Session Chairs: Alban Puech & Etienne Vos (both from IBM Research)
Building Your Own Neural Solver from Scratch with OpenGridFM
Alban Puech – IBM Research & ETH Zürich in the Main Auditorium
https://zoom.us/j/94726375664?pwd=xhYFSv3QqMNExvclWDZ3HyJlN8V2HZ.1
This hands-on tutorial introduces our end-to-end framework released through the OpenGridFM project, enabling the development and deployment of GridFM models for steady-state grid analysis.
Participants will work with notebooks to generate synthetic but realistic power-flow and optimal-power-flow datasets using gridfm-datakit, including load variation, generator-dispatch variation, topology perturbations, and configurable grid cases. They will then use gridfm-graphkit to train and evaluate GENCO, a unified graph neural solver, in a low-code environment and compare its performance against classical solvers.
The tutorial is designed for power-system practitioners, researchers, and developers who want to contribute to OpenGridFM or apply its tools in their own
workflows. No prior experience with the project is required; familiarity with Python and basic power-system concepts is helpful.
Tutorial on using gridfm-app-distribution
Mikka Kisuule – INESC TEC in B035
https://teams.microsoft.com/meet/3787830886026105?p=jl674u9voElqq5Lh5t
The tutorial aims to provide an in-depth overview of the gridfm-app-distribution repo and guide participants on how to use it in practice, with a hands-on demonstration of deploying a small dummy model. The tutorial will be divided into 3 parts:
Part A will provide a general overview of the repo, discussing the repo layout and how to access it; the structure of the input data files; the input configuration file; the Datamodule and the normalisers; the model architecture; the VLD loss function; and the specialised reconstruction task.
Part B will focus on a practical end-to-end implementation guide for using the application, beginning with local installation, tuning parameters for the model architecture and training, and how they are loaded into the NameSpace, data loading of the model, settings for the VLD learning and their impact on training and model performance, how to run model’s training, finetuning, prediction, and evaluation tasks, and the evaluation of the model’s prediction accuracy.
Part C will be a hands-on demonstration that will be interactive, with participants installing the repo (locally or on a remote Jupyter notebook/Google sheet), tuning a dummy model to train on a small randomised dataset, and going through all the steps, from data loading to reconstruction tasks and evaluation of the model’s prediction accuracy. In this part, the participants shall be guided one-to-one to overcome any challenges in deploying the model.
The final part of the tutorial will be a discussion of the way forward for the repo, detailing potential applications to develop in it, the need for more training datasets, and the fine-tuning of pre-trained models on real-world data.
From Forecast to Dispatch: Measuring What Probabilistic AI Is Worth in Electricity Markets
Louiza Ait Mouloud – Foresight Grid Inc. in B025
https://teams.microsoft.com/meet/3756787205470726?p=iFyPGTlfnfTA7YUmKI
Grid foundation models are usually judged by forecast error. Market participants judge them by money: a forecast matters only if it changes a decision and that decision earns more. This gap explains why many accurate models never reach an operator’s desk. This tutorial closes that gap using public ERCOT data. Participants take a probabilistic price forecast, turn it into a battery bidding decision under uncertainty, and settle that decision against real prices. We show how to value a forecast in dollars against a realistic benchmark, how a model with worse average error can earn more because it is right in the hours that matter, and why tail calibration drives value during scarcity. Winter Storm Fern, the first major scarcity event after ERCOT’s real time co-optimization redesign, serves as the running case: a single week in which forecast quality translated into hundreds of thousands of dollars of difference on one node. Participants leave with an open notebook that links any probabilistic forecast to a settled market outcome, and a clear method for showing operators, investors, and utilities what their model is actually worth.
PowSyBl and Dynawo in Practice: Generating Steady-State and Dynamic Power System Datasets for Foundation Models
Nicolas Lair – Artelys in B031
https://teams.microsoft.com/meet/3456396837887418?p=SOnhmLLzEYa3jxPEpZ
Hands-on tutorial on PowSyBl and Dynawo, two open-source industrial-grade tools for power system simulation, and how they combine in GridFM-Datakit to generate steady-state and dynamic datasets — including randomized initial states and dynamic events — for training grid foundation models. No prior experience with PowSyBl or Dynawo is required.