Stability and Error Structure Analysis of Generative Adversarial Network–Based Architectures in Time Series Anomaly Detection

Abstract

The rapid expansion of Internet of Things ecosystems and interconnected network infrastructures has intensi%ed the need for robust time series anomaly-detection methods. Generative adversarial network (GAN)-based approaches have gained increasing attention due to their capacity to model complex data distributions and learn representations of normal system behavior without explicit supervision. However, adversarial training may introduce instability and sensitivity issues, particularly in sequential anomaly-detection tasks. This study presents a controlled comparative evaluation of three GAN-based architectures, GANAD, N-GAN, and Net-GAN, under uni%ed experimental conditions. All models are assessed on the Numenta Anomaly Benchmark artificial time series anomaly benchmark using identical preprocessing and an F1-optimized threshold calibration protocol to ensure methodological consistency. Performance is evaluated across ten random seeds using precision, recall, and F1-score to capture both effectiveness and robustness. The results reveal distinct architectural behaviors. The reconstruction-driven GANAD model achieves the highest mean F1-score (approximately 0.32) and maintains stable precision across runs, indicating consistent convergence and controlled false-positive rates. In contrast, the discriminator-centric Net-GAN architecture attains near-complete sensitivity (mean recall ! 0.96) but exhibits substantially reduced precision due to elevated false alarm rates. N-GAN demonstrates intermediate performance with greater variability across initializations, reflecting sensitivity to adversarial training dynamics. These findings highlight an intrinsic stability–sensitivity trade-off in GAN-based anomaly-detection frameworks. The study provides empirical evidence that architectural design fundamentally determines error structure and operational suitability, offering practical insights for deployment-oriented anomaly detection in time series environments.

Cite this article as: B. Çorbacıoglu and S. D. Odabaşı, "Stability and error structure analysis of generative adversarial network–based architectures in time series anomaly detection," Electrica, 26, 0062, 2026. doi: 10.5152/electrica.2026.26062.