A Multimodal Foundation Model for Power Grid Operations
Grid operators drown in heterogeneous signals — SCADA streams, alarms, switching records, outage histories, weather feeds, operator notes — yet have no easy way to test a control action or a flexibility strategy before deploying it on the real network, nor to quickly get a plain-language read on “what’s happening and why.” This internship builds a multimodal foundation model for power-grid operations: a single model that learns the grid’s behavior from heterogeneous operational data, can simulate future scenarios and intervention outcomes, and doubles as a natural-language grid intelligence assistant that explains grid state and advises on cost-optimal control actions.
What you’ll work on:
Data unification & literature review (Month 1). Unify heterogeneous operational data (topology/asset models, SCADA and event streams, alarm logs, switching records, outage histories, weather) into a common multimodal representation; review recent foundation and world-model architectures for time series, graphs, and text.
Multimodal grid-state encoder (Months 2-3). Design and pretrain an encoder that fuses time-series measurements, event/alarm logs, and network topology into a shared latent “grid state” representation, using self-supervised objectives (masked reconstruction, cross-modal contrastive alignment).
Scenario simulation & intervention modeling (Months 3-4). Extend the encoder into a generative world model that rolls forward future grid states conditioned on candidate control actions or flexibility strategies, enabling replay of past incidents and “what-if” testing of interventions before deployment.
Grid intelligence assistant (Months 4-5). Ground a language model in the learned grid-state representation to generate natural-language explanations, summaries, and incident reports, and to answer operator queries about current or past grid state.
Control advisory & evaluation (Months 5-6). Couple the world model with an optimization or reinforcement-learning layer to propose cost-optimal control actions under operational constraints; benchmark scenario fidelity, explanation quality, and advisory recommendations against historical incidents and operator ground truth.
Deliverables:
- A multimodal foundation/world model for power grid operations, trained on heterogeneous SCADA, event, topology, and weather data, supporting scenario simulation and safe replay of interventions.
- A grid intelligence assistant prototype delivering natural-language explanations and reports of grid state, together with cost-aware control advisory recommendations.
Profile sought:
- Enrolled in a Master’s program in Computer Science, Electrical Engineering, or a related field
- Strong programming skills in Python, with hands-on experience in deep learning frameworks (PyTorch preferred)
- Interest in foundation models, multimodal learning, sequence models/transformers, or graph neural networks; prior exposure to LLMs or power systems is a plus
- Autonomous, rigorous, and comfortable working at the interface of machine learning and power systems engineering
How to apply: Send your CV and your Master transcripts by email.