Zoubeyr Farah

The increasing vulnerability of power systems to cyber-attacks has underscored the need for robust data-driven anomaly detection methods. However, existing datasets used to evaluate these methods have several limitations, including restricted availability, limited attack scenarios, and a lack of fidelity assessment. To address these challenges, we first propose a hybrid framework that combines model-based and data-driven methods to generate a dataset under normal operation based on IEEE bus systems (4-, 14-, 39-, and 57-bus systems). The framework solves the direct current (DC) power flow problem to obtain initial measurements. A public load profile is scaled and distributed across these values to generate time-series data. A Kalman filter is integrated for state estimation and filtering. A neural network adaptively fuses the Kalman-filtered data with the original time-series data, generating accurate, temporally correlated measurements that reflect system dynamics. It is trained using physics-informed ground truth derived from DC power flow equations and Kirchhoff’s principles to ensure physical consistency. Second, we simulate two attack scenarios: 1) unpoisoned data, where cyberattacks evade traditional Bad Data Detection (BDD), and 2) data poisoning, where adversarial examples are generated using a Projected Gradient Descent (PGD) attack to improve the robustness of data-driven detection models through adversarial training. Finally, we evaluate supervised and unsupervised machine learning and deep learning models on the proposed dataset under both scenarios. Evaluation results show that the proposed dataset is physically consistent and temporally coherent, offering higher fidelity compared to the benchmark dataset. Moreover, the results reveal that deep neural models remain vulnerable to adversarial threats; however, the proposed poisoned data enhances their robustness through adversarial training.