Adaptive Contamination Estimation for Single-Cluster Anomaly Detection in PV Systems Using K-Distance Elbow Analysis

Abstract

Accurately estimating the contamination parameter—the proportion of anomalies—is a critical challenge in unsupervised anomaly detection, especially for univariate, single-cluster data as found in photovoltaic (PV) systems. This paper proposes a novel, automated framework that estimates contamination using the first derivative of the k-distance curve to detect the transition from normal to anomalous regions. The estimated contamination is then used to configure unsupervised models such as Isolation Forest, DBSCAN, KMeans, and Local Outlier Factor. To validate the effectiveness of the proposed estimation strategy, the framework is applied to real-world current data from a 65 MW PV plant under uniform irradiance conditions. Results demonstrate that the adaptive contamination closely matches actual anomaly ratios and significantly enhances detection accuracy—particularly when used with Isolation Forest—without requiring manual parameter tuning.