Feature Engineering for Semi-supervised Electricity Theft Detection in AMI

Cyber attacks targeting Advanced Metering Infrastructure (AMI) of smart grids like electricity theft can cause significant financial losses to utility companies. Data-driven electricity theft detection methods based on machine learning techniques have become popular in recent years due to the massive amount of data and information collected by smart meters in AMI. Those methods are mainly based on unsupervised learning or supervised learning, both with limitations. In this paper, we proposed to use semi-supervised outlier detection for data-driven electricity theft detection, which only uses normal energy usage data to train detection models. Nine semi-supervised outlier detection algorithms were investigated in our study. To improve the performance of those algorithms, we performed feature engineering to extract 20 time-series features from energy load profiles, which help detect malicious changes in the load curve or energy usage amount. The performance of semi-supervised detection algorithms with and without feature engineering was evaluated with a real-world smart meter dataset under eight different False Data Injection (FDI) attacks. The results demonstrate that the proposed feature engineering offers a significant performance improvement for semi-supervised outlier detection algorithms.