Physics-Informed Graph Neural Operator: Dynamic State Estimation Under Sparse PMU Measurements

Physics-informed graph neural operator estimates global generator dynamic states from sparse PMU measurements in one forward pass, with high accuracy and low cost.
The paper proposes a physics-informed graph neural operator for power system dynamic state estimation under sparse PMUs. A graph attention encoder maps measurements to node features, and a neural operator decodes all generators' rotor angles and speeds. The swing equation is injected as a hard or soft constraint with learned correction, and training uses only measurement reconstruction error. On four transmission systems, the operator is the most accurate among estimators without scenario optimization or state labels, beats hard-constrained MHE on the 118-bus system, and matches or exceeds optimization baselines at far lower cost under model errors. Limitations include the tuned Kalman smoother's edge on the smallest system under large errors and unvalidated real-world PMU noise and latency.
Power system dynamic state estimation (DSE) is a critical component of secure grid operation, but traditional methods rely heavily on measurement density and often struggle under sparse phasor measurement unit (PMU) configurations. A recent paper proposes a Physics-Informed Graph Neural Operator that aims to estimate the rotor angle and speed trajectories of all generators across the grid—including those not directly measured—in a single forward pass using only intermittent measurements at a small number of generator buses.

Problem Background: The Estimation Challenge Under Sparse Measurements
Dynamic state estimation requires real-time knowledge of key state variables such as generator rotor angles and speeds, but the installation cost of PMUs limits their deployment density in large power grids. Traditional Kalman filter-based methods degrade sharply when measurements are sparse or the model contains errors, while optimization-based moving horizon estimation (MHE) achieves higher accuracy but requires iterative solving for each scenario, incurring a heavy computational burden. The paper seeks to answer: can a trained neural operator replace per-scenario optimization to achieve high-accuracy, low-cost dynamic state estimation under sparse measurements?
Moving Horizon Estimation (MHE) is a rolling-optimization-based state estimation method: at each time step, it uses measurement data over a fixed-length past time window and solves a constrained optimization problem to estimate the current system state. Compared with Kalman filtering, MHE can explicitly handle state constraints (such as generator output limits and the swing equation) and nonlinear models, and therefore achieves higher accuracy in power system dynamic state estimation. The cost, however, is that an optimization problem must be solved iteratively at every time step, with computation growing with system size and horizon length, making real-time requirements difficult to meet. The paper attempts to replace this per-scenario optimization with a trained neural operator, thereby greatly reducing online computational cost while maintaining accuracy.
Method Design: Graph Attention Encoding and Swing Equation Constraints
The core of the method is a graph attention encoder that maps PMU measurements to graph node features, followed by a neural operator that decodes the trajectories of all generators. Physical knowledge is injected in two ways: first, the swing equation is imposed as a hard constraint, forcing the output to satisfy power system dynamics; second, it is used as a soft constraint with a learned correction term to absorb model errors. This design allows the network to be trained without state labels—the loss function is built directly on measurement reconstruction error, requiring no ground-truth states as supervision.
Training data comes from simulated measurements, using a masking training strategy: when multiple PMUs are present, the input of one is randomly masked while its contribution to the loss is retained, forcing the network to learn to infer global states from partial measurements. This self-supervised training avoids dependence on state labels and enhances the model's robustness to missing measurements.
The swing equation is the core dynamical equation in power system transient stability analysis, describing how synchronous generator rotor angles and speeds change over time. Its standard form is: ( M_i \frac{d\omega_i}{dt} = P_{m,i} - P_{e,i} - D_i(\omega_i - \omega_{ref}) ), ( \frac{d\delta_i}{dt} = \omega_i - \omega_{ref} ), where ( \delta_i ) is the rotor angle, ( \omega_i ) is the speed, ( M_i ) is the inertia constant, ( D_i ) is the damping coefficient, and ( P_{m,i} ) and ( P_{e,i} ) are mechanical power and electromagnetic power, respectively. This equation couples generator dynamics with network power flow and is the physical foundation of dynamic state estimation. The paper embeds the swing equation as a hard constraint in the network output, meaning the estimated trajectories must strictly satisfy this dynamical relationship; as a soft constraint, a certain deviation is allowed, and a learned correction term absorbs model errors (such as inaccurate parameters or unmodeled dynamics). This flexible combination of hard and soft constraints enables the neural operator to balance physical consistency against robustness to model errors.
Graph Attention Network (GAT) is a neural network that operates on graph-structured data, using an attention mechanism to assign different weights to each node's neighbors and thereby aggregate neighborhood information. In power systems, buses or generators can be treated as graph nodes and transmission lines as edges; GAT can adaptively learn which neighbors are more important for estimating the current node's state. In the paper, the graph attention encoder maps PMU measurements (such as voltage and current phasors) to node features and then, through multi-layer attention aggregation, captures the grid's topology and local measurement patterns. Neural operators are a class of models that learn mappings between function spaces, aiming to approximate an operator (such as the mapping from measurement functions to state trajectory functions) rather than merely fitting finite-dimensional vectors. This gives neural operators resolution independence, enabling them to handle grids of different sizes and topologies and to generalize to generators not directly measured under sparse measurements. Together, the two form the core architecture of the physics-informed graph neural operator.
Experimental Validation: Comparison Across Four Transmission Systems
The paper conducts simulation tests on four transmission systems and compares against multiple baseline methods. Results show that among estimators that neither optimize per scenario nor use state labels, the neural operator achieves the highest accuracy. The only exception is the smallest system under large model errors, where a tuned Kalman smoother performs better. As a detail, on the 118-bus system, a network trained with simulated states actually underperforms the operator used directly—the latter is faster and more accurate than hard-constrained moving horizon estimation.
In the presence of model errors, the operator matches or exceeds the accuracy of optimization baselines at far lower cost. Moreover, the results still hold on a detailed generator model whose exact parameters are only indirectly visible through voltage phasors, indicating that the method adapts well to parameter uncertainty.
The Kalman smoother is an offline or post-processing version of the Kalman filter that uses not only measurements before the current time but also future measurements to estimate historical states, and thus is typically more accurate than the filter. In dynamic state estimation, the Kalman smoother requires known system models and noise statistics and computes state estimates through forward-backward recursion. When model errors are large, a parameter-tuned smoother can partially compensate for model inaccuracy, which is why it outperforms the neural operator—not optimized for that scenario—on the smallest system. The paper notes this exception, showing that traditional methods still have advantages under specific conditions and suggesting that the neural operator's performance depends on how well the training data match the test scenario. In addition, hard-constrained moving horizon estimation means that physical constraints such as the swing equation are strictly satisfied in the MHE optimization problem, whereas soft-constrained MHE allows constraint violations with penalties. The paper compares the neural operator with hard-constrained MHE on the 118-bus system; the latter has stronger physical consistency but high computational cost, while the neural operator achieves faster and more accurate estimation in a single forward pass.
Significance and Limitations
This work demonstrates the potential of physics-informed neural operators for power system state estimation: a single forward pass completes full-state estimation, avoiding per-scenario optimization and opening the door to real-time applications. The two injection modes for physical constraints also offer flexibility—hard constraints guarantee physical consistency, while soft constraints with learned corrections tolerate model errors.
The limitations are equally clear: on the smallest system with large model errors, the traditional tuned Kalman smoother still prevails; the impact of network size and training strategy on performance is not monotonic, and on the 118-bus system training with simulated states actually hurts accuracy, indicating that transferability and scaling require further study. In addition, the experiments are based on simulated data, and issues in real PMU measurements such as noise, latency, and communication outages have not yet been tested in real-world scenarios.
Conclusion
The physics-informed graph neural operator offers a new path for dynamic state estimation under sparse PMUs: encoding measurements with graph attention, constraining outputs with the swing equation, and training in a self-supervised manner, it demonstrates a balance of accuracy and efficiency across multiple transmission systems. Although traditional methods still have advantages under specific conditions and real-world deployment remains to be validated, this approach points to the broad potential of combining physical knowledge with neural operators for real-time power system analysis.
One-line Summary
A physics-informed graph neural operator estimates global generator dynamic states from sparse PMU measurements in a single forward pass, with high accuracy and low cost.
Paragraph Summary
The paper proposes a physics-informed graph neural operator for power system dynamic state estimation under sparse PMU configurations. The method uses a graph attention encoder to map measurements to node features, then decodes rotor angle and speed trajectories of all generators via a neural operator. The swing equation is injected as either a hard constraint or a soft constraint with a learned correction, and training relies only on measurement reconstruction error without state labels. Simulations on four transmission systems show that the operator is the most accurate among estimators that use neither scenario optimization nor state labels, outperforms hard-constrained moving horizon estimation on the 118-bus system, and matches or exceeds optimization baselines at far lower cost under model errors. Limitations include the tuned Kalman smoother's advantage on the smallest system under large errors and the lack of validation for real PMU noise and latency.
Related articles

AI Writes Your Code—Do You Still Need to Learn Git? Five Reasons
AI coding agents handle Git operations automatically—is it still necessary to learn Git? Five reasons: team communication, reviewing agent PRs, monitoring code changes, fallback when AI goes down, and the joy of learning—plus how the focus of learning Git is shifting in the agent era.

Anthropic Reportedly Prices IPO at $83 Per Share, Valuing Company at $1.5 Trillion
Anthropic reportedly prices IPO at $83 per share, above the expected range, with a fully diluted valuation of $1.5 trillion. This article analyzes the signal significance of the pricing and its impact on the AI industry's financing environment.

Open-Source Code Hosting Platform Codeberg Suffers Outage, Sparking Debate in Developer Community
Open-source code hosting platform Codeberg.org suffered an outage, sparking Hacker News debate over nonprofit platform stability and open-source infrastructure dependency risks.