Time Series Forecasting to Detect Anomalous Behavior in Multiphase Flow Meters

As the world transitions toward more sustainable energy sources, many global energy productions still rely on oil and natural gas. Under these circumstances, the efficient management of reservoirs is paramount to ensuring a continuous flow of resources necessary to support both our economy and the transition itself. This paper introduces a novel approach to swiftly identify anomalies in multiphase flow meters, which are crucial instruments for reservoir management. By leveraging well-known technologies and algorithms, our method ensures reliability and availability during plant operations, providing a practical solution that is readily deployable in production settings. Specifically, we utilized the capabilities of Temporal Neural Networks to forecast signals, then compared the predicted values with actual ones using an improved thresholding technique. This approach stabilized the anomaly score, making it more interpretable for human operators, reduced potential false negatives, and generated additional insights into the nature of detected anomalies. We tested our method on a proprietary dataset containing both real and synthetically reproduced faults. Among our findings, we underscored the importance of avoiding the well-known issue of network adaptation to faults, which exploits the small-time scale correlation arising from the varied placements of sensors in multiphase flow meters.