A multi-task learning framework for real-time integrated FDIA localization and state correction of cyber–physical power systems

The improved situational awareness capabilities of the power system have also exposed significant cybersecurity vulnerabilities, underscoring the critical need for developing effective approaches against attacks. However, existing methods fail to provide any effective information other than classification results during the detection stage, making it difficult to directly guide decision-making. In this paper, a computationally efficient false data injection attack (FDIA) model is formulated based on a modified AC power flow, which constructs stealthy attack vectors that cannot be recognized by any residual-based detection methods through refined physical constraints. Then, in order to diagnose the corruption on node states caused by data manipulation, a spatiotemporal graph encoder–decoder (ST-GED) is proposed based on a multi-task learning framework. The model performs real-time contamination detection and state correction simultaneously, which is completely immune to erroneous measurements. Specifically, in the spatiotemporal blocks of the encoder and decoder, spatial information is captured through an edge-aware message passing module that introduces a multi-head self-attention mechanism, and GRU is leveraged to learn temporal features. Numerical experiments demonstrate that the proposed ST-GED has excellent performance on both detection and state correction tasks. • Formulate a computationally efficient stealthy FDIA model based on a modified AC power flow. • Realize FDIA localization and state correction through a multi-task learning framework. • Develop an edge-aware graph attention message passing mechanism to incorporate line measurements explicitly.