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

Abstract

Abstract Pressure measurement from permanent downhole gauges (PDHGs) during extended shut-ins is a key piece of information that is often used for model calibration and reserve estimate in deep-water gas reservoirs. A key challenge in practical operation has been the failure of permanent downhole gauges (PDHGs) within the first few years of operation. To overcome this challenge, innovative modeling solutions have been developed to enable accurate bottom-hole well pressures to be calculated. Both the physics-based model and machine learning model are developed to predict PDHG pressure and temperature measurement from the wellhead and other measurements during well shut-in events. These models are calibrated and blind-tested with data collected while the PDHG is still functioning. The key to the physics-based model is our model of the gas temperature profile within the wellbore and cool-down rate on shut-in, which has been inspired and validated by independent OLGA transient well simulations. In addition to physics-based models, machine learning (ML) models have also been developed, which directly perform regression on available data. The physics-based model is shown to be able to predict PDHG pressure and temperature accurately by capturing two key physics. Firstly, on well shut-in, a significant gas density gradient develops, as the wellhead (usually on the ocean floor) cools rapidly, while the bottom-hole temperature remains high. This changing temperature profile along the wellbore path has been accurately captured by a data-driven temperature model in the physics-based model. In addition, the decline of temperature along the wellbore has been observed to depend on the well's production history. With higher/longer production of hot gas through the wellbore and additional heating of wellbore surrounds, a slower cool-down rate is observed. A data-driven decline-curve model is devised within the physics-based model and has been shown to successfully capture this dependency. Compared to the physics-based model, the machine learning model is much simpler to devised. It also has the flexibility to incorporate new input features other than those that we can physically interpret. Multiple ML models are tested, and the random forest has shown the best performance. The accuracy of the ML model is comparable to that of the physics-based one. In this work, novel physics-based models and ML models are presented and compared for estimating PDHG measurement from wellhead measurements. The build-up pressures from PDHG during well shut-ins are the principal input for reservoir surveillance, analysis, and optimization (SA&O) activities, including material balance estimates of gas-in-place and reservoir simulation history matching.