Maintaining ESP operational efficiency through machine learning-based anomaly detection

Abstract

During oil extraction from wells, Electric Submersible Pumps (ESP) often work in Periodic Short-Term Activation (PSA) mode. While efficient, this operational mode is prone to transitioning into abnormal states that can lead to significant operational losses. Therefore, we should aim for prompt detection of these transitions.

This study introduces a novel machine learning-based approach for detecting anomalies in ESP operations by leveraging a pre-established library of generalized anomalies and comparing them with current operational modes. Unlike conventional methods that rely on extensive telemetry, our approach focuses on frequency and load time series data, utilizing Dynamic Time Warping (DTW) to identify similar anomaly patterns. The advantage of DTW in this context lies in its superior ability to align sequences with temporal shifts, making it particularly suitable for the erratic data generated by ESP operations, which are often more complex and less predictable than those handled by traditional techniques.

We have integrated this methodology into real-world SCADA systems to provide early warnings and facilitate timely interventions that optimize oil production. Testing on real data from 1,667 wells in Western Siberia, the proposed approach shows high effectiveness with an accuracy of 93% and an F1 score of 88%. The integration of the anomaly library with automatic feature extraction significantly enhances the quality of predictions, swiftly identifying potential issues and recommending operational adjustments to maintain efficiency and profitability.