Spatio-temporal dynamic graph neural network-based missing measurement recovery method for power system state estimation

High-quality measurement data are essential for reliable power system state estimation, yet SCADA and field monitoring devices frequently experience both random and continuous missingness due to sensor degradation, communication failures, and device outages. Although notable progress has been made in missing data recovery, many existing approaches remain tailored to low missing rates, underutilize node–edge electrical coupling, and rely on single-scale temporal modeling, which limits their robustness under the hybrid and high missingness conditions commonly observed in real grids. To address these limitations, this study proposes STGNI, a spatio-temporal graph neural network designed for heterogeneous missingness without predefined assumptions. The model integrates: (1) a missing-aware representation module that suppresses unreliable inputs; (2) a topology-informed dual-channel spatial module that captures coupled node–edge dependencies; and (3) a cross-scale temporal mechanism that jointly models short-term variations and long-range periodicity. Comprehensive experiments on the NREL-118 system and real provincial grid data show that STGNI consistently outperforms state-of-the-art baselines across 10–90% missing ratios, achieving substantial reductions in MAE and RMSE. The reconstructed measurements further enhance multiple state estimation algorithms, demonstrating the practical value of STGNI for improving data reliability and strengthening grid monitoring resilience. • Hybrid-missingness-aware STGNN models random and continuous missingness in power systems via joint node–edge–mask learning. • Mask-enhanced dual-channel attention and dual-branch spatial framework strengthening node–edge collaboration under severe missingness. • Bidirectional ODEs + Informer for cross-scale modeling of short-term fluctuations and long-range periodicity in real grids.