PhD Position, Foundation models for the electric power grid of the future

The power system is changing, largely driven by the energy transition and climate change. The large shares of renewables, both centralized and distributed, are posing new challenges to system operations that are yet to be fully understood. In this context, it is evident that the operation, control, and planning of power systems will soon be pushed to their limits. Therefore, new computational methods and approaches are needed to better tackle the challenges posed by increased uncertainty and complexity. Machine learning (ML) and artificial intelligence (AI) methods have shown promise for such purposes in a wide spectrum of industries, with significant breakthroughs in computer vision, natural language processing, and intelligent control. This PhD project aims to develop foundation models (FMs) for the electric power grid. FMs are advanced AI models developed through self-supervised learning, most often based on transformer architectures, that generalize across various tasks after initial training on large datasets, enabling efficient adaptation to specific applications with minimal annotated data.

The objectives of this PhD research project are to:

1. Develop scalable foundation models for the electric power grid (GridFM).
2. Fine-tune the models for a variety of downstream tasks.
3. Assess its application potential in cooperation with industry.

This thesis aims to answer the following research questions:

R1: Which ML concept is the most promising in developing a GridFM? In particular, the thesis aims to investigate:

R1.1 Can physics-informed learning play a role in building GridFM?

R1.2 Can we exploit neural operators in building GridFM?

R1.3 Can we leverage Reinforcement Learning (RL) in building GridFM?

For the latter, the thesis will investigate whether current LLM (e.g., ChatGPT and DeepSeek) architectures, which rely on RL, can also be used to develop a robust GridFM architecture.

R2: How can the mixture of experts (MoE) paradigm be utilized in building GridFM? The current success of DeepSeek has demonstrated that the MoE Transformer architecture can significantly improve inference efficiency and model scalability. Developing and testing such an architecture will be crucial to GridFM's development.

R3: Which downstream tasks will benefit the most? The thesis will identify a set of related power system tasks for which a single MoE will be developed.

Additional research questions:

R4: What datasets and data structures will be of most benefit to GridFM? The theses will contribute to the gathering and generation of real and synthetic datasets, respectively. This data will aim to better reflect real-world conditions while ensuring applicability across various operational planning and optimization tasks.

This work will be performed using currently developed or in-development tools at RRE. Namely, the PowerGraph dataset will be initially used and further extended. The current modeling tool for grid analyses, Cascades and its surrogates, developed using graph neural networks (GNN), will serve as a starting point for future modeling developments. In addition, the candidate will leverage our experience and models developed using RL and physics-informed neural networks. This work will further benefit from our ongoing collaborations with IBM and other partners.

Job description

Profile

• Fully funded position with competitive salary according to ETH standards
• Access to state-of-the-art computational infrastructure
• Interdisciplinary and international research environment
• Perspectives for career development
• International team culture and team composition