Generative Pretraining Transformer Based False Data Injection Attack Detection Framework for DC Microgrid Under Uncertain Operating Condition

Traditional False Data Injection Attack (FDIA) techniques typically rely on prediction-based models that forecast Renewable Energy Source (RES) generation. This reliance leads to increased computational demands, added complexity, and greater uncertainty. Recent advancements in generative models with zero shot learning have provided alternatives to apply autoregressive modeling to other domains. In this paper, Generative Pretraining Transformer (GPT) based model is adapted for current measurement forecasting. The choice of GPT is motivated by its ability to capture complex temporal dependencies in sequential data and leverage generative pretraining to model nonlinear patterns in historical measurements, while also handling unknown inputs more robustly, which is an area where traditional prediction models often struggle. The forecasted values are compared with actual values based on sample of past measurements and any deviation past tuneable threshold can be used for detection of FDIA. Furthermore, the sample captures current trends and contextual information about system dynamics, enabling efficient prediction of future measurements even when current readings are corrupted by FDIA or exhibit high variability through zero-shot learning. Various scenarios were examined by incorporating FDIA while evaluating false positives and computational demands. Results indicated proposed framework provides good detection with less computational overhead and complexities.