Using Supervised Machine Learning Algorithms to Predict BHA Walk Tendencies

Abstract Directional drillers have faced numerous challenges due to the right and left BHA walk tendencies. Walk, build, and drop tendencies increase sliding time thus reducing ROP and increasing CPF. Drilling tortuous wellbores are more prone to NPT, increased torque and downhole drag. It was noted that bit side forces and the tilt angle influence the DLS of the wellbore. In this study, novel predictive analytic models were developed to understand the factors that influence BHA assembly walk tendency as well as uncover the hidden relationships between several different features influencing the walk tendencies. Sixty-eight wells are included in this study with an initial model training and testing being executed on eight wells. A blind test was later performed on 57 wells with 149 different BHAs. The model accounted for the number and locations of the various components in the BHA and their different types. The modified BHAs are assumed to be a continuous beam supported by PDC bits, PDM, stabilizers, and an assembly, mirroring the contact points of the BHAs, and the wellbore. For simplification purposes, the assembly assumes that all three components are made of non-magnetic material with comparable OD, linear weight, and material. The assembly was based on the fact that these components had the same bending stiffness due to similar material and thus elasticity. Seven different ML models were experimented with to determine the lowest MAE. They included Gradient Boosting Machine, Random Forest, Artificial Neural Networks, and Adaboost. The attributes included mud density and formation type. Bit variables were composed of: OD, gauge length, length of inner and outer profile, TFA, manufacture, cutter size, and blade count. For PDM: location, OD, LW, length, bit to bend distance, and bend angle. The stabilizer included location, blade count, gauge length, gauge OD, LW, and stabilizer to bit depth and assembly specifications. Moreover, hole size, block height, hookload, WOB, ROP, differential pressure, mud flow rate, SPP, GR, Annular pressure loss, MSE, ECD at Bit, Bit RPM, Bit TOR, and bit aggressivity. The survey data had MD, Inclination, azimuth, and finally DLS. The models showed that the side forces in the form of seven dominant factors were the main culprits in influencing the walk direction of the drill bit. There was a highly significant relationship with a MAE of 14.7% between stabilizer location, gauge OD, PDM bit to bend distance, bit gauge, PDM differential pressure, ROP, WOB, inclination, and Hookload. These results prove to be a great advantage in controlling the drilling direction and reaching the target zone in minimal time. The optimized machine learning model helped optimize rotatory drilling time, ROP, smoother wellbores, and Lower CPF overall.