Correlation-Exploiting False Data Injection Attack and Corresponding DRL-Based Detection Countermeasure in Partially Observable Grids

In partially observable power grids, limited knowledge of grid topology and correlated nodal injections presents significant challenges to cybersecurity, complicating the development of stealthy false data injection attacks (FDIAs). To address these issues, this paper proposes an optimization-based and correlation-aware FDIA that remains stealthy under AC state estimation and bypasses bad data detection (BDD) by explicitly incorporating spatio-temporal dependencies among nodal injections. The attack strategy employs a reduced-dimensional measurement matrix and an error covariance matrix to model injection correlations, thereby minimizing operational constraint violations. Correspondingly, a coupled Cubature Kalman Filter approach is characterized to enhance state estimation accuracy and reduce conservatism. To counter such evolving attacks, a deep reinforcement learning (DRL)-based detection mechanism is developed. Extensive validation on IEEE 14, 30, 57, and 118-bus systems demonstrates the effectiveness and robustness with 99.92%-100% accuracy of the proposed method.