A machine learning-based predictive model for real-time monitoring of flowing bottom-hole pressure of gas wells

Abstract

Prediction of the flowing bottom hole pressure (FBHP) of gas–water two-phase flows is of great importance in optimising the production of gas and reducing down-time in unconventional gas wells. Unlike the case for conventional gas wells, prediction of FBHP for unconventional gas wells, particularly coal seam gas (CSG) wells, has not been studied.Monitoring of FBHP typically is done using a downhole pressure sensor placed close to the bottom of the well. Replacing a failed pressure sensor or recalibration of a pressure gauge, which is required frequently throughout the life of a well for reliable measurement of FBHP, requires interruption of the gas production at a high cost. A low-cost and reliable model for continuous prediction of FBHP, would enable smooth operation of CSG wells in the event a pressure gauge fails, without interrupting the well production.This work presents predictive models for real-time and reliable prediction of FBHP using surface and subsurface data, acquired from 91 CSG wells in Australia over 5–19 month production periods. Two sets of models are developed; one for specific wells using data from that individual well and another one for a group of wells. Three different modelling approaches, multiple linear regression, linear mixed-effects and gradient boosting regression tree (XGBoost) are implemented. The XGBoost modelling outcomes show promising results with the best mean absolute percentage error (MAPE) of 10% and 11.7% for specific well models and multiple well models, respectively.