Abstract:

To address the data security issues faced by remote terminal units in smart substations under false data injection attacks（FDIAs），a bad data recovery strategy based on spatio-temporal multi-view learning was proposed. This approach was designed to enhance the accuracy of situational awareness and operational reliability of power systems，with a focus on overcoming the limitations of traditional methods that ignore spatio-temporal correlations，rely heavily on physical modeling，and lack robustness against coordinated attacks. A centralized FDIA model fitting the attack characteristics of substations was established，followed by the construction of a four-view learning framework integrating temporal dynamics，spatial topological correlations，feature interactions，and physical constraints. By leveraging an attention mechanism，adaptive fusion of multi-source features was achieved to precisely reconstruct the tampered data. Furthermore，multi-scenario simulation experiments were designed，covering critical nodes at different electrical locations and continuous multi-node attack scenarios. Simulation results indicated that the reconstruction errors of the proposed method for voltage and power measurements were significantly lower than those of traditional methods. Additionally，the proposed strategy demonstrated the capability to maintain optimal recovery accuracy and system-level physical consistency，even under extreme attack scenarios. The proposed strategy effectively improves the attack resistance and recovery reliability of substation measurement data. Consequently，it can provide critical technical support for resilient power grids.

Key words: smart substation, remote terminal unit, false data injection attack, data recovery, spatio-temporal multi-view learning