Univariate Time-Series Anomaly Detection in Crude Oil Production Volumes Using Unsupervised Learning Techniques

Abstract

Accurate and timely detection of production anomalies is critical for optimizing well performance and preventing unplanned shutdowns in crude oil operations. However, many producing fields lack comprehensive multivariate sensor data, limiting the effectiveness of traditional monitoring approaches. This study presents a univariate time-series anomaly detection framework that leverages unsupervised learning techniques to identify abnormal patterns in crude oil production volumes using 71 months of national production data (January 2020 to November 2025). Historical production data were preprocessed, normalized, and transformed into rolling statistical features to capture temporal dynamics. Four unsupervised algorithms: K-Means, DBSCAN, Isolation Forest, and Autoencoder neural networks were implemented and evaluated for their ability to distinguish normal from anomalous production behavior. The Isolation Forest and Autoencoder models demonstrated superior sensitivity, exhibiting the strongest alignment with high-confidence anomaly consensus and effectively capturing subtle deviations in production patterns. The framework successfully detected significant anomalies without relying on labeled data, utilizing only historical production time series and engineered features. This approach provides a robust and data-efficient solution for production monitoring in sensor-limited environments and establishes a scalable foundation for future integration into real-time digital oilfield systems.