SHRED: New AI Reconstructs Power Grid State with Just a Few PMUs
A shallow recurrent decoder that works without a physical model—even under nonlinear faults.
A team led by Andrea Pomarico (Politecnico di Milano) and J. Nathan Kutz (University of Washington) introduces SHRED (SHallow REcurrent Decoder), a machine learning architecture designed for dynamic state estimation in power systems. Traditional methods like Kalman filter variants suffer from high computational complexity, sensitivity to model inaccuracies, and degradation under nonlinear conditions. They also require optimal PMU placement—a costly constraint. SHRED sidesteps these problems by learning a direct mapping from sparse measurements to the full system state, without relying on a detailed physics model.
Validated on the IEEE 39-bus test system under short-circuit disturbances, SHRED consistently outperforms a state-of-the-art shallow decoder baseline when only a limited number of PMUs are available. The model maintains high reconstruction accuracy even under severe noise and nonlinear faults, making it a practical alternative for wide-area measurement systems. The results suggest that utilities can achieve reliable real-time grid monitoring without massive PMU deployment, potentially reducing infrastructure costs while enhancing situational awareness.
- SHRED uses a shallow recurrent decoder to reconstruct the full power system state from sparse PMU measurements.
- The method requires no accurate physical model and is insensitive to PMU placement—unlike Kalman filters.
- Validated on IEEE 39-bus system under short-circuit disturbances, it outperforms a benchmark shallow decoder in sparse-measurement scenarios.
Why It Matters
Enables accurate grid monitoring with fewer PMUs, cutting infrastructure costs and improving resilience during faults.