or bulk by labels. Looking at you, issues labeled “question” or “discussion”.

[ziolai]

Greetings,

What kind of public domain simulations tools should we develop further to bridge the gap between the loading guide and measurements in the field? How should next-generation digital twins of transformers in the distribution grid look like?

How can the loading guide be extended to incorporate phenomena such as the the magnetic field induced by the windings, the losses induced by the magnetic field, the spatial distribution of the temperature, the lamination of the ferromagnetic core, the harmonics above 50 Hz, phase-unbalance, the thermal degradation of the oil and insulating material and the ventilation of the substation?

What can be gained by complementing the current implementation in numpy with an implementation on higher level of abstraction using scipy (among alternatives) in terms of model analysis, model flexibility, model maintainability, computational efficiency and extendability?

What approaches are available (e.g. least-squares fitting, Kalman filtering, neural networks) to complement models based on the loading guide with either measurements in the field or high-resolution numerical models?

How can academics contribute addressing the above issues? How should the assignment on this topic for bachelor and master students outline at 1 be extended further?

We love to hear from you. Thanks. Kind wishes, Domenico Lahaye.

[GroenteLepel]

Dear @ziolai, thanks for the questions! Great to see you be enthusiastic. I'm not really fit to answer most of your questions regarding the first two paragraphs, for that I will ask @ImkedeMan for more info.

Regarding the implementation of the model: for now we want to keep the computation in this model as simple as possible with a numpy core. This gives us a benefit of readability and speed, with the "loss" being "lesser" accuracy. We've used pandas before, but since the migration to numpy have reached much better runtimes. It might be worth looking into Polars to see what the benefits might be to make it more accessible for the datasets that are typically used for this model.

As for complementing our model, we use internally a simple scipy optimisation to scale the load profile we provide to the model until we reach provided load limits or thermal limits. This can of course be hugely expanded upon with more complex solutions, but we've found that the improvements in errors do not yet benefit our use cases.

I'm not sure if this example handles your mentioned "higher level of abstraction", so do clarify if it misses your point!

[martijn-janssen]

Dear @ziolai, thanks a lot for your valuable questions!
They demonstrate the many potential model improvements possible in the future and the good part is that we are not alone to investigate / implement them :-)
I can add to @GroenteLepel's answers that it is useful to keep evaluating the balance between:

• relevance of the model to the use-cases
  • Asessing (future) cyclic grid loadability (example paper [1] (this may include risk based asset management in the future by making use of the aging estimation)
  • Real time system operations (using 48h current profile forecast OpenSTEF)
  • Other use-cases
• explainability of the code
• complexity of the code
• speed of the code

[1] CIRED 2023 Cyclic loadability of entire HV/MV-substations, [online]: https://digital-library.theiet.org/doi/abs/10.1049/icp.2023.0973?download=true

[ziolai]

Dear Martijn-Janssen,

Many thanks for your input. Much appreciated.

Can you pls. send me a copy of your paper (the library does not provide immediate access).

Thanks again.

[ziolai]

Very nice!

I am tempted to proceed as follows:

Implement the equation for the top oil temperature (Theoretical Documentation => Model Description) as a first order scalar ordinary differential equation in DifferentialEquations.jl (an alternative for scipy) and proceed as follows:

1/ implement a disturbance (source term) that is not captured by the fixed time-step sampling of dt = 15 min (here fixed is important, not the actual value) and that is captured by a variable time step method;

2/ implement the absence of any disturbance to show that fixed a time step is doing more than required (i.e., the opposite of previous scenario);

3/ implement an extension to two coupled scalar equations (e.g. top oil temperature and hot spot temperature) (both physics based instead of by scaling one from the other) to show how easy this can be implemented;

4/ implement a combination of DifferentialEquations.jl and interval arithmetic to quantify how uncertainty in the input translates to uncertainty in the output. See e.g. https://docs.sciml.ai/SciMLTutorialsOutput/html/type_handling/02-uncertainties.html

5/ implement a combination of DifferentialEquations.jl and least-squares fitting to obtain an estimation of a coefficient in the model input (e.g. oil viscosity) from measured data of the model output (e.g. temperature). See e.g. https://docs.sciml.ai/DiffEqParamEstim/stable/tutorials/ensemble/

6/ Show the machine code that DifferentialEquations.jl generates under the hood to show that not only the code is fast, but explain you why it is fast.

How does this sound to you?

[GroenteLepel]

Thanks so much for this writeup, it's cool to see what you already have in mind for this codebase. We're still in the early stages, so we don't have a solid "vision" or "roadmap" set up for the model, so apologies from our side if we cannot put it to words properly, yet. We have discussed our goals verbally with the team already thoroughly, and have established that right now, we are mainly focussing on the robustness, consistency and trustworthiness of our model. Anything that adds to that would be great!

If I understand your list correctly, that would mean that points 1, 2 and 4 would contribute to that, but I must confess that it is still quite difficult to comprehend as a software engineer!

Would you be so kind as to make a new topic for these suggestions in the 💡Ideas channel? I can also make it for you, but this way you will stay the owner of the topic!
With this topic, we can keep the discussions for each of those central, keep this topic clear for newcomers, and establish which project we would mostly benefit from. With that new topic, you can get some space to provide more detail on these steps, we can ask some more questions, and eventually we can potentially decide what to include in ours or in your roadmap! You can just start by copy/pasting what you've already written down, and answering to this message, that would be great.

Again, thank you for your excitement!

[ziolai]

Thanks! I proceeded as you suggest. Talk soon.