Artificial lift optimization is one of the core activities Production Engineers and Techs are asked to perform on a daily basis. Rod lift is the most widely used artificial lift type, deployed on horizontal and vertical wells alike. Due to the widespread use of rod pumps, consistent best practices have been established over time. When those operational best practices are adhered to, meaningful increases in field profitability of even the lowest producing wells have followed.
Despite their differences in the horizontal development and vertical legacy well context, industry best practices to optimize rod lift around efficiency, and thus profitability, have a consistent logic and methodology. First, wells are diagnosed as underpumping, dialed in, or overpumping. Second, based on the categorical classification, a consistent remediation workflow is applied based on the available levers to pull that either increase production or lower the number of damaging strokes into the system.
The primary daily operational set point levers to optimize rod lift wells are on/off time, SPM, and pump fillage. Following the Pareto principle, these levers represent the lion’s share of the value-added changes that an engineer or tech can make to optimize wells. Thus, the third and final step is to apply a change to the control system (POC, VFD, timer) and observe the before-and-after to judge success.
Ideally, this workflow is performed for every well on a regular basis. However, this involves significant time and resources to accomplish, and even when done daily can still neglect over 99% of the strokes put into a system during a well’s run-life. The reality is that field personnel have too many wells, too little time, and not enough of the right data or technology to enable a step-change in profitability.
The question, then, is: why have humans perform such a Sisyphean task? Or rather, in the era of modern technology, why are humans still tasked with carrying out a repetitive, defined logic workflow such as set point changes when we are much better suited to solve more complex and nuanced field problems? Wouldn’t E&P operations be better served following the manufacturing model by letting machines automate optimization set point changes? This would allow operators, techs, and engineers to do what they do best: solve problems, design solutions, and think strategically, while keeping wells dialed in at all times.
This post (originally released as a whitepaper) outlines how machine learning combined with rod pump domain expertise delivers operators a step-change in operating leverage and optimization capabilities in the context of vertical wells. If you are interested in reading more about horizontal well optimization.
In the context of conventional vertical wells, there exist vast fields comprised of thousands of unconnected and uninstrumented rod pump wells. Each of these wells makes a few barrels of oil per day, which has traditionally restricted the economic case for retrofitting with conventional automation technologies beyond a percent timer. As a result, wells are visited wwith an ‘every well, every day’ mentality, every day’ mentality, where routines generally trump criteria-based route prioritization. Sometimes these fields are closely bunched together, while other times wells span across hundreds of miles. This means if wells go down after the daily check, they are down until the next day.
In reality, operations teams are generally in reactive mode and focused on addressing critical, non-well-related activities at the central batteries or across gathering systems. With hundreds of wells to look after per pumper, these wells may not get looked at for weeks. We have spoken with multiple operators who manage legacy fields; one engineer had over 1,000 wells under his direction, and another said when his field informed him that a well was down, he realized he had no idea when it was actually last pumping. Adding to this is the lack of downhole visibility or remote control capabilities. Optimization efforts are slow and inconsistent as a result, generally following rules of thumb to fill the gap. There is no well data to view, no feedback loop to quantify the impact, and no confirmation a change was ever made. This severely restricts any value creation beyond basic maintenance.
Similar to a timer, our High-Resolution Adaptive Controller (HRAC) devices are tied into the cross/soft starter panel of the electric motor. Originally, Ambyint provided real time, remote visibility of on/off well status and torque trending. This allowed us to satisfy base field monitoring requirements.
We have now advanced our functionality to alarm on periods outside of the timer, which we call alert grace periods. If a timer is set for 1 hour of downtime and the alert grace period is 15 minutes, a user would receive an alarm after 75 minutes. He or she would know the well is down for an unplanned reason, and a visit is required.
From there, our team created and released the functionality to replace the onsite timer. Now operations and engineers have the ability to control on/off cycle times remotely at their desk or on their phone like any other well possessing significantly more technology. By pairing this with torque analysis trends in the platform and validating changes with production, users can systematically track performance and optimize their low value wells in the same platform as their higher value wells.
With a consistent and reliable edge device and reliable flow of clean data in our end-to-end platform, the fundamentals are in place for machine learning to provide adaptive control for stripper wells. This allows us to layer on machine learning techniques such as neural net and anomaly detection without the data reliability or sensor issues that have traditionally created problems in applying these algorithms, and moves another step closer to our goal of autonomous well operations.
Look at the figures below to see the similarities to how value is added in the field in the operational hierarchy of needs versus how high-end data science is enabled. The two figures mirror each other. This is why we systematically built our solution from the bottom-up to increasingly satisfy key pain-points and fulfill the needs of legacy fields.
Machine learning algorithms are able to look at every stroke and identify well behaviors automatically. But, much like humans, machines also require training by domain experts to become truly effective at their jobs. Studies across multiple industries show that even the most accurate machine learning systems involve “humans in the loop”, which can account for up to 20% of the solution. This is primarily accomplished either through helping label training data sets or correcting inaccurate predictions to continually refine the algorithm. We subscribe to this methodology across both of these functions, having production engineers collaborate with developers and data scientists alike.
In the example below, for the comparison to electric motor torque, a machine algorithm simultaneously looks at multiple patterns to detect if a human is walking or running. This parallels on/off cycle well optimization, in which a computer is able to detect underpumping (walking) or overpumping (sprinting) conditions through torque trends. In the underpumping conditions, torque (y-axis) stays consistent through every cycle, indicating that the well is not pumping off. The timer on-time is extended until torque starts to drop and then shuts off. Conversely for overpumping, torque falls off immediately so timer off time is extended to allow for a full pump. The screen-cap below shows regular torque patterns falling off without any downtime in between. This indicates an overpumping well where downtime should be introduced to reduce fluid pounding conditions and save electricity.
A machine can execute well optimization set point adjustments more quickly, objectively, consistently, and routinely. This is a win-win for field personnel because, since they are unburdened by routine optimization changes, they can focus on delivering higher value tasks. Additionally, management is happy because wells are dialed in, which means they produce more oil and fail less frequently.
We are currently deploying hundreds of devices to stripper wells across multiple operators–both large, public independents and smaller, private equity backed–across basins in Texas & New Mexico. This provides an unparalleled data set to deploy, field test, and refine adaptive control algorithms while delivering real field value today through implementation of a ‘pump by exception’ operating model.