The project seeks, as a proof of concept, to use machine learning to create a surrogate model from a” black box” model of an AC/DC converter. The black box model and the surrogate are to be of the type used in PSCAD, a type of electromagnetic transient (EMT) simulation software.
The aim is to create a surrogate model that has sufficient accuracy that it can be used by ESO in stability studies.
The surrogate model must be available as source code that can be recompiled so it can work on all future software systems, and it must be able to run at different time steps to ensure compatibility with other converter models (whether surrogate or black box).
Benefits
- Avoids issues with incompatible black box models.
- A single surrogate model could represent many diverse converters connected to the same substation.
- At present converter manufacturers do not maintain models long-term, so models cease to work as the simulation environment is updated. In contrast the source code for surrogate models is available, so they can simply be re-compiled to work with the latest software.
- Potential for faster execution of studies, as surrogates avoid the “slowest ship in the convoy” effect.
- Potential for simpler, standardised, converter models without software and legal complexities. This would make building EMT grid models much easier.
- Avoidance of “lock-in” with particular EMT modelling software – surrogate models can be re-compiled to work with any new software.
Learnings
Outcomes
The project has succeeded in providing a proof of concept:
> A “surrogate model” has been created that was trained to reproduce the behaviour of a wind farm. This wind farm was in turn composed of several black-box models connected by transformers and cables.
> The surrogate model was trained by recording the behaviour of the original wind farm. No assumptions were made about the wind farm or the black-box models within it.
> The surrogate model was implemented as a custom component in PSCAD using the Fortran and C languages. It can run alongside black-box models.
> All code and parameters used by the surrogate model are available to NESO, allowing the model to be recompiled for use in any future simulation environment.
> Tests in PSCAD showed good accuracy. In nearly all cases the surrogate’s real and reactive power outputs tracked those of the original manufacturer provided model closely. Exceptions were mostly situations where the surrogate correctly predicted instability with a limit cycle but had results offset by a few miliseconds from the original model, or where the surrogate correctly predicted chaotic behaviour. Importantly, in none of these cases did the surrogate model incorrectly predict stability when the original model was unstable or vice versa.
Lessons Learnt
The lessons learned were the following:
> Data driven neural networks were the most suitable approach in comparison with Physics Informed Neural Networks (PINN) due to their flexibility
> Training data needed to be tuned to capture the unstable behavior of the model in this use case, in contrast to usual machine learning use cases where instability is something to be removed
> Generic model testing in PSCAD before moving on to black box training and testing was not needed due to how well the data-driven approach preformed avoiding the need of introducing domain knowledge to the model
> The timestep of the surrogate model needed to be flexible enough to run slow enough for the fastest model intended to use in a study so while high timesteps are possible they may not be always practical if the intention is to use it alongside normal EMT models
> We identified a tradeoff between training time and neural network timestep, high training model timesteps introduced high frequency noise components while shorter timesteps resulted in larger data sets and subsequently training times
> Composition of training data needed to account for noise rejection behavior through different amounts of noise infused and clean data in combination with testing
