Enhancing pressure gradient prediction in multi-phase flow through diverse well geometries of North American shale gas fields using deep learning

To ensure the efficient operation of production wells and to avoid flow assurance issues, accurately predicting pressure gradients, Δp/ΔL, in wells is crucial. This study explores the relationship between Δp/ΔL and production well geometries. The data from 29,047 wells across five shale gas basins is utilized: Bakken, Barnett, British Columbia, Marcellus, and Delaware. These basins have distinct shale gas reservoir properties and diverse well geometries. Each well's geometry was characterized by ten parameters, and k-medoids clustering was applied to identify geometry patterns. However, this method overlooked minor geometric variations impacting Δp/ΔL. An adaptive factorization network model, a deep learning approach, was employed to address this. This model demonstrated significant predictive accuracy, as evidenced by high determination coefficients, R2, comparing actual and predicted Δp/ΔL in test data (overall 0.904; Bakken 0.840; Barnett 0.848; BC 0.844; Marcellus 0.872; Loving-Delaware 0.914; Reeves-Delaware 0.880). The method's reliability in forecasting Δp/ΔL is supported by its R2 values of ∼0.8 and the mean absolute percentage error of ∼5 %, indicating its applicability in newly developing basins. This methodology offers reliable Δp/ΔL predictions for all possible well geometry scenarios in shale gas fields but also eliminates the need for additional multi-phase flow simulations, enhancing design and production plan.