Evaluating GAN-LSTM for Smart Meter Anomaly Detection in Power Systems

TL;DR

The paper tackles anomaly detection in nonstationary smart-meter time series by evaluating a GAN–LSTM framework within a unified preprocessing and evaluation pipeline on the LEAD dataset, which contains one year of hourly consumption from $406$ buildings. It adopts a fixed $60$-hour window, trains only on normal windows, and uses latent-space optimization with an AnoGAN‑style inversion to produce per-window anomaly scores, achieving a peak F1 of $0.89$ and accuracy of $89.73\%$ on $206$ test buildings. The GAN–LSTM outperforms six baselines (statistical, kernel-based, reconstruction-based, and GAN-based), demonstrating the effectiveness of adversarial temporal modeling for asset monitoring, non-technical loss detection, and situational awareness in real-world power distribution networks. The work highlights practical utility in grid operations and points to future extensions to multivariate, spatio-temporal, and context-informed anomaly detection, along with improving inference efficiency.

Abstract

Advanced metering infrastructure (AMI) provides high-resolution electricity consumption data that can enhance monitoring, diagnosis, and decision making in modern power distribution systems. Detecting anomalies in these time-series measurements is challenging due to nonlinear, nonstationary, and multi-scale temporal behavior across diverse building types and operating conditions. This work presents a systematic, power-system-oriented evaluation of a GAN-LSTM framework for smart meter anomaly detection using the Large-scale Energy Anomaly Detection (LEAD) dataset, which contains one year of hourly measurements from 406 buildings. The proposed pipeline applies consistent preprocessing, temporal windowing, and threshold selection across all methods, and compares the GAN-LSTM approach against six widely used baselines, including statistical, kernel-based, reconstruction-based, and GAN-based models. Experimental results demonstrate that the GAN-LSTM significantly improves detection performance, achieving an F1-score of 0.89. These findings highlight the potential of adversarial temporal modeling as a practical tool for supporting asset monitoring, non-technical loss detection, and situational awareness in real-world power distribution networks. The code for this work is publicly available

Figures (4)

• Figure 1: Overview of the GAN-LSTM anomaly detection pipeline. Top: preprocessing and windowing. Center: adversarial training of the LSTM-based generator and discriminator on normal windows only. Bottom: test-time latent-space optimization with frozen networks to compute anomaly scores and final classification.
• Figure 2: Confusion matrix for the GAN-LSTM model on the test windows.
• Figure 3: Full-year consumption sequence for a sample building with detected anomalies (red) and ground-truth anomalies (green).
• Figure 4: Zoomed view of samples 200--400 showing anomaly detections in a dense irregular region.