Probabilistic Load Forecasting

Interview with:
Florian Oppermann
Math.-techn. SW-Devoloper at the Institute for Automation of Complex Power Systems
RWTH Aachen University

Interviewer: Florian Oppermann, the Probabilistic load forecasting is one of the key results of the Platone project whose development was coordinated by the consortium partner RWTH Aachen University. Let´s assume I am a member of the R&D department at a German DSO and I mostly spend my days developing forecasting solutions required for the energy management systems used in our different field trials. In one of our field trials with a considerable amount of PV installations and residential households, we are using the scaled standard load profiles and the PV forecast provided by the weather data provider to provide forecast data to our energy management system. However, we are quite willing to improve the forecast accuracy as we have managed to collect measurements for the last three years.

Florian Oppermann: This seems to fit the specifications of ProLoaF very closely. ProLoaF provides a software package centered around a deep learning model that is specifically designed for sequence to sequence forcasting. Additionally, the package provides easy ways for data pre-processing and hyper-parameter-tuning without the need to get deep into the software.

Interviewer: I would like to ask a couple of questions in this regard: Is the ProLoaF service a stand-alone service or does it depend on other services?

Florian Oppermann: The software package can be installed as standalone python package directly from PyPi. The forecasting service depends on the availability of some other components like a timeseries database and a message broker. These can however easily be installed in one go using helm charts, so that the user can view it as closed off package.

Interviewer: Are there any tutorials? Before installing the service, can I give it a try via some Jupyter notebook examples?

Florian Oppermann: Yes, the package is available on Github where there are also Notebook tutorials available. These are both educational and instructional depending on whether you want to know more about usage or the inner workings.

Interviewer: What is the main language used for this service? What are the minimum SW and HW requirements?

Florian Oppermann: The package is exclusively written in Python, using PyTorch as deep-learning framework. For the service, we chose FastAPI as API framework. For the interaction with the REST-API, any http client can be used independent of programming language.

For the installation of the service a running kubernetes setup is required and for anything but the simplest of trials it is recommended to run the worker on a machine with a CUDA enabled graphics card as the model training can take a long time otherwise. Once a model is trained, there is not need for GPU access. RAM usage depends on the used model configuration, we have run demo versions on a 6GB machine, so this is possible though not recommended.

Interviewer: What about previous experiences of using this service in other projects?

Florian Oppermann: The core model of the forecasting package was developed during the Coordinet project and is in use in the internal software used by EON Sweden and it might be picked up again in the BeFlexible project. The package is also available via Allianders OpenSTEF though with reduced features.

Interviewer: What about support? Are you or somebody at your organization available to grant help if needed? How can I get in touch with you?

Florian Oppermann: While I am not the core author, I am one of two main contributors to the package and the sole developer of the service, so feel free to contact me via Mail (florian.oppermann@eonerc.rwth-aachen.de) or the LFEnergy Sogno slack channel,  for Sogno. The service version of ProLoaF is currently not used in other projects apart from Platone. However, more refinement of the API is more than welcome.

Interviewer: Can we have also support for the SW development? Like debugging, calibration, etc.? Is the SW a stable version?

Florian Oppermann: Of course, we are always willing to help and answer questions and are especially happy about feedback that improves our software. Note however that we are a small team and that resources to work independent of research projects are limited. The project itself is open source and development and contribution to the project from non-RWTH members are welcome.

Interviewer: If an ICT department is very reluctant about using open source SW, as they take the security precautions very seriously: I assume the SE and other services of DSOTP use dependencies that might be prone to vulnerabilities. Has the dependency scanning been done for the DSOTP services?

Florian Oppermann: We are using both DependaBot and code scanning for vulnerability detection.

Interviewer: Is there a specific publication that can explain to me the mathematical background behind deep-learning algorithms of ProLoaF?

Florian Oppermann: Yes, we have two publications on this topic:

Probabilistic Load Forecasting for Day-Ahead Congestion Mitigation (G. Gürses-Tran, H. Flamme and A. Monti),
2020 International Conference on Probabilistic Methods Applied to Power Systems (PMAPS), Liege, Belgium, 2020

Advances in Time Series Forecasting Development for Power Systems Operation with MLOps (Gürses-Tran, Gonca, Monti, Antonello),
Forecasting 4(2), 2022

Interviewer: Thanks a lot for the interview, Florian Oppermann!