Downhole Pressure Prediction for Deepwater Gas Reservoirs Using Physics-Based and Machine Learning Models

Abstract

Summary Pressure measurement from permanent downhole gauges (PDHGs) during extended shut ins (SIs) is a key piece of information that is often used for model calibration and reserve estimation in deepwater gas reservoirs. A key challenge in practical operation has been the failure of PDHGs within the first few years of operation. In this work, a physics-based data-driven (PBDD) model and machine learning (ML) models are developed to predict PDHG pressure and temperature measurement from the wellhead and other measurements during well SI events for deepwater dry-gas wells. During SI events, the wellbore cools down, resulting in increased gas density and bottomhole pressure (BHP). In the PBDD model, the temperature profile in the well is modeled with a piece-wise linear model as derived from wellbore simulations. The temperature decline during cooldown is captured using a decline-curve model, with the decline-curve parameters dependent on the location. The dependency of the cooldown effect on past production is captured with a linear model. Model parameters in the PBDD model are calibrated with data. In the ML models, multiple methods are tested, and the best performing method is picked based on cross-validation results. Two use cases are considered in this work. The first case (single well) involves predicting future SI BHP and temperature based on past PDHG measurement of the same well. Both the PBDD model and the ML model show good accuracy in blind tests for this use case. The second case involves predicting SI BHP and temperature of a well based on PDHG measurement of other wells. The PBDD model sees reduced accuracy in temperature prediction but is still reasonably accurate, while unphysical behavior is observed for the ML model even though the cross-validation score is high. It is concluded that, comparing the two types of models, the PBDD model is constrained by physics and thus the result is more interpretable and reasonable even when extrapolating. It also can provide the entire temperature and pressure profile during SIs. However, it does come with a series of assumptions (such as dry gas with no liquid content) and needs to be modified when the problem changes. On the other hand, the ML model is easier to construct and extend to other cases but is not bounded by physics so the result could be unphysical when extrapolation occurs.