Spatial-Temporal Attention Based Interpretable Deep Framework for FDIA Detection in Smart Grid

Abstract

False data injection attacks (FDIA) destroy the integrity of information transmission by evading the bad data detection mechanism, and thus affects the stability of power cyber-physical systems (PCPS). Existing studies simply introduce complex neural network models for FDIA detection, ignoring spatial-temporal correlation and interpretability of neural networks. As a result, the accuracy and reliability of false detection may be negatively affected. To address the challenges above, this paper proposes an interpretable deep learning framework based on the spatial-temporal attention mechanism. Firstly, based on the gated recurrent unit (GRU), a dual attention mechanism is designed by combining spatial and temporal features of deep neural network to dynamically mine the potential correlations between the FDIA detection and the input features. Besides, the quantification of attention weights is introduced to interpret the spatial-temporal correlations between normal and attack data, which can effectively enhance the interpretability and reliability of detection results. Finally, based on the IEEE 14-bus test system and real operation data, simulations are conducted and the results show that the proposed STAGN model can detect FDIA effectively, has higher accuracy and stability than the latest detection models, and also has reasonable interpretability.