In this work, we analyze an extensive dataset of wells in the Delaware Basin to identify key factors affecting the performance of the new (child) wells. We propose depletion-volume-overlap idea as a new method for describing the spatial region of interference among communicating wells. This idea replaces the conventional 1D spacing which can have limitations when applied to non-ideal field situations and extends the concept to several wells with production interference. We formulate "depletion score" as an interference metric that includes the combined effect of reservoir depletion and spacing for tight oil wells. Data compiled for over 13,000 wells in the Delaware Basin was utilized to determine important features controlling well interference and to guide us to refine our definition for the depletion score. Using a machine-learning model, we assess the impact of the individual features on the productivity of new infill wells during the first year of production.
The depletion score is conceptualized based on our results of 360 physics-based simulations, each conducted for a pair of parent/child wells considering different spacing and depletion conditions. The depletion score is defined as a non-linear function that depends on: (1) the overlap of depletion volumes around the parent and child wells; (2) the parent's cumulative hydrocarbon production; and (3) a fitting parameter p. We consider the damage to child-fracture propagation often caused by alterations to rock mechanical stress, which occur as a result of depletion around the parent well. We assume three scenarios: severe, moderate, and no geomechanical-induced damage. In addition to the depletion score, eleven features are defined and calculated for the child wells to serve as the input for different machine-learning (ML) regression models. The goal of the ML models is to infer the relationship between the input variables and the productivity of the child well relative to that of the parent well.
Our results show that the interference metric is predictive of child-wells performance relative to the parent's performance (R2= 84% - 89%) when compared against the simulation results. However, when fitted to the field data, the interference metric is not as predictive and explains only 9.8% of the variance. Production from the field depends on additional factors such as reservoir quality and completion conditions. As the additional parameters are considered, the random-forest model explains 97.9% of the variance for the training data and 38.8% for the testing data following an 80/20 train/test data splitting scheme. We interpret the results of the machine-learning model using SHapley Additive exPlanations (SHAP), which provides insights on the importance of individual features and their interactions in explaining the prediction. The results suggest that depletion score has the greatest impact on the child well's performance. This is followed by the injected fluid ratio, the original oil-in-place, and the proppant concentration ratio.