Objectives: 
The aim of this research paper is to explore the potential of implementing a data-driven digital twin of an offshore Oil and Gas (O&G) facility's Power Generation System (PGS) and its associated Electrical Distribution System (EDS) for reducing emissions. This paper presents a methodology to forecast gas turbine fuel consumption and thermal efficiency through a machine learning (ML) workflow, using historical operational data from a set of dual gas turbines.

Methods:
The data-driven model utilized operational data from a process historian and bypassed traditional 1st principle-based evaluations to provide a more operationally responsive model. Different ML methodologies, such as Linear Regression, Random Forest, XgBoost, and LSTM Neural Network were applied for building the data-driven GT model. The model was trained on data spanning over a period of 5 years, sampled at 5-minute intervals. The final model could perform auto-tuning of various model parameters after a period of 1 year to automatically improve accuracy without the need for manual intervention.

Applications:
The implemented integrated Digital Twin acts as a real-time performance monitoring and scenario planning tool. The Real-Time Performance Monitoring workflow tracks and monitors the GTs' current overall operational performance characteristics against the ML model predicted performance, particularly in terms of fuel consumption and emissions. The Scenario Planning workflow allows users to configure hypothetical power/energy demand saving scenarios by adjusting varying electrical loads from the Electrical Distribution System and observing its impact on the subsequent fuel consumption and emissions calculations. This provides a better line of sight in reducing future emissions than simple annual cumulative GHG reporting.

Results:
A data-driven ML model was developed, representing the steady-state performance of the main power generator gas turbines. The ML model provided an output result as the predicted measure of fuel consumption, utilizing input features such as electrical load setpoint, ambient conditions, and internal gas turbine parameters. The modelled fuel consumption was used to estimate the overall thermal efficiencies of the gas turbines, along with GHG emissions calculations (including CO2, CH4, and N2O). Various ML models were considered during the model building phase, starting from Linear Regression, Random Forest, and XgBoost to more complex techniques such as Neural Networks (NN). An auto-tune model was subsequently developed to maintain the accuracy of model performance and avoid any model drift when the model suddenly encounters new/live data after deployment.

Novel/Additive information:
The implementation of a time-series forecast incorporating features such as seasonality and spatial analysis, along with operational parameters, proved to be a unique approach in predicting real-time performance characteristics and What-if scenario planning of the GTs, along with the complete Power Generation System. The method provides a framework and blueprint for any future data-driven digital twin development, particularly for complex proprietary systems with a multitude of data points. This data-driven method has the potential to significantly contribute to reducing GHG emissions in the offshore Oil and Gas industry.