Deep learning-based situation detection, comprehension, and projection for power systems under false data injection attacks

With the continuous expansion and increasing intelligence of power systems, numerous state estimation methods based on deep learning have been proposed. However, most studies have primarily focused on estimating a limited set of state variables, lacking a deeper comprehension of system operating conditions. Additionally, models trained on specific datasets are unable to adapt their parameters and structures once deployed in real operations, rendering them ineffective in responding to unexpected events. This limitation becomes especially apparent when faced with False Data Injection Attacks (FDIAs), as the compromised measurements notably diminish the accuracy of model predictions. To address these, this paper presents a real-time detection, comprehension, projection model for power systems, leveraging the concept of situation awareness. The proposed model combines the Chebyshev network with the Long Short-Term Memory (LSTM) network. In the situation detection phase, malicious data within the measurements is identified and localized. An adaptive preprocessing layer is designed to reduce the weight of data from compromised nodes, mitigating the impact of FDIA. Subsequently, in the situation comprehension and situation projection phases, the current and future system state variables and risk situations are predicted, providing an in-depth comprehension and anticipation of system operational status. Experimental results demonstrate that the proposed model exhibits high robustness and accuracy when tested on IEEE 39 bus and IEEE 118-bus systems, thereby enhancing the resilience of deep learning models against FDIAs during real-time operations.