Attentional Convolutional Neural Network Based on Distinction Enhancement and Information Fusion for FDIA Detection in Power Systems

A false data injection attack (FDIA) is one of the major threats to power systems, and identifying false data is critical to the stable operation of power systems. However, false data that closely resemble normal data hinder the accuracy of existing detection methods, and their performance further declines when exposed to ambient noise. To address these challenges, this paper proposes an attentional convolutional neural network based on distinction enhancement and information fusion (DEIF-ACNN) for FDIA detection. Firstly, by minimizing the loss of reconstruction and discrimination, this paper designed an autoencoder with a discriminator for normal data (NAE), which had the characteristic of producing a small loss for normal data. Secondly, the trained NAE is utilized to compute the feature correlation matrix between the original and reconstructed data to enhance the distinction between normal and false data. Finally, to enhance feature extraction and suppress ambient noise interference in detection, DEIF-ACNN incorporates a convolutional block attention module (CBAM) to emphasize key feature channels and highlight crucial regions in the feature matrix. Experimental results show that DEIF-ACNN outperforms other FDIA detection methods on IEEE-9, IEEE-14, and IEEE-118 bus power systems, achieving an accuracy of 99.22%, 99.83%, and 100.00%, respectively. In addition, the method exhibits the best robustness under different noise environments, and its accuracy is maintained at about 80%.