In Part 4 of this series, we introduced causal modeling, and its utility in debiasing horizontal well datasets. In addition to algorithmic methods, there are also feature engineering improvements we have implemented to provide better information to models about a well’s surroundings. In Part 4, we simulated 4 different spacing scenarios: 4, 6, 8, and 10 wells per section per zone. Another way to describe this is 1320, 880, 660, and 528 feet between wells. Yet another way to describe this is 16, 24, 32, 40 wells in the cube. If there are many different ways to describe pad development to fellow humans, how should we communicate spacing information to a machine learning model?
In this post, using the same pad location and layouts from Part 4, we will compare the results of 3 different geometric descriptions of well spacing. All 3 scenarios used causal models to generate their forecasts, with identical geology and completions information feeding the models. The only differences were in the spacing descriptions used:
- Linear neighbor distance:
the features passed to the model are the straight line distances to neighboring wells. This corresponds to the 528, 660, 880, and 1320 feet descriptions. The spacing features used here are distance to farther lateral neighbor and distance to farther stagger neighbor, in units of feet. - Nonlinear transformed neighbor distance:
We still use the distance to neighbors as the basis, but the distance in feet is transformed to a nonlinear function, making the variable impact higher with increasing proximity. These features also account for partial overlap between neighbors – although that is not a consideration for the idealized pads shown here. - Voronoi polygon:
This method computes a bounding box around each well, to inform the model of an available stimulated reservoir volume. It is also nonlinear, increasing the production penalty superlinearly as the available volume decreases.
Test Setup and Spacing Scenarios
Now, we examine our same Delaware test pad simulated with the 3 different causal model results. Relative to Part 4, we have added 2 spacing scenarios at the extremes. For this study, we ran (2, 4, 6, 8, 10) wells per section per zone, equivalent to (2640, 1320, 880, 660, 528) lateral feet between wells.
Observations and Key Differences
The biggest difference between model results here is the rate of change of degradation as wells are tightened together. In the case of both the linear and nonlinear distance models, we classify wells as staggers or lateral neighbors, which makes the model results highly dependent on how we choose to describe spacing information. Using a drainage polygon around each well removes the need to make specific decisions about lateral and stagger well classifications, and attempts to describe a stimulated reservoir volume around each well. In Figure 2 we see the results of different spacing scenarios on our Delaware Basin test pad.
Figure 1: Winerack scenarios in the Delaware Basin. Scenario with 2 wells per section per zone not pictured. As more wells are added and the cube is tightened, the Wolfcamp zones experience more crowding with their proximity to the next formation. Bone Spring targets do not experience this impact, being 1000 feet away in depth from the next zone.
Figure 2: Production declines with tightened spacing across 3 different model types. The model using linear spacing features (1A and 1B) shows a consistent decline in slope as wells are tightened together. This creates a diminishing degradation as the distance between wells becomes smaller. The entire curve has a positive 2nd derivative. In 2A and 2B, we show the results of nonlinear transformations of distance features. This model shows a larger magnitude of spacing degradation, but puts a strong weight on the stagger distances. The close proximity between the Wolfcamp zones forecasts a large degradation relative to the Bone Springs, which only experience lateral neighbor interference. In 3A and 3B, we see the results of the Voronoi polygon method.Conclusion
Our comparative analysis reveals that while each method has merits, the choice of spacing description impacts forecast accuracy significantly. Linear and nonlinear models provide insight into lateral and stagger effects, but may require precise spacing classifications. Meanwhile, the Voronoi polygon method simplifies spacing descriptions and prioritizes reservoir volume, offering a comprehensive approach when direct classifications aren’t feasible.
Key Takeaways
- A Delaware Basin test pad compared three geometric well-spacing feature sets with identical geology and completions inputs in causal models.
- Five spacing scenarios were tested at 2, 4, 6, 8, and 10 wells per section per zone, equivalent to 2640, 1320, 880, 660, and 528 feet.
- Linear neighbor-distance features produced a consistent decline in slope as spacing tightened, yielding diminishing incremental degradation at closer well distances.
- Nonlinear distance transformations increased the magnitude of spacing degradation and weighted stagger distances more strongly than lateral-neighbor distances.
- Nonlinear distance models forecast larger Wolfcamp degradation than Bone Spring due to close inter-zone proximity versus ~1000 feet depth separation for Bone Spring.
- Voronoi polygon features removed lateral-versus-stagger classification by modeling available stimulated reservoir volume and applying superlinear penalties as volume decreases.
