Graph Neural Networks for Resilient State Estimation Under Degraded Grid Monitoring
Grid operators are increasingly flying blind: PMUs, SCADA units, and smart meters malfunction or drop offline without warning, and classical state estimators silently degrade — or fail outright — once observability is lost. This internship tackles both sides of that problem with modern deep learning: learning where to place sensors so the grid stays observable under plausible failure scenarios, and learning to estimate state robustly when some of those sensors go dark or start lying. Rather than treating placement and estimation as separate combinatorial and least-squares problems, you will cast the grid as a graph and use graph neural networks (GNNs) to solve them jointly and adaptively, with uncertainty quantification giving operators a real-time confidence signal instead of a silent failure.
What you’ll work on:
Grid-as-graph formulation & literature review (Month 1). Formalize the network topology, existing sensor layout, and historical measurement streams (including fault and communication-loss events) as attributed graphs; review recent GNN-based state estimation, physics-informed neural networks, and deep reinforcement learning approaches to sensor/actuator placement.
Learned sensor placement under failure scenarios (Months 2-3). Design a GNN encoder that scores candidate sensor locations for observability, combined with a reinforcement-learning or combinatorial-optimization policy that selects the minimum-cost sensor set guaranteeing observability across a defined library of failure scenarios (sensor dropout, communication loss, correlated outages).
Robust, uncertainty-aware state estimation (Months 3-5). Train a physics-informed GNN state estimator that ingests incomplete and noisy measurements, using self-supervised or generative imputation (e.g. masked autoencoders, diffusion-based imputation) to reconstruct missing readings, and conformal prediction or Bayesian deep learning to produce calibrated, bounded uncertainty on each state estimate.
Anomaly detection & stress-testing (Month 5-6). Add a detection head that flags malfunctioning or spoofed sensors from residual and uncertainty patterns; stress-test the full pipeline (placement + estimator + detector) on historical and synthetically degraded scenarios, and benchmark against classical WLS/Kalman-filter baselines.
Deliverables:
- An optimal sensor placement strategy that minimizes deployment cost while guaranteeing observability under a defined set of failure scenarios.
- A robust, uncertainty-aware GNN state estimator with built-in sensor-malfunction detection, benchmarked against classical estimators, packaged as a reproducible prototype.
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 or PyTorch Geometric/DGL preferred)
- Interest in graph neural networks, uncertainty quantification, or power systems; prior exposure to state estimation or optimization 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.