[URTeC 2020]The Impact of Interwell Spacing Over Time: A Machine Learning Approach ( id 2800)

ABSTRACT::

Objectives/Scope: The decision to downspace wells can lead to significantly higher return on a given land acquisition, or catastrophic losses if drilling and completion costs cannot be recouped by each individual well. As operators push the boundaries of interwell spacing, the negative impacts may not be observed in the early production time. We demonstrate a machine learning approach to understanding the impacts of spacing, geology, and completion intensity on production rates through time.

Methods, Procedures, Process: We created multi-input-multi-output (MIMO) decision tree-based models to predict the first 2 years of wellhead production for gas, oil, and water at 30 day increments. This leads to an individual well prediction of 72 data points. The input variables we utilized were completions intensity (proppant and fluid), interwell spacing calculations based on directional surveys, and geological interpretations based on well logs in the Williston basin. After implementing an 80/20 test/train split, we examined the results of several test wells which were severely downspaced compared to the basin average. We then examined the Shapley feature importance values across time for spacing, completions, and geology.

Results, Observations, Conclusions: For wells with large completions designs and tight spacing, we notice a uniform trend across the two basins. Large completions designs have a strong positive impact on oil production in the first 5 months of production, while spacing variables have minimal impact. In contrast, between 5 and 24 months into production, the impact of high quality acreage and large completions is diminished, while the impact of spacing becomes strongly negative. We observe spacing impacts changing as much as 30% moving from IP30 to IP720, while completions and geology impacts can be reduced by as much as 75%.

Novel/Additive Information: Isolating the impact of well spacing and timing from geology and completions can be very difficult in shale basins where downspacing is still being tested. Furthermore, the negative impacts of downspacing may not be fully revealed until 6 months or more into a well’s production history. Machine learning approaches, combined with feature importance metrics, can help inform operator decisions on downspacing by both creating more accurate cumulative production forecasts, but decline curves as well.