"Electricity anomaly detection dataset"

"This dataset, named \"Electricity Anomaly Detection Dataset,\" is designed for research on electricity theft detection using smart meter load data. It contains normal electricity consumption data and six representative types of anomaly (theft) behaviors: Dynamic Reduction, Peak Clipping, Random Reduction, Single-Point Zeroing, Peak Shifting, and Interval Zeroing. The total sample size is 7,000, with each of the six anomaly types contributing 1,000 samples and normal samples also providing 1,000, making the dataset perfectly balanced across categories. Each sample records a user's load curve over a 24-hour period with a sampling interval of 15 minutes, resulting in 96 consecutive load values per sample. The data can be used for binary classification (normal vs. anomaly), multi-class classification (identifying the specific theft technique), or unsupervised anomaly detection (training only on normal data). The balanced class distribution and realistic anomaly injection make this dataset suitable for benchmarking machine learning models for non-technical loss detection in power systems. Researchers are encouraged to use precision, recall, and F1-score as primary evaluation metrics due to the practical imbalance of theft events in real-world scenarios. The dataset is shared in \".npy\" format for easy integration with common data science tools.s categories."