Enhancing Locational FDIA Detection in Smart Grids: A Hyperparameter Optimization Analysis

In cyber-physical power systems, mitigating the detrimental effects associated with the stealthy nature of False Data Injection Attacks (FDIAs) is imperative. This paper introduces a deep learning model based on a 1-D Convolutional Neural Network (1-D CNN) for simultaneously localizing and detecting FDIAs. To enhance the predictive accuracy of the CNN model, the hyperparameters must be tweaked. An analysis focused on hyperparameter optimization is performed by comparing various optimizers, including random search, hill climbing, and Bayesian Optimization (BO). To this end, the number of filters, kernel size, and number of layers of the CNN structure are optimally tuned using Graphics Processing Unit (GPU) resources. The findings of this study present evidence that the BO method is perfectly tailored for the hyperparameter tuning task with minimum error in comparison with competitive models. The efficiency of the CNN-BO method is demonstrated with two benchmark representative examples from the IEEE 14-bus system and the IEEE 118-bus system. For the IEEE 14-bus system, BO uses 256 filters, five kernels, and four layers; the IEEE 118-bus system uses 128 filters, three kernels, and eight layers. The CNN-BO method outperforms other optimization algorithms by achieving an impressive locational detection accuracy of 96.67% for the IEEE 14-bus system.