In this paper, we present a successful implementation for the development and deployment of a system that automates and simplifies the surveillance and optimization workflows for the gas lifted wells at a platform in the Gulf of Mexico.
As with other "digital oil field" type systems, the success of the capability deployed, called Gas Lift Optimization Workflows (GLOW™), required the use of diverse and more-than-capable library of techniques, algorithms, and methods that already exist in the fields of optimization, machine learning, signal processing, physical modeling, and numerical simulation. We adopted a philosophy in GLOW of relying on our understanding of the underlying physics as much as possible to achieve our goals.
We found that most of our problems, like the diagnosis of steady-state gas lift injection through multiple valves or solving for the optimal allocation of gas lift to every well, could be solved to an appropriate level of accuracy with physical models that are automatically matched to well-test and real-time data coming from the field using non-linear regression techniques. We resorted to using statistical modeling approaches, like Naïve Bayes classification for slugging detection or Kalman Filtering for reservoir pressure prediction, in situations where uncertainty precluded the use of physical models. The benefits of this system include production uplift, increased operator efficiency, optimal use of gas lift experts, and timely mitigation of issues
The key challenge in developing GLOW™ was to figure out how to stitch some of these methods together and create a system to enable operators and gas lift specialists to monitor and optimize all the gas lifted wells given limitations in data quality and coverage.