Dynamic State Estimation from Sparse PMUs via Physics-Informed Graph Neural Operators
Authors: Huy Hoang Le, Haoguang Wang, Christian Moya, Marcos Netto, Guang Lin
Organizations: School of Mechanical Engineering, Purdue University, West Lafayette, IN 47907 USA · Department of Mathematics, Purdue University, West Lafayette, IN 47907 USA · Department of Electrical and Computer Engineering, New Jersey Institute of Technology, Newark, NJ 07102 USA
This paper proposes a physics-informed graph neural operator for power system dynamic state estimation. From a sparse, intermittent measurement window at a few generator buses, the operator estimates the rotor-angle and speed trajectories of all machines, unmeasured ones included, in one forward pass. A graph-attention encoder processes the measurements, and the swing equations are imposed as a hard constraint or as a soft one with a learned correction that absorbs model errors. It is trained on simulated measurements, without state labels. With several phasor measurement units, masked training hides one from the input but keeps it in the loss. On four transmission systems, simulations show that the operator is the most accurate estimator that neither optimizes per scenario nor uses state labels, except on the smallest system at large model errors, where tuned Kalman smoothers are better. Networks trained on simulated states are more accurate on the two smallest systems but not on the 118-bus system, where the operator is faster and more accurate than hard-constrained moving-horizon estimation. Under model errors, it matches or outperforms an optimization-based benchmark at a fraction of its cost. The results hold for a detailed machine model with exact parameters seen only through voltage phasors.
Figures & tables
Fig. 1: The problem on the WSCC 9-bus system: from the sparse window of its only PMU (machine 2), the graph neural operator (GNO) estimates the trajectories of all machines.
Fig. 2: The GNO of Section III (settings of Section V , window of Fig. 5 (a)): a GATv2 branch encodes the PMU window; the physics decoder integrates the swing equations. Dashed: GNO-soft only; dotted: training gradient.
Property
[ 20 ]
[ 22 ]
[ 23 ]
[ 33 ]
[ 35 ]
[ 12 ]
[ 16 ]
Ours
Estimates machine states
–
–
–
–
✓
✓
✓
✓
Nonlinear transients
✓
–
–
✓
–
✓
✓
✓
Machines without a sensor
–
–
–
–
✓
✓
✓
✓
No state labels
–
–
–
–
✓
✓
–
✓
No per-window optimization
–
✓
✓
✓
∘
–
✓
✓
Handles model error
∘
–
–
–
∘
∘
–
✓
TABLE I: Positioning against representative prior work (✓: yes; ∘ : partly, see text; –: no).
Fig. 3: How machine 3 is reached on the window of Fig. 5 (a): (a) Kron admittances ∣Yij∣ and GNO-hard’s attention into machine 3; (b) machine 2’s samples and bins ( 7 ), and both machines’ trajectories; (c) ∂δ^2(t)/∂x^3(0) through the hard unroll.
Fig. 4: Estimators by trust in the nominal model (columns) and computation per window (rows). SHRED, trained on state labels, lies outside the grid.
Method
RMSE of δ (deg) vs. model error ε
Inference
0%
5%
7%
10%
15%
20%
time (ms)
EKF
2.93
3.05
3.23
3.11
3.61
3.74
99
UKF
2.87
3.02
3.14
3.06
3.57
3.74
167
ERTS
0.39
0.64
0.81
0.95
1.37 ∗
1.72 ∗
97
GNO-hard (ours)
0.39
1.27
1.54
2.33
3.46
4.12
110
GNO-soft (ours)
0.39
0.59
0.68 ∗
0.94
1.64
2.23
123
TABLE II: WSCC 9-bus: angle RMSE (deg) at the unmeasured machine versus model error ( N=50 ; static: 6.3 to 7.5 ), and inference time.
Fig. 5: Two 9-bus windows at 7% : (a) median gap between hard-MHE and GNO-soft; (b) GNO-soft’s largest error. Grey band: truth; error axis clipped at ±10∘ ; legends: per-window RMSE.
Fig. 6: WSCC 9-bus. (a) Angle and (b) speed versus model error; (c) noise and (d) sparsity at 7% (dashed: training value). Bands: ±1 SEM. Dotted: GNO-soft trained with random sample loss (one seed).
Fig. 7: RMSE relative to each system’s static baseline (value at the dashed line) against single-CPU time per scenario. Dotted: prior mean, no measurements. Spans: hard-MHE to GNO-soft cost ratio.
Fig. 8: Angle RMSE versus model error on Kundur and the 39-bus system (9-bus: Fig. 6 (a)). UKF omitted (never below 5.8∘ ). Bands: ±1 SEM.
System
Method
0%
7%
15%
20%
ms
Kundur 4-mach 1 PMU
EKF
9.21
6.90
7.00
7.74
66
UKF
9.05
6.37
6.43
7.30
375
ERTS
6.95
5.13
5.55
6.13
78
Hard-MHE
0.72 ∗
1.67
3.22
3.42
868
GNO-hard
1.00
2.18
2.81
3.38
108
GNO-soft
0.95
1.46
2.27
2.91 ∗
120
TABLE III: Angle RMSE (deg) at the unmeasured machines versus model error ( N=30 ; static: 6.4 to 9.0 on Kundur, 5.8 to 7.5 on the 39-bus), and inference time. SHRED: supervised, untimed.
Fig. 9: Angle RMSE on every scenario of Table IV (log scale). Box: quartiles (dashed: hard coupling); bar: median; diamond: mean; whiskers: 5 th to 95 th percentile.
System
Method
mean
median
90th pct.
ms
Detailed 9-bus flux-decay + AVR 1 voltage PMU
EKF
2.54
2.16
3.74
1638
UKF
2.11
1.92
3.03
2752
ERTS
0.56
0.52
0.84
1386
Hard-MHE
0.60
0.54
0.98
2339
GNO-hard
0.52
0.43
0.97
540
GNO-soft
0.51
0.40
0.93
667
TABLE IV: Angle RMSE (deg) at the unmeasured machines and inference time: detailed 9-bus model ( N=60 , exact parameters, static 8.77 ), IEEE 118-bus system ( N=50 , static 4.44 ). Asterisks concern the mean.
Fig. 10: Amortization versus the learned prior. (a) 39-bus, exact model; GNO-soft retrained for each K . (b) Nonzero model errors pooled; rows: PMUs/non-reference machines; bars: share of the gap to GNO-soft the prior closes.
KIOS Center of Excellence, University of Cyprus · CEMSE, King Abdullah University of Science and Technology (KAUST) · Department of Electrical and Computer Engineering, University of Cyprus
Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany · Institute of Electrical Energy Systems, Friedrich-Alexander-Universität Erlangen-Nürnberg, Germany