3rd In-Person Workshop: Foundation Models for the Electric Grid

Feb 11, 2025, 8:30 AM → Feb 13, 2025, 12:00 PM America/Chicago

Room 1416 (TCS Conference Center)

Kibaek Kim (MCS)

Foundation Models for the Electric Grid
February 11-13, 2025 | Argonne National Laboratory

We are excited to announce the 3rd in-person workshop, Foundation Models for the Electric Grid, hosted at Argonne National Laboratory from February 11-13, 2025. This event will gather over 70 international stakeholders from academia, industry, and government agencies to drive forward the development and application of AI and maching learning foundation model technologies for critical challenges in electric grid systems. 

Workshop Goals:

• Explore the transformative potential of foundation models for electric grid systems.
• Address key challenges and opportunities identified in the positional paper, "A Perspective on Foundation Models for the Electric Power Grid", recently published in Joule.
• Foster collaboration among leading researchers, practioners, and policymakers to advance the state of the art in electric grid technologies.
  

Building on the success of our previous workshops held in Yorktown and London earlier this year, this workshop will provide a dynamic platform for knowledge exchange, collaborative discussions, and future vision-setting for AI in electric grid systems.

Featured Workshop Highlights:

• Keynote Presentations: Hear from thought leaders in AI and electric grid research.
• Technical Sessions: Covering data, model development, computation, and real-world applications.
• Panel Discussions: Insights from national laboratories, system operators, utilities, and industry stakeholders.
• Networking Opportunities: Engage with peers to foster new collaborations.

Who Should Attend:

• Researchers and Scientists: Focused on AI, ML, and electric grid systems.
• Industry Professionals: Exploring foundation models for grid optimization.
• Policymakers and Government Representatives: Shaping the future of electric grid technologies.

Organizing Committee:

• Argonne National Laboratory:  Kibaek Kim, Adrian Maldonado, Emil Constantinescu
• IBM Research:  Hendrik Hamann Jonas Weiss
• Hydro-Quebec:  Francois Miralles
• NRECA: Venkat Banunarayanan

Registration and Contact Information:

For questions, please contact the organizing committee at [email protected]. 

Registration Process: Registration involves two steps:

Step 1: Click the "Register Now" button at the bottom of the page to register for the workshop.

Step 2: Complete an Argonne visitor registration form. This process can take up to two weeks for approval.

• Foreign Nationals: Please provide all necessary documents listed in the registration form. Argonne administrative staff will contact you for further information if needed.

REGISTRATION DEADLINE: 

• Foreign Nationals Registration Closed: Monday, January 27, 2025
  
• US Citizens: Monday, February 3, 2025

 

Tue, February 11

8:30 AM → 9:00 AM Welcome and Opening Remarks

Speaker: Kibaek Kim (Argonne National Laboratory)

9:00 AM → 9:30 AM Vision and Progress Towards GridFM

Foundation models (FMs), pre-trained on large datasets and adapted to a broad set of applications, are revolutionizing the field of artificial intelligence (AI). Powerful FMs for language and weather have recently emerged, proving that such models can understand complex physical systems. In this presentation we review a moonshot concept, which are FMs for the electric grid (GridFMs). In an initial implementation, we introduce a GridFM pretrained on load flow data. We show good reconstruction performances and demonstrate that a GridFM pretrained on various grid topologies generalizes well to new grids. Finally, we discuss downstream applications, future development, roadmaps and opportunities of GridFM.

Speakers: Hendrik Hamann (IBM), Thomas Brunschwiler (IBM Research)

9:30 AM → 11:30 AM Grid Use Cases and Needs: Presentations

Convener: Venkat Banunarayanan (National Rural Electric Cooperative Association)

9:30 AM Adoption of AI for power systems at Hydro-Québec

Under the pressure of the energy transition and the digital transformation, there has been an intensification of global research efforts on AI for power systems in recent years. This led to the creation of a new practice and solutions applicable to a wide set of problems for the generation, transmission and distribution of electricity to customers. This presentation will give an overview of some of the current and future applications of artificial intelligence for power systems at Hydro-Québec to address the challenges faced by the company. While AI for specific tasks, like electric load forecasting for example, reached a level of maturity high enough to be used in our everyday operations, recent progress in generative modeling has the potential to widen applications of AI. Yet, Hydro-Québec’s network has some particularities due to Québec’s geography: most of the production is located in the North of the province, most of the load is located in the South and the network is not synchronized with its neighbours. This is a good test case for the generalization abilities of foundation models for power grids. We will show how Hydro-Québec Research Center structured its internal research initiative, called Voltaire, to participate in the development and test this new generation of generative AI algorithms.

Speaker: Sylvain Nadeau (Hydro-Quebec)

9:50 AM Digital Twins Enabling the Transition into the Grid of Tomorrow

The grid of today is one that is complex, unpredictable, and dynamic. These traits will be exponentially exacerbated, and new challenges will arise for the grid of tomorrow. The processes and technologies internal to electric utilities must adapt and evolve to prepare for these challenges and even enable utilities to improve their KPIs compared to today. This presentation will highlight Southern California Edison's vision and progress relating to digital twins. The presenter will define digital twins, the proposed architecture and implementation considerations, existing digital twin efforts within the company, and potential collaboration opportunities with industry to achieve this vision.

Speaker: Abder Elandaloussi (Southern California Edison)

10:10 AM Leveraging Digital Twins for Managing Grid Complexity

As the electric grid evolves, managing its complexity becomes increasingly challenging. Digital twins offer a powerful solution by providing a virtual, real-time model of the grid, enabling the simulation, monitoring, and optimization of the system. This presentation will explore the transformative value of digital twins in today's dynamic grid environment, with a particular focus on cyber-physical security and other critical needs.

Digital twins enhance grid management by improving cyber-physical security, optimizing operations, and facilitating predictive maintenance. Despite their benefits, developing and maintaining digital twins presents several challenges, including the tedious task of building and updating models and integrating diverse data sources.

To address these challenges, AI-based model development and maintenance offer promising solutions. This presentation will delve into the benefits and challenges of digital twins, highlighting Southern Company's work in this area. Attendees will gain insights into how digital twins can revolutionize grid management and the potential of AI to overcome existing challenges, paving the way for a more resilient and efficient electric grid.

Speaker: Clifton Black (Southern Company)

10:30 AM Grid AI in Practice: Balancing the Roles of General and Narrow Models for Grid

We will share our journey and evolution of constructing pragmatic Machine Learning (ML) and Artificial Intelligence (AI) models for ComEd’s grid. The focus will be on how this process necessitated the establishment of both general and narrow AI foundational models. We learned the most effective approach is to concentrate on impactful problems which then guides the selection and development of the most suitable tools. Despite the stringent industrial constraints that demand practical applications at each stage, our vision was guided towards both Narrow AI and General multimodal AI. This journey has led us to unanticipated theoretical depths. For instance, we have recognized the need for a higher dimensional graph topological framework. This framework, equipped with a multimodal dynamic federated learning, contributes to the creation of digital twins. These digital twins are key to unlocking solutions for critical practical grid applications, such as next-generation conditional maintenance and investment optimizations. We will explore the balance between practicality and theoretical depth in the realm of Grid AI, and how this balance can lead to innovative and efficient solutions.

Speaker: Norayr Matevosyan (ComEd)

10:50 AM From research to utility operations – Realizing value in the Era of Foundation Models and Generative AI

AI has been successfully used within the energy industry for decades however for most utilities, scaling beyond a limited number of use cases remains elusive. Foundation models and Generative AI holds the potential to break through many of the challenges holding us back but also introduces new challenges that must be considered. As we advance research with initiatives like GridFM, and move AI closer to grid operations, we must consider how to leverage this technology with the trust, reliability, and security necessary for critical infrastructure. Collaboration between industry participants on the utilization of AI insights in operations will be fundamental to moving beyond research, allowing us to leverage this technology to meet the emerging needs of our industry.

Speaker: Bryan Sacks (IBM)

11:10 AM The Grid Is Not An Island

NextEra Analytics (NEA) serves the R&D function for NextEra Energy which build and operates renewable energy plants across US and Canada. The design of the solutions for our customers requires NEA to simulate processes and optimize decisions across the various interacting systems of the electrical grid, electricity and gas markets, regulatory and government, customer clean energy goals and constraints, transportation network, water sources, land use etc. I will demo the software platform that NEA has developed that brings together information about these systems to serve the various other detailed modeling tools we develop.

Speaker: Sriharsha Veeramachaneni (NextEra Analytics)

11:30 AM → 12:00 PM Grid Use Cases and Needs: Panel Discussion

Convener: Venkat Banunarayanan (National Rural Electric Cooperative Association)

12:00 PM → 1:00 PM Net-Working Lunch: TBD

Convener: Kibaek Kim (MCS)

12:00 PM Keeping Data Private While Learning Together: Advances in Privacy-Preserving Federated Learning for the Grid

Speaker: Kibaek Kim (MCS)

1:00 PM → 2:30 PM Data Challenges: Presentation

Convener: François Miralles (Hydro-Quebec)

1:00 PM Leveraging the Sparse Data for Large Power Transmission Blackouts with Generative Statistical Models

Large blackouts are important because they have a high impact on our society, and, although rare, are not rare enough to be low risk. Indeed, the heavy tails in their statistics make them high risk. The observed data for large blackouts is limited, which poses challenges to analysis and prevents the training of AI models. However, large blackouts combine different processes at multiple time and spatial scales, and we can learn this anatomy and extract from observed utility data statistical descriptions of the initiation and propagation of outages in the grid. This leads to sampling from generative models that can produce typical patterns of outages. We illustrate such a generative model for the fast propagation of protection system outages caused by protection system misoperations. The generative model can generate synthetic samples with statistical properties similar to the original data. One objective for the new generative models is to leverage the limited observed outage data in order to generate synthetic data samples that can be used in training AI models that account for extreme events in the power grid.

Speaker: Ian Dobson (Iowa State University)

1:20 PM Creating Realistic Synthetic Electric Grids to Promote Open Science in Power Engineering

Before the advent of synthetic electric grids, public test cases for electric transmission grids were limited to the IEEE test cases and similar datasets. While these have served the community well, they do not match the size, complexity, or structure of today’s bulk electric grids. Industry grid models, however, are not publicly sharable because of critical energy infrastructure information (CEII) designation and similar restrictions. To address these challenges, over the last few years new methodologies have been developed to create synthetic (fictitious) electric grid models that better match the size, complexity, and structure of actual grids, while being free of CEII. This presentation discusses some of the latest research efforts in building synthetic grids and introduces some of the most recent public datasets available for large-scale electric grid simulation research, including those used as evaluation cases for ARPA-E’s Grid Optimization (GO) Competition.

Speaker: Adam Birchfield (Texas A&M University)

1:40 PM Power System Decarbonization Modeling Using the Argonne Low-carbon Electricity Analysis Framework (A-LEAF)

The Argonne Low-carbon Electricity Analysis Framework (A-LEAF) is a comprehensive power system simulation platform. A-LEAF integrates advanced tools for long-term generation and transmission expansion planning, production cost simulation, and probabilistic reliability assessment. This presentation will introduce the key capabilities of A-LEAF including the climate-informed decision-making, flexible power system representation, and temperature-dependent generation outages sampling along with results from past analyses.

Speaker: Jonghwan Kwon (Argonne National Laboratory)

2:00 PM Towards the development of a Digital Twin of the European Energy Infrastructure

Digital Twin is emerging as a key technology supporting and improving grid planning and operation. While different interpretation for the definition of a digital twin can be found in literature, there is an urgent need to come to common understanding and interoperable implementations. The presentation focuses on the current status and ongoing development in the European context. While it is not thinkable to have a single solution for all the grid operators in Europe, it is important to define common elements making the different implementations interoperable. The TwinEU project, with the involvement of more than 70 partners, is working towards the definition of a common architecture and interoperable solutions for data exchange. The element of data exchange is particularly relevant and critical. In this context the adoption of a data space approach will be presented as a key enabler to bring models from different operators together.

Speaker: Antonello Monti (Fraunhofer FIT and RWTH Aachen University)

2:30 PM → 3:00 PM Data Challenges: Panel Discussion

Convener: François Miralles (Hydro-Quebec)

3:00 PM → 3:15 PM Group Photo

3:15 PM → 4:45 PM Breakout Discussion: Data Challenges

Convener: François Miralles (Hydro-Quebec)

3:15 PM → 4:45 PM Breakout Discussion: Grid Use Cases and Needs

Convener: Venkat Banunarayanan (National Rural Electric Cooperative Association)

4:45 PM → 5:00 PM Adjourn

6:00 PM → 9:00 PM Group Dinner (optional)

Wed, February 12

8:30 AM → 9:00 AM Research Highlights from Mathematics and Computer Science Division

Speaker: Valerie Taylor (MCS)

9:00 AM → 9:30 AM Keynote

9:00 AM TBD

Speaker: Francis Alexander (Argonne National Laboratory)

9:30 AM → 9:40 AM Break

9:40 AM → 11:30 AM Computations in AI/Grid: Presentations

Convener: Adrian Maldonado (MCS)

9:40 AM AuroraGPT

AuroraGPT is a foundation model for science, trained on the Aurora supercomputer at the Argonne Leadership Computing Facility (ALCF). This model is designed to process and generate human-like scientific text, with the goal of supporting scientific discovery. Trained on a large corpus of scientific literature, this model has the potential to assist researchers in exploring, summarizing, and generating new hypotheses, making it a useful tool for various scientific domains.

Speaker: Sam Foreman (Argonne National Laboratory)

10:00 AM Globus: Enabling an Interoperable Ecosystem for AI in Science

Globus is a hybrid cloud platform designed to support advanced machine learning (ML) and artificial intelligence (AI) development, deployment, and research across distributed cyberinfrastructure (CI). With Globus services, users can manage remote computation, data management, indexing, search, and automated workflows without regard for the distributed resources used, from scientific instruments and sensors through to supercomputers and cloud nodes. By using Globus, researchers can significantly reduce the complexities associated with managing remote and distributed research processes. In this presentation, I will demonstrate how Globus serves as a platform for building cutting-edge AI services, such as federated learning, model training, and distributed inference.

Speaker: Kyle Chard (University of Chicago)

10:20 AM Edge to Power (E2P): Enhancing Energy Systems through Edge and AI Innovations

With the rapid advancement of digitization, energy systems have experienced significant improvements through innovations. As we move toward a more dynamic and complex network, real-time analysis, AI, and decentralized systems are essential for enhancing efficiency and reliability. Addressing this issue, many studies focus on centralizing data in a cloud system to develop innovative monitoring, control, and protection approaches for both conventional and inverter-based resource (IBR) systems.
The presentation discusses Hydro-Québec's new R&D project (E2P), which aims to develop post-cloud technologies by 2050. E2P focuses on enhancing integrated edge advancements and AI/data-driven applications.
The objective is to create a flexible, scalable system capable of swiftly incorporating new innovations within an automated design, testing, and deployment chain. This system would cover functions for plants, substations, feeders, and customer applications, addressing current limitations and eliminating silos to ensure a more resilient and efficient future energy system while addressing potential challenges.

Speaker: Vincent Choiniere (Hydro-Quebec)

10:40 AM Foundation Models for Time Series Prediction in Electric Power Grids

Recent successes with large language models (LLMs) for natural language processing have prompted exploration into their potential for time series prediction using numerical data. Chronos is a recent framework that pre-trains LLM models for predictions on time series data from various domains. The authors show that the framework can lead to successful zero-shot predictions for unseen time series data in several scenarios. In our research, we investigate the application of the Chronos models to analyze the frequency information from electric power grids to assess the effectiveness of time series predictions using this LLM-based framework. Based on our initial results, we fine-tuned the model with specific time series data from power systems to enhance its predictive accuracy.

Speaker: Russell Bent (Los Alamos National Laboratory)

11:00 AM Scalable Multi-Period AC Optimal Power Flow Utilizing Gpus

There has been a growing interest in solving multi-period AC OPF problems, as the increasingly fluctuating electricity market requires operation planning over multiple periods. These problems, formerly deemed intractable, are now becoming technologically feasible to solve thanks to the advent of high-memory GPU hardware and accelerated NLP tools. This study evaluates the capability of the ExaModels.jl and MadNLP.jl tools for GPU-centered nonlinear programming to tackle previously unsolvable multi-period AC OPF instances. Our numerical experiments, run on an NVIDIA GH200, demonstrate that we can solve a multi-period OPF instance with more than 10 million variables up to 10−4 precision in less than 10 minutes.

Speaker: Mihai Anitescu (Argonne National Laboratory)

11:30 AM → 12:00 PM Computations in AI/Grid: Panel Discussion

Convener: Adrian Maldonado (MCS)

12:00 PM → 1:00 PM Net-Working Lunch: TBD

Convener: Jonas Weiss (IBM)

12:00 PM We Built GridFM 0.5 – From Concepts to Our First Working Foundation Model

Speaker: Jonas Weiss (IBM)

1:00 PM → 2:40 PM AI Models and Architecture: Presentations

Convener: Emil Constantinescu (MCS)

1:00 PM Foundation models for weather and climate

Seen from a distance, the problem of foundation models for power systems has some similarity with that of foundation models for atmospheric physics. In both cases we are dealing with a system whose fundamental laws are known and simulation is possible, yet computationally expensive. Moreover, in both cases one frequently deals with incomplete information about the current state of the system. Given the ongoing rapid developments in the field of FMs for weather and climate, this talk will detail these developments with a view towards what one can learn for the case of power systems.

Speaker: Johannes Schmude (IBM)

1:20 PM Grid and Energy Sector Resiliency to Hydrological and Meteorological Extreme Events: Foundational Models for Resiliency with AI Models

Developing a general-purpose AI model for infrastructure resiliency assessments requires integration across a variety disciplines sectors and data heterogeneity and inference for a range of needs. When we talk about resilience to hydrological and meteorological hazards, hazard forecasting at time scales extending from few minutes to 100’s of years is necessary for addressing resiliency of grid / infrastructure for the immediate threat to develop and implement plans for low frequency and high impact events. These forecasts also need to be at a scale that is relevant to the infrastructure. We are making progress in building large Foundation model for medium range weather forecasting (1-10 days) with increasingly higher spatial resolution. Extending these models to seasonal-to-sub seasonal, interannual, decadal and climate scales is a challenge that is being addressed now. Integrating these hazard models with infrastructure resiliency models (i.e. fragility curves), socioeconomic models to estimate vulnerability of the infrastructure and exposure of the population to impacts for estimating risk and developing strategies for mitigating that risk is the goal of resiliency models. Initial attempts at integrating AI models developed across sectors is an active area of research. One such model our group is working is building a coupling between AI weather models and individual based decision models at Urban scale. The possibilities for integrating infrastructure models and scaling these types of models to regional and national scale is an emerging opportunity.

Speaker: Rao Kotamarthi (Argonne National Laboratory)

1:40 PM The Interplay between AI and Power Systems

This presentation examines two primary dimensions of the evolving interplay between AI and power systems. First, AI’s increasing electricity demand poses both short-run operational and long-run planning challenges, making it necessary to reform and modernize the existing electric grid. Second, new AI capabilities present an unprecedented opportunity to enhance the efficiency, reliability, and adaptability of power systems, from real-time grid control to long-term expansion planning.

Speaker: Le Xie (Harvard University)

2:00 PM Physics-informed Learning and Analytics for Power System Risk Assessment and Control with High Penetration of Inverter-based Resources

The increasing penetration of stochastic and uncertain inverter-based resources (IBRs), such as wind and solar PVs, has a considerable influence on the power system dynamics, operation, and optimization, causing reliability and resiliency concerns. On the other hand, the power industry is transforming itself from a hierarchical, passive, and sparsely sensed engineering system into a flat, active, and ubiquitously sensed cyber physical system. The emerging multi-scale data from phasor measurement units, SCADA, smart meters, weather, and electricity markets offers tremendous opportunities and challenges for the industry to dynamically learn and adaptively control the smart grid. This talk will present a new physics-informed learning framework to fully unlock the potentials from data while respecting physical constraints, including physics-informed estimation, inference and learning frameworks for power system security assessment, control, and uncertainty quantification in presence of different types of uncertain resources. Various transmission and distribution system applications will be presented to highlight the benefits of this framework.

Speaker: Junbo Zhao (University of Connecticut)

2:20 PM Building smarter grids with AI: Insights from EU projects

This talk will showcase novel concepts, algorithms, and applications primarily developed within European projects that leverage AI-based systems in smart grids. It will introduce three key concepts: the interdisciplinary framework for AI-driven decision systems in critical infrastructures, developed in the AI4REALNET project (https://ai4realnet.eu); the data-driven evolving models for managing power systems with high renewable energy integration, under development in the ENFIELD project (https://www.enfield-project.eu); and the Testing and Experimentation Facility (TEF) concept from the AI-EFFECT project (https://ai-effect.eu), demonstrated through a use case on reconstructing low-voltage grid topology and electrical parameters. The presentation will also highlight additional industry-relevant use cases with significant potential for foundation model applications, such as smart alarm management.

Speaker: Ricardo Bessa (INESC TEC)

2:40 PM → 3:00 PM AI Models and Architecture: Panel Discussion

Convener: Emil Constantinescu (MCS)

3:00 PM → 3:30 PM Tour: Argonne Leadership Computing Facility

Discover the cutting-edge of scientific innovation at the Argonne Leadership Computing Facility (ALCF), a U.S. Department of Energy Office of Science user facility. Home to Aurora, the world’s fastest AI supercomputer capable of performing over one quintillion calculations per second, the ALCF enables revolutionary breakthroughs in fields ranging from climate science and materials research to quantum information and nuclear engineering. Explore how this next-generation exascale supercomputer, built in partnership with Intel and Hewlett Packard Enterprise, powers U.S. leadership in supercomputing and transforms science at an unprecedented scale.

3:30 PM → 5:00 PM Breakout Discussion: AI Models and Architecture

Convener: Emil Constantinescu (MCS)

3:30 PM → 5:00 PM Breakout Discussion: Computations in AI/Grid

Convener: Adrian Maldonado (MCS)

5:00 PM → 5:15 PM Adjourn

Thu, February 13

8:30 AM → 9:00 AM Research Highlights from Energy Systems and Infrastructure Analysis Division

Speaker: Henry Huang (Argonne National Laboratory)

9:00 AM → 9:45 AM Special Discussion on Partnerships

Conveners: Kibaek Kim (Argonne National Laboratory), Hendrik Hamann (IBM)

9:00 AM Panelists

Speakers: Alexandre Parisot (Linux Foundation Energy), Ed Kovler (PJM Interconnection), Jack White (EPRI), Saba Khalid (ARPA-E)

9:45 AM → 10:00 AM Break

10:00 AM → 11:45 AM Roadmap to Collaboration: Building GridFM Together

This interactive session will bring the workshop to a productive close by collecting insights and planning the path forward. Breakout leads will present their proposed roadmaps to address challenges and opportunities in building foundation models for the electric grid. A focused presentation will outline governance strategies to ensure sustained collaboration. The session will conclude with a collaborative action planning discussion, where participants will identify immediate next steps, establish or refine working groups, and set timelines for follow-ups, ensuring momentum continues beyond the workshop.

Speakers: Kibaek Kim (MCS), Thomas Brunschwiler (IBM Research)

11:45 AM → 12:00 PM Adjourn