Machine-Learning and Physics-Based Models Compared in Downhole Pressure Prediction in Deepwater Reservoirs

Abstract

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of paper OTC 31758, “Downhole Pressure Prediction for Deepwater Gas Reservoirs Using Physics-Based and Machine-Learning Models,” by Jincong He, SPE, Matthew Avent, and Mathieu Muller, Chevron, et al. The paper has not been peer reviewed. Copyright 2022 Offshore Technology Conference. Reproduced by permission. Pressure measurements from permanent downhole gauges (PDHGs) during extended shut-ins often are used for model calibration and reserves estimation in deepwater gas reservoirs. A key challenge in practical operation has been the failure of PDHGs within the first few years of operation. To overcome this challenge, modeling solutions have been developed to enable accurate bottomhole well pressures to be calculated. In the complete paper, novel physics-based models and machine-learning (ML) models are presented and compared for estimating PDHG measurement from wellhead measurements. Introduction One way to predict PDHG shut-in pressure is with a full-physics wellbore simulator. However, those simulators usually require input parameters such as heat conductivity through the reservoir, itself highly uncertain in the field. In addition, wellbore simulation can be expensive computationally to run. The authors investigate two alternative modeling approaches to predict future PDHG shut-in data based on historical PDHG measurement from the same, or different, wells. The first approach is a physics-based data-driven (PBDD) method in which the temperature profile along the wellbore is characterized by a piecewise linear model, the cooldown process during shut-in is characterized by a decline-curve model, and the dependency of the cooldown on previous gas production is characterized by a linear model. During training, the parameters in these models are calibrated by historical PDHG shut-in data before gauge failure. Once trained, the model can be used to predict downhole measurement during future extended shut-in events. The second approach is to use ML methods. Using historical data, this approach trains a black-box model or function that directly projects the wellhead measurements and time since shut-in to downhole measurements. Both approaches are shown to be accurate when calibrated on historical data and used to predict future shut-ins of the same well. However, when they are used to predict shut-in pressure and temperatures of a new well based on other existing wells, the behavior of the ML model can be nonphysical and erratic, while the behavior of the PBDD model is more constrained and interpretable.