Abstract: To address the coexistence of missing values and anomalous distortions in power grid operational data, as well as the limited efficiency of long-sequence processing, an anomaly detection and iterative reconstruction strategy based on an Iterative Masked State-Space Variational Autoencoder (IMSM-VAE) is proposed. First, Centrality Distribution Feature Mapping (CDFM) is employed to integrate static nodal topology attributes with dynamic measurement features. Second, a bidirectional selective state-space module is introduced to model long-range temporal dependencies with linear time complexity. Finally, latent-space Mahalanobis distance is used to iteratively update anomaly masks, and anomalous measurements are reconstructed using bidirectional temporal context.