Abstract
This paper proposes a virtual metering (VM) system using machine learning and well production data for estimation of emulsion, produced gas, and vapors for steam-assisted gravity drainage (SAGD) production system. The system consists of four types of virtual meter models for every well: Emulsion Rate VM, Produced Gas Rate VM, Steam Vapor Rate VM, and Flash Vapor Rate VM. Emulsion Rate VM is used to estimate the oil rate output, while the other three VM types are used as the risk level indicators. For optimal well production and well management, operations need to consider both oil production and risks at the same time. Providing both well rate estimation and ability to assess risk at near-real time is a challenging task for large production fields, especially when the data is sparse and there is uncertainty in the well data. We propose a comprehensive system that utilizes a scalable framework to develop and deploy data-driven but process-informed virtual meters at scale and at near-real time. Specifically, in this paper, we describe a) a robust well data-preprocessing pipeline developed to continuously process the stream of input and target data for online learning. b) a state-of-the-art modeling framework designed to combine and segment large amounts of training data from the production filed. c) a scalable and flexible framework to ensure stable model performance with low data latency. d) a streamlined domain-specific model monitoring and MLOps process to enable tuning and retraining of the models by monitoring the performance and data drift.