Authors: Diego Alberto Junca Rivera, Julian Ricardo Bohorquez Gutierrez, Evgeniya Dontsova, John Estrada Giraldo, Jesus Martinez Ferreira, and Jerry Webb, Enovate Upstream
Abstract
Using multidimensional analysis of historical drilling parameters combined with deep learning (DL) techniques, consistent ROP improvement in different drilling environments can be achieved. This discussion is focused on how offset well data can be properly analyzed and modeled to generate valuable outputs that improve not only drilling performance but also sustainability across the entire upstream. The workflow starts by history-matching comparable wells based on different variables such as well shape, hole size, wellbore design, downhole tools, etc. This first step is heavily based on the collaborative effect between subject matter experts working closely with data scientists to ensure only the right variables that are known to effect ROP are used and that a stable and scalable model can be acheived for different drilling environments. The process then moves to the data scientists where using multilayer perceptron models and random forest techniques allow determination of the ranking of features that affect ROP the most. The top tier features are then used to train a machine learning (ML) model to determine the average threshold of historic performance. Once the threshold is known, ML is again used to determine the optimal combination of drilling parameters that yield above-average ROP performance. This process is then repeated for each formation type and hole size. The performance range of the historical offset well data is then reviewed to determine the recommended threshold and fl score to output the highest modeling performance and the output parameter recommendations are then uploaded on a dashboard for real time guidance.