Knowledge-Guided Machine Learning Method for Downhole Gauge Record Prediction in Deep Water Gas Field

Abstract A 1500-meter deepwater gas field in the South China Sea consisting of several wells has been installed with permanent downhole gauges (PDGs) to monitor downhole pressure. However, the high-temperature and high-pressure conditions pose a significant risk of PDG failures during the first few years of operation, which are prohibitively expensive to repair or replace. Therefore, it is essential to develop efficient models that can provide accurate estimations of the bottomhole pressure. In this paper, we proposed a method for predicting downhole pressure by integrating physics-based models with machine learning models. First, with the historical temperature and pressure data collected from the wellheads and the PDGs, a physics-based model was built on the basis of the wellbore multi-phase flow theory to depict the gas-liquid pipe flow behavior. Afterward, machine learning methods, XGBoost and Feedforward neural network were utilized to build data-driven model for predict bottomhole pressure. Finally, based on these two models, a knowledge-guided machine learning (KGML) model that integrates the physical knowledge and the data-driven model was established to predict the bottomhole pressure. The domain knowledge from the physics-based models was incorporated into the loss function as an adaptive weight such that the physical constraints were enforced during training. Based on the real-world data collected from two wells of the deepwater gas field in the South China Sea, a set of experiments were carried out to evaluate the proposed method for downhole pressure prediction. Compared to the pure data-driven model, the KGML model significantly reduces the distribution shift bias in pressure, improving the Mean Absolute Percentage Error (MAPE) of the pressure prediction by approximately 50%. Therefore, the KGML model can quantitatively describe the complex relationship between the wellhead temperature or pressure data and their downhole counterparts. Moreover, KGML exhibits robust performance across different well pads in the gas field, implying its extensibility to address the PDG failure challenge in a variety of deepwater gas wells. Consequently, the model can provide an efficient and economical approach for downhole measurements in deepwater regions.