Data-driven recurrent neural network model to predict the rate of penetration

Abstract

The Rate of Penetration (ROP) is a vital parameter in drilling operations. Due to the complex relationship between the parameters affecting ROP, accurate prediction of ROP is hard to be obtained analytically. In this study, a recurrent neural network model was developed to estimate ROP using Plastic Viscosity (PV), Mud Weight (MW), flow rate (Q), Yield Point (YP), Revolutions per Minute (RPM), Weight on Bit (WOB), nozzles total flow area (TFA), and Uniaxial Compressive Strength (UCS). The data were collected from more than 2000 wells drilled worldwide. The network architecture was optimized by trial and error. The data were categorized into three sets; 70 % for training, 15 % for validation, and 15% for testing. The created network predicted ROP with an average R² of 0.94. With this tangible prediction method, oil and gas companies can better estimate the time of well delivery as well as optimizing ROP by altering the controllable input parameters affecting the ROP model. Artificial intelligent methods have shown their potential in solving complex problems. The oil and gas industry can benefit from artificial intelligence, especially with the large data sets available, to better optimize the drilling process.