Early Detection of Hydrate Formation in Deepwater Drilling Using an Ensemble of LSTM Networks

Abstract

A key safety problem for the offshore drilling industry is the sudden formation of gas hydrates, which cause blockages in deep water production pipelines. These blockages disrupt well operations, damage critical safety equipment, and can lead to rig explosions. Gas hydrates can form within minutes. Once formed, their removal can take weeks or months. The removal process is expensive and dangerous. The solution primarily used by oil companies to prevent formation - continuously pumping thermodynamic inhibitors down pipelines - can be unreliable, expensive, damaging to hydrocarbon quality, and environmentally unsafe. This research proposes a novel approach to this problem – to predict hydrate blockages in advance. A model was created, consisting of an ensemble of long-short term memory networks. Each network was created with a different architecture, allowing it to exploit a unique pattern in the data, and predictions were averaged to make the ensemble prediction. It was trained on multivariate time series data from 284 different well instances, with five sensor values monitored every second at critical places in the production pipeline. The model learned complex, nonlinear relationships from the sensor values over time and predicted the likelihood of hydrate formation with a high accuracy of 91.5% and an AUC ROC score of 0.96. It vastly outperformed a separate, benchmark time series Logistic regression model that was created. The model is cost effective and can be retrained for specific oil wells and is robust for a variety of different operating conditions due to its ensemble nature. In summary, it can potentially reduce environmental, worker safety, and efficiency problems in the oil industry.