Since approximately 2012, large multistage completions in horizontal wells have significantly boosted oil production in the Bakken Formation. As of the time of writing, more than 18,000 stimulated wells have been completed in Bakken using various completion designs.
Operators have experimented with well spacing, as well as the size and intensity of stimulation jobs by injecting increasingly large volumes of fluid and proppant to maximize oil production while minimizing costs. However, wells that received larger treatments did not always perform as expected. Furthermore, the aggregated impact of various interrelated completion design parameters and reservoir characteristics was not fully understood, leaving room for improvement in completion optimization evaluations.
Publicly accessible completion and production data, meticulously collected by the North Dakota Industrial Commission (NDIC), combined with advancements in data science, have created an excellent opportunity to optimize drilling and completion strategies using statistical data analysis and predictive modeling. Observations and experiences from thousands of producing Bakken wells can now be analyzed and interpreted using data mining techniques.
This article presents examples of data science applications in completion optimization calculations in Bakken production.
The initial optimization study analyzed drilling and completion results from more than 12,000 oil-producing wells across the Bakken in North Dakota using 2020 data (URTeC 3723843).
Simple interpretations of the relationship between production performance and completion parameters, based on bivariate (two-dimensional) scatterplots, proved challenging due to the nonlinear nature of these dependencies.
Data-mining techniques capable of accommodating nonlinear relationships and complex, incomplete information were applied to identify optimal completion practices. The dataset used in this analysis included eight publicly available completion design parameters:
1. Perforated interval
2. Injected fluid volume
3. Proppant amount
4. Stage count
5. Injection rate
6. Injection pressure
7. Proppant type
8. Completion type
These parameters were used to predict well performance, measured by cumulative 6-month oil production. Predictive modeling was performed using the gradient boosting (GB) data-mining tool, in which multiple decision trees described the variability of the multidimensional dataset. The initial predictive modeling conducted for the entire Bakken play demonstrated acceptable performance for the training model (evaluated by R²), but the test model—using 20% of all data—showed weaker performance.
This phenomenon, known as overfitting, occurred when a single statistical model struggled to accurately predict well production performance across various Bakken locations which are characterized by heterogeneous geology and reservoir properties.
