Calculating a reliable per-well estimate of reserves and future production matters more than ever, especially in unconventional resources. And especially with today’s uncertain economics.
Good geodata is essential. But how much data is enough?
Research suggests that machine learning not only provides an advantage over traditional decline curve analysis, but also reduces the amount of current production data needed for a clear outlook. And the use of computer models can open a window for applying geological knowledge toward a better understanding of ultimate production and reserves.
“An advantage of machine learning algorithms is that the algorithm can take whatever approach it wants to declining a well,” said Ted Cross, director of product for Novi Labs in Austin, Texas.
Cross is an author for the paper “How much data is needed to create accurate PDP (proved developed producing well) forecasts?” with his Novi Labs colleagues Alex Cui, reservoir engineering adviser, and Austin Lim, reservoir engineering intern. Their research will be the subject of a presentation Wednesday morning, June 14, at the Unconventional Resources Technology Conference in Denver.
The Research Project
In their findings, they note that “the rigidity of the curve-fit workflow DCA models use does not allow DCA forecasts to react appropriately to unexpected deviations from an ideal hyperbolic decline. Meanwhile, ML models can learn and react appropriately to change the wells’ forecast.”
Three models were created for the PDP research project: a limited ML model using only production history, a full ML model using all available input and a DCA model.
To replicate a real-world scenario, the researchers withheld pieces of production data from the models and then reintroduced them. That helped determine the minimum decline history needed to create accurate forecasts for both the ML and DCA methods.
They found that the full ML model could forecast cumulative production with a median absolute percent error of 14.4 percent with only 90 days of decline data. The DCA model provided a median error of 16.1 percent with the same 90 days of data.
Also, program-level analysis showed that the ML models had a peak aggregate percent error between 23 and 29 percent, substantially better than the DCA peak APE of 54 percent.
Errors decreased over time for all the models as the amount of available production data history increased, converging at a median error of about 1 percent after 1,080 days of decline data.
“When the well is older in its life, its production is very stable and it’s easier to fit a curve to it, from a reservoir engineering viewpoint,” Cross said.
According to the researchers, the improvement in forecast reliability with limited data came from the ability of ML modeling to leverage additional information like geology, completions and spacing.
“One of the interesting things about ML models is that they’re able to do fairly well with simple geological data,” as simple as depth, longitude and latitude entries, Cross said.
Additional geological observations “especially offer help when you’re looking at new areas or areas that aren’t as heavily drilled. We’ve seen a wide variety of geological information become useful,” he observed.
The Human Factor
Cross said he has noticed “an almost iterative process between the ML model and the geological interpreters.” When geoscientists see what data the ML models use as a prediction basis, they focus on those areas.
Other cases are the opposite: What geologists think is important information can turn out to be of only minimal value in estimating future production. And sometimes the model response reveals incorrect human interpretation and input.
“I’ve seen this happen a couple of times, where the model says, ‘Hey, this area of the basin where there’s a lot of water saturation is good for production.’ When something like that happens, it lets you know something is amiss,” Cross said.
For computer skeptics looking for a catch in the ML advantage: There is a catch, of sorts. ML and computer modelling might produce more accurate results using a lower amount of recent-data input, but they require lots – and lots – of historical production data for ML training.
To make a reliable prediction for a current well, an ML model has to “learn” what has happened with similar wells in the past. Cross said the Novi Labs PDP research was conducted in the Bakken play in the Williston Basin, where there is an abundance of drilled wells and production data.
Unconventional wells that appear to have a predictable future can shift to a different production pattern when other wells are drilled and completed nearby, a challenge for ML. It’s a puzzle the industry has grappled with for years.
“The tricky part is that you can’t predict the future very easily for the new wells coming online,” Cross observed.
Other Forecast-related Content
Several other oral presentations at URTeC will address methods of predicting future production and reserves in unconventional resources, particularly in two Tuesday June 13 sessions on “Applications in Reserves Estimation and Production Forecasting.”
The morning session incudes the presentations “Data-driven oil production prediction and uncertainty quantification for unconventional asset development planning through machine learning” and “Predicting hydrocarbon production behavior in heterogeneous reservoir utilizing deep learning models.”
Novi Labs will examine a different kind of production diagnostics in its Thursday, June 15 morning presentation, “Revealing the production drivers for refracs in Williston,” with a goal of helping to screen refrac-candidate wells more effectively and to design appropriate refracture completion jobs.
That analysis also incorporates a machine-learning approach. Cross said that, in general, refracturing programs boost output and can extend the economic life of unconventional wells.
“From most of what we’ve seen they usually make sense – a successful refrac does usually result in a nice uplift in production and works out economically,” he noted.
The downside comes from operational challenges, he said. Not all refrac programs work out and not all pay off. As of right now, refracturing unconventional wells remains far from a proven winner.
“I will say, there’s still a lot of work being done on this,” he added.
The Data Transition
Cross has seen a great deal more adoption of computer modeling in the oil and gas industry in recent years.
“The whole industry got much more excited about computing about six years ago, but it’s taken time to integrate that,” he observed.
Companies now have spent years building databases, training staff and adopting new computing and modeling approaches, he noted. Also, almost all petroleum engineers and geoscientists coming out of college now have data training.
“It’s a nice transition that’s happening,” he said.