Integrated reservoir simulation and machine learning for enhanced reservoir characterization and performance prediction

This thesis presents a comprehensive study on reservoir simulation and machine learning techniques for improved understanding and prediction of reservoir behavior. The research focuses on the Sarir C-Main field and utilizes various data sources including seismic cubes, well logs, base maps, check shot data, and production history. The methodology involves the development of static and dynamic models through processes such as data quality control, log interpretation, seismic interpretation, horizon and surface interpretation, fault interpretation, gridding, domain conversion, property and petrophysical modeling. Additionally, well completion, fluid model definition, and rock physics functions are established. History matching and prediction are performed using simulation cases, and machine learning techniques including data gathering, cleaning, dynamic time warping (DTW), long short-term memory (LSTM), and transfer learning are applied. The results obtained through Petrel simulation demonstrate the effectiveness of depletion strategy, history matching, and completion in capturing reservoir behavior. Furthermore, machine learning techniques, specifically DTW and LSTM, exhibit promising results in predicting oil production. The study concluded that machine learning approaches, such as the LSTM model, offer distinct advantages. They require significantly less time and can yield reliable predictions. By leveraging the power of transfer learning, accurate predictions can be achieved efficiently when limited data are available, offering a more streamlined and practical alternative to traditional reservoir simulation methods.