Perpustakaan Digital ITB

Flowing bottomhole pressure governs how a producing well is evaluated and optimized, yet continuous downhole measurement is rarely available throughout its operating life. This study embeds the Beggs-Brill pressure-gradient calculation within the architecture of a neural network. Gravitational, frictional, and acceleration components are computed deterministically along the wellbore, and the network learns one daily correction factor for each before the corrected gradients are integrated to gauge depth. Daily records from five wells in the Volve Field, 5,305 observations between February 2008 and September 2016, were partitioned chronologically into training, validation, and testing subsets. Three reference configurations were built on identical partitions: a vertical lift performance model using the Petroleum Experts 4 correlation, conventional learning models trained on surface variables and well identity, and hybrid models receiving the physics estimate as an input. Over the held-out test period the physics-embedded network reached a mean absolute error of 2.58 bar, a mean absolute percentage error of 1.08 per cent, a bias of ?0.35 bar, and a coefficient of determination of 0.97, without requiring simulator output at inference. The physics baseline overpredicted systematically, while purely data-driven learning proved unreliable in one well despite a low overall error. Learned factors remained near unity, with the gravity term corrected most strongly. Embedding the gradient computation preserves the physical structure of the traverse while allowing field data to refine it.