Abstract
The industry remains focused on achieving efficiency gains through accessing and processing production, asset and original equipment manufacturer (OEM) data, and applying machine learning (ML) principles to arrive at improved outcomes. As a service provider, we are experiencing an increased activity level related to hybrid analytics which involves embedding high-fidelity physics-based models together with ML models to improve outcomes. Critically, rather than addressing one piece of equipment, or a subset of a facility, companies focus on economies of scale, and seek asset-wide understanding of implications to equipment condition when changing operating parameters.
Deployment of a Dynamic Digital Twin provides production and maintenance engineers with a ‘single source of truth’ for information (i.e. P&ID, PFD, OEM information, maintenance history, etc.) integrated with high-fidelity physics-based models for subsea and topsides processes. Field sensors measuring hydrocarbon quantity, quality and other physical properties are integrated to provide real-time and historic data. Physics-based dynamic process models are first calibrated to match the field sensor data and then used to generate synthetic data for training ML models. A high-fidelity model generates virtual measurements where field sensors are not available. Access to such high-quality virtual measurements presents a paradigm shift for upstream analytics, as ML algorithms now have access to larger datasets for training. This improves quality, allowing for proactive planning and improved uptime leading to increased facility uptime by predicting equipment failure and enabling condition-based maintenance (CBM).
In our work with major oil and gas operators, we have observed that maintenance engineers until now have struggled, because enough field sensors are not always available to support the ML algorithms, leading to less specific assumptions and lower quality results. By taking advantage of a Dynamic Digital Twin - containing the asset structure, visualization and models - hybrid analytics were applied to continuously improve predictions, thereby increasing facility uptime.
In this paper, we present a few case studies of applying hybrid analytics with some oil and gas operators to enable virtual flow metering, prediction of unplanned equipment shutdown and prediction of optimum operating parameters for increased facility uptime. Examples presented demonstrate the integration of historic and real-time measurements with the physics-based process and multiphase flow models, and ML algorithms such as Autoencoder (AE), Long Short-term Memory (LSTM) neural networks and Reinforcement Learning.
