An Integrated Machine Learning Algorithm for Unconventional Flowing Bottomhole Pressure Prediction During Dynamic Gas Lift Operation

Abstract Machine learning algorithms have been used to predict flowing bottomhole pressure (FBHP), but past research often neglected data characterization, affecting accuracy. We developed an integrated ANN model by embedding and categorizing multiphase flow physics to improve FBHP predictions under varying conditions. We collected data from 16 deep shale oil wells in the Permian Basin, Texas. Inputs for the machine learning model were derived from multiphase flow physics analysis, including oil/gas/water gravities, injection depth, wellhead pressure and temperature, well temperature gradient, liquid flow rate, gas-liquid ratio, and water-oil ratio. The physics-based ANN model was created using a simulation dataset and field data, while the data-based ANN model was developed directly from the 16 wells. Both models predicted FBHP for four new wells over their entire period and one new well with early months’ data. We also explored the combination of an unsupervised clustering model with the physics- and data-based ANN model for FBHP prediction.