An Explicit Machine Learning Model for Predicting Sandface Pressure in Complex Shale Reservoirs

Abstract Predicting sandface pressure in shale is inherently complex due to the fact that the reservoirs are highly heterogeneous in nature. Shale reservoirs, defined by ultra-low permeability (nano-Darcy range) and convoluted pore structures, usually require advanced techniques like hydraulic fracturing to enable production. This complexity arises majorly because of the variability in permeability, presence of natural fractures, and multiphase flow dynamics that exist within the reservoirs, making conventional modeling approaches inadequate for accurate sandface pressure predictions. In this paper, we utilized four machine learning algorithms, viz., Gradient Boosting Machine (GBM), Decision Tree (DT), K-Nearest Neighbors (KNN), and an Artificial Neural Network (ANN), to effectively predict sandface pressure using the Vertical Lift Performance (VLP) method across five shale reservoirs: Eagle Ford, Haynesville, Bossier, Marcellus, and Marcellus-upper. Unlike the conventional practice of Inflow Performance Relationship (IPR), where multiple petrophysical inputs are used, this work focused on a simplified set of parameters, including porosity as the sole petrophysical property and ten other production parameters: casing, tubing, and line pressures (psia); gas and oil production (MMscf and stb); gas, oil, and water volume (MMscf, stb); measured and true vertical depth (ft). A dataset containing 9,561 data points was used in developing four robust computational intelligence models. ANN consistently outperformed GBM (R²= 0.9350; MSE = 0.0710), DT (R²= 0.9485; MSE = 0.0562), and KNN (R² = 0.1550; MSE = 0.9219), demonstrating superior predictive accuracy as evidenced by the highest R² value (0.9626) and lowest MSE value (0.0053524). This is because of the neural network's ability to model non-linear relationships and manage high-dimensional data. The results also show that porosity, when used alongside production parameters, can suffice for accurate sandface pressure predictions, reducing the reliance on extensive petrophysical datasets, which are often not readily available. Four major features make this work exceptional: (i) The developed model is presented explicitly, supporting an easy deployment in software apps; (ii) computational cost analysis was conducted to evaluate the model's scalability, observing that the developed ANN model requires a memory footprint of only 424 bytes to run on a software application; (iii) the use of a connection weights algorithm for sensitivity analysis of each input parameter and their influence on the model's performance; and (iv) a step-wise guide on how oil and gas industry practitioners can deploy the model in a wellhead monitoring facility on any rig. Given that an understanding of pressure behavior is crucial for optimizing reservoir performance, this research is very relevant for oil and gas operators whose aims are to improve hydrocarbon recovery, reduce sandface completion costs, and enhance overall well productivity.