Spatio\u2011Temporal Power System Time Series with Adversarial Attack Scenarios

This dataset provides spatio-temporal time series data that is both physically consistent, respecting power system balance, and temporally coherent that captures the power system dynamics. It is generated based on the IEEE standard bus systems (4, 14, 39, and 57 buses) by fusing Kalman-filtered time series with noisy measurements through physics informed neural network guided by DC power flow equations to produce high-quality data suitable for data-driven anomaly detection. The dataset includes data under normal operation that can be  for simulating new threats and prediction, as well as two attack scenarios namely unpoisoned and poisoned data scenarios for evaluating data-driven methods. In the unpoisoned attack scenario, it provides training and testing sets for both supervised and unsupervised learning, including attack types namely False Data Injection Attacks (FDIA) and three types of Replay Attacks. In the poisoned attack scenario, 1% of the training data is adversarially perturbed for each bus system using Projected Gradient Descent (PGD) to simulate feature-space poisoning and enable robust model training. Additionally, the dataset contains four stealthy FDIA vectors for each bus system. We also include the source code for simulating FDIA and Replay attacks, the code for the neural network used to fuse Kalman-filtered and noisy time series, and four state-of-the-art machine learning and deep learning models selected for benchmarking. This dataset serves as a valuable resource for researchers working on anomaly detection in power systems.