همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Adaptive Deep Learning Framework with Uncertainty-Aware Drift Detection, and Adversarial Robustness for Anomaly Classification in Power Systems State Estimation

نویسندگان :

Seyed Mohammad Shobeiry¹ Mohammad Taghi Ameli² Mohammad Sadegh Sepasian³

1- دانشگاه شهید بهشتی 2- دانشگاه شهید بهشتی 3- دانشگاه شهید بهشتی

کلمات کلیدی :

Anomaly Classification،Adaptive Deep Learning،Power Systems State Estimation،Concept Drift

چکیده :

Power System State Estimation (PSSE) is increasingly threatened by False Data Injection Attacks (FDIAs) and Sudden Load Changes (SLC) under evolving operational conditions driven by Concept Drift (CD). To address these challenges, this study introduces an Adaptive Deep Learning (ADL) framework that integrates uncertainty-aware drift detection and adversarial robustness for anomaly classification in non-stationary power systems. The proposed framework employs multi-scale temporal attention to capture dynamic behaviors across multiple time horizons, applies Bayesian uncertainty estimation via Monte Carlo Dropout to distinguish epistemic from aleatoric uncertainty, and adopts FGSM-based adversarial training to enhance resistance to evasion attacks. A memory buffer with class-specific importance scoring ensures continual adaptation while avoiding catastrophic forgetting. An experimental evaluation on the IEEE 14-bus system shows that the ADL model maintains 83.6% accuracy under severe concept drift (only 3.4% below its baseline), while conventional deep learning models drop by 28.3%. It also achieves 97.06% accuracy under adversarial conditions, exhibiting strong generalization to various attack types. Overall, the proposed framework delivers robust, uncertainty-aware anomaly detection, enhancing cybersecurity resilience in modern power grids.