Predictive Maintenance Optimizes Gas Well Production

Mike Thurber

November 13, 2017


To offset the fluctuations in the cost of oil, the oil and gas industry looks for improved efficiencies in all parts of its production chain. Predictive analytics leverages the large volumes and variety of historical well data to find critical patterns to improve performance, reduce losses, enable operators to be more proactive in field operations, and reduce operational costs.

This article describes how sensor analytics and predictive maintenance helped to prioritize gas well intervention to reduce downhole freeze events and reduce remediation cost.  

The Challenge

Natural gas wells have a propensity to “freeze” (or “shut-in” due to hydrate obstruc­tion), especially during harsh winter months, losing or deferring gas production and incurring intervention costs to disperse the obstruction and/or install a down­hole pump. Predicting when each natural gas well is at risk for freezing and when it is likely to need a downhole pump installed is a challenging problem, but know­ing the risks in advance enables effective prevention.

The client engaged Elder Research to discover and communicate actionable insights about well production and propensity to freeze (or experience pump failure) based on well design, envi­ronmental conditions, well interventions, and other data gathered by the client. The goal was to use predictive maintenance to optimize intervention resources to reduce operational costs and natural gas deferrals.

The Solution

Over a terabyte of operational well data was harnessed to identify and predict well freeze events on about 1,700 wells. The sensor data was mostly from well­head and separator tanks and included temperatures, pressures, tank levels, and flow rates. Also, operational reports of freeze events were provided, and records of well design, well environment, and production history played predictive roles in the models. The data was large, diverse, noisy, often erroneous, and incomplete. To improve model performance many rules were developed and ap­plied to clean the data and impute estimated values when data feeds were down or un­reliable. With input from the client key ratios that affect the well’s propensity to freeze within the next three months were calculated.

Elder Research modeled well freezing and the need for plunger pump installation to predict the impact of interventions on gas production volume. Several modeling algorithms — including decision trees, neural networks, regression, and ensembles of multiple techniques — were applied, tuned, and evaluated. Figure 1 shows the freeze model performance compared to random decisions while Figure 2 demonstrates the benefit of knowing when pump installation and activation should to be completed.

Well Freeze case Study_Freeze Model vs Random.png

 Figure 1. Freeze Model performance compared to random decisions

Using an Agile model building process, the insights about predictive relationships were reviewed on regular intervals with client subject matter experts. The team applied the predictive models to all wells with adequate data avail­able and predicted which wells would improve produc­tion if an effective intervention were made. Custom data visualizations were designed to illustrate import­ant relationships.

Well Freeze case Study_Plunger Pump Predictions.png

Figure 2. Total gas production deferrals (MSCF, or million standard cubic feet) versus recommended time to plunger installation date. The recommended time after start of well production is before the 50th percentile date. This demonstrates the benefit of knowing when pump installation and activation should to be completed. The required information is available within one month of start of production.


The models were thoroughly vetted by the client using a blind validation process and revealed actionable insights about well production and freeze propensity. The Plunger Pump model produced useful and flexible maintenance predictions for optimizing the timing of plunger pump installation and was pro­jected to reduce production deferrals by 15-20%. Testing of the Downhole Freeze model revealed that it had a 70% success rate. Its insight helped the client prioritize well interventions to reduce well freeze remediation costs, with projected savings of 600,000 MSCF (million standard cubic feet) in deferred gas volume annually.

Download the case study Using Sensor Analytics to Predict Natural Gas Well Freezing


Learn more about Energy Analytics

Learn more about Sensor Data Analytics applications 

Read the blog Analytics Best Practices – Agile Data Science

About the Author

Mike Thurber Mike Thurber is an analysis professional who listens carefully to partners to master an organization’s objectives and challenges, and he has a passion for extracting relevant and valuable insights from available data in a collaborative setting. As a trusted data science consultant, he clearly communicates deep analytical insights to managers and leaders regarding decision alternatives to help them improve key outcomes. He has 20 years of experience modeling causal relationships between potential actions and desired outcomes. He has 30 years of experience procuring and transforming historical data for descriptive analysis, statistical testing, predictive modeling, and deep learning. As a Principal Scientist at Elder Research, a highly regarded data science consultancy, he has delivered a broad range of advanced analytic solutions across many industries, as well as training, mentoring, and leading other data scientists. Mike’s work has ranged from estimating the profitability, risk, and responsiveness of credit card prospects, to identifying which infants will be negatively impacted by a Cesarean delivery. He has gleaned insights on how complex consumer choices impact sales, modeled individual healthcare providers rank in achieving desired patient outcomes, and calculated fraud risk and identified emerging fraud types. His projects have shown how call center interactions affect customer retention, measured the effect of targeted messages in political campaigns, forecasted debt recoveries at the account-holder level, modeled maintenance events on natural gas wells, and predicted propensity of past benefactors to make voluntary donations. Finally, especially in the last five years, Mike has been teaching principles and best practices of data science to a broad professional audience of emerging and experienced data scientists, with an emphasis on predictive and prescriptive modeling in an AI setting.