If you want to catch a wily mouse, you need to build a better mouse trap. If you want to tap into increasingly more difficult reservoirs in the search for oil and gas, you need to drill a better well. When it comes to global hydrocarbon reserves, the low-hanging fruit was picked over years ago. What’s left and available for exploration is generally ultradeep, super-hot, overpressured, or some combination of that trio.
Loss of well control can lead to disaster. The world remembers 20 April 2010, when 11 workers lost their lives onboard the BP-operated semisubmersible Deepwater Horizon. A blowout occurred on the Macondo well in the deepwater Mississippi Canyon area of the US Gulf of Mexico. The rig was consumed by fire and ultimately sank. The Macondo well flowed uncontrolled oil and gas into US Gulf waters for 87 days.
In paper OTC 35818, a study led by researchers at Louisiana State University (LSU), Blade Energy Partners, and Intellicess, aimed to address the gaps in the current understanding and application of leading indicators for loss of well control.
The study employed methodologies including data assimilation (DA) and change point-Bayesian networks (CP-BN) to develop an integrated real time kick detection and state estimation framework and evaluated the effectiveness and reliability of the identified indicators
Data was gathered using surface and downhole gauges and transducers in addition to a set of fiber optic sensors that were installed outside of tubing string and used to record the temperature and vibration during the migration/circulation of gas influxes. The system architecture is designed to leverage the strengths of each component while compensating for individual limitations, ultimately delivering more reliable and accurate kick detection.
The CP-BN algorithm forms the foundation of the rapid detection capabilities. The algorithm processes multiple data streams, including real time measurement data and driller’s memos, through a data pipeline that begins with natural language processing of operational notes. The processed data undergoes measurement noise filtering before entering a time-series-based CP-BN detection module. These refined data streams feed into a Bayesian network model that assesses the probability of kick events.
Paper Number: SPE-0625-0005-JPT