One of the key points in the construction of smart oil and gas fields is the effective utilization of data. Virtual Flow Metering (VFM), as one of the representative research directions for digital transformation, can obtain real-time production from oil and gas wells without the need for additional field instrumentation, utilizing pressure and temperature data obtained from sensors and employing multiphase flow mechanism models. The data-driven VFM demonstrates a commendable capacity in capturing the nonlinear relationship between sensor data and flow rates, while circumventing the necessity for rigorous analysis of the underlying mechanistic processes. However, this approach also faces the problem of poor model interpretability and uncertainty in the reliability of the output results. To enhance the reliability of data-driven models, this study proposes a hybrid model that integrates knowledge into the data-driven model. We added a constraint containing prior knowledge to the Long Short-Term Memory neural network to guide data-driven model training and established a Knowledge-Guided Predictive Model (KGPM) suitable for VFM. Through a series of comparative experimental analyses, our proposed model has demonstrated exceptional proficiency in flow rate prediction, with a Mean Absolute Percentage Error of 3.211% and 1.141% for the two experimental wells. This research contributes to the optimization of VFM techniques, making a significant contribution to the efficient construction of intelligent oil and gas fields.
© 2024 The Authors. Published by American Chemical Society.