Optimizing Reservoir Management Through Strategic Data Acquisition and Machine Learning-Enhanced Water Saturation Surveys

Data acquisition is a cornerstone of effective reservoir management, playing a critical role in understanding and optimizing the extraction of hydrocarbon resources. For reservoir management engineers, the process of acquiring data involves several types, each tailored to specific needs and conditions. These include production data, well logs, and seismic data, among others.

This data is utilized to construct a resilient proxy model utilizing machine learning algorithms, leading to precise predictions of field water production. Moreover, these runs are completed in a matter of seconds rather than days, unlike numerical commercial simulators, aiding decision-making during critical times in reservoir management.

In the context of managing water resources within a reservoir, water saturation surveys conducted in observation wells are invaluable. These surveys help delineate the movement of water across the field, particularly in layered reservoirs where water dynamics can be complex. By mapping the shape and flow of water, engineers can more effectively calculate displacement efficiency and tailor production strategies to maximize output and minimize water production. The traditional approach to conducting periodic saturation log surveys can be both time-consuming and costly, typically constrained by the limited budget allocated for such activities annually. To address these challenges, this work has utilized a machine learning approach to optimize survey efforts. In this context, a numerical reservoir model was considered for mimicking the measurements of the saturation by generating various scenarios using space-filling design of experiments, viz., Latin Hypercube Design (LHD). The gained data from the numerical simulations was used for developing the proxy model based on various machine learning approaches, including tree-based (CatBoost) and neural network schemes (MLP and CFNN). The obtained results demonstrated the robustness of the implemented proxy model for predicting water saturation profiles.

Validating the machine learning (ML) model against synthetic data generated by a black oil a compositional numerical model demonstrates a robust and accurate prediction of water breakthrough and water production in the field. This underscores the robustness of the experimental design used to orchestrate the numerical runs and the ML models developed in this investigation. The study highlights that the complexity of the ML algorithm escalates for highly intricate and heterogeneous reservoirs to enhance the accuracy of water prediction in the field. All ML algorithms performed effectively, with CatBoost proving particularly adept at capturing reservoir heterogeneity and complexity. The accuracy of the ML models has been assessed across various conditions by altering the bottom hole pressure and their effect on the different production rate during ten years water injection Our work flow demonstrate the successful application of our developed proxy model, resulting in reliable predictions of water production of producer wells. Additionally, computational time has been reduced by more than 10-fold compared to reservoir simulation runs.

This predictive power allows for strategic selection of wells for detailed logging, based on significant changes in water saturation, thereby potentially reducing survey costs by 30-50%. This innovative model represents a significant advance in reservoir management, offering a way to minimize time and financial expenditure while maximizing the effectiveness of saturation surveys.

Fahd Mohamad Alqahtani

Optimizing Reservoir Management Through Strategic Data Acquisition and Machine Learning-Enhanced Water Saturation Surveys

geologist, geology, prediction, numerical data, artificial intelligence, reservoir, water injection, proxy model, water saturation, accuracy