Robust computational approach to determine the safe mud weight window using well-log data from a large gas reservoir

Abstract

A key parameter when drilling for gas and oil is to determine the safe mud weight window (SMWW) to ensure wellbore stability as part of a quantitative risk assessment (QRA). This study determines SMWW by predicting acceptable upper and lower limits of the bottom hole pressure window during over-balance drilling method. A novel machine learning method is developed to predict SMWW from ten well-log input variables subject to feature selection. 3389 data records from three South Pars gas field (Iran) wells include data from: uncorrected spectral gamma ray; potassium; thorium; uranium; photoelectric absorption factor; neutron porosity; bulk formation density; corrected gamma ray adjusted for uranium content; shear-wave velocity and compressional-wave velocity. Combinations of these well logs are tuned to provide predictions of the SMWW, measured in terms of subsurface pore and fracture pressures, using machine learning (ML) algorithms hybridized with optimizers. The ML algorithms assessed are multiple layer extreme learning machine (MELM) and least squares support vector machine (LSSVM), hybridized with genetic (GA) and particle swarm (PSO) optimizers. This new algorithm (MELM) incorporates special features that improve its prediction performance, speeds up its training, inhibits overfitting and involves less optimization in the model's construction. By combining MELM with PSO, its optimum control parameters are rapidly determined. The results reveal that the MELM-PSO combination provides the highest SMWW prediction accuracy of four models evaluated. For the testing subset MELM-PSO achieves high prediction performance of pore pressure (RMSE = 12.76 psi; R2 = 0.9948) and fracture pressure (RMSE = 15.71 psi; R2 = 0.9967). Furthermore, the model demonstrates that once trained with data from a few wells, it can be successfully applied to predict unseen data in other South Pars gas field wells. The findings of this study can provide a better understanding of how ML methods can be applied to accurately predict SMWW.