Deep learning-based sensitivity analysis of the effect of completion parameters on oil production

Abstract

Given that the goal of every hydrocarbon well is to maximize the amount of oil and/or gas produced, optimization of well completion and reservoir parameters is crucial for the development of an unconventional field. A good optimization tool is sensitivity analysis. Artificial Neural Networks (ANNs) are a fairly nascent but powerful technique that can be used to capture the effects of well completion and reservoir parameters on hydrocarbon production within a similar formation. This is because ANN models have a proven record of capturing underlying and complex correlations between complex dependent and explanatory variables. In this study, we propose using location, Volume of Fluid per Foot of Perforated Interval, Pounds of Proppant per Foot of Perforated Interval, Average Porosity, Average Water Saturation, and Average Permeability as input to train an ANN model that forecasts the first six months of oil production. We then opt to use the ANN model as a basis to explore the effect of various well completion and reservoir parameters (Volume of Fluid per Foot of Perforated Interval, Pounds of Proppant per Foot of Perforated Interval) on oil production. The dataset used consisted of 464 wells from the middle Bakken. 323 wells were used for model training, 69 for validation, 69 for testing, and 3 wells for sensitivity analysis. The average performance of the ANN model with the root mean squared error of 4109.31 bbl and R-squared of 0.78 suggests the workflow described in this study is a viable way to anticipate the oil production of a stimulated horizontal well. The sensitivity analysis then portrays a feasible way to infer production values, should completion and stimulation parameters change. • Production data are used to calibrate and evaluate reservoir potential. • A reservoir's capability is assessed using completion parameters. • Deep learning is used to relate reservoir and completion parameters to production. • A varied dataset is key to the representation of formation property distribution.