Applying Machine Learning Methods to Identify Reservoir Features That Highly Affect Water Encroachment Based on Transforming Unsupervised to Supervised Model

Abstract Machine learning methods are increasingly applied to identify reservoir features that significantly impact water encroachment in heterogeneous oil reservoirs. This study applies machine learning to identify reservoir characteristics most responsible for rapid water encroachment, leading to an increasing water cut trend. In heterogeneous and naturally fractured reservoir, multiple sets of features contribute to higher water cut trends, governed by the areal and depth variations in rock and fluid properties. The study leverages machine learning’s capabilities to quickly detect these trends and link them to highly influential features. The approach begins with an unsupervised model that clusters wells based on their water cut trends, followed by supervised models that identify the reservoir features most significantly affecting these trends. Wells with similar water cut trends are grouped using time series clustering on oil producers, incorporating dynamic data such as oil, water, and total fluid rates; cumulative oil, water, and fluid production; water-oil ratio; water cut; bottom-hole pressure; and wellhead pressure. Additionally, data and information from water injection wells, including water injection rate and injection pressure, is used to enhance clustering based on areal patterns. To integrate dynamic and static parameters and establish their connections, a supervised model uses the generated clusters as target and static parameters as inputs. Input data includes matrix and natural fracture properties such as porosity, permeability, rock types, picks and fluid properties, all of which exhibit variations. Data from these static models are prepared using harmonic, geometric, and arithmetic averaging methods to ensure proper input for the supervised model. Classification and Regression Tree (CART) models are employed to determine the relative importance of each variable, identifying those with the highest positive or negative impact on water encroachment. Based on these variables and their associated weights, water influx contour maps are generated for each cluster. This novel approach not only identifies water encroachment trends in clustered areas but also pinpoints the static properties that most significantly impact these trends using supervised learning models. By utilizing CART, the relative importance of these properties per cluster is identified, allowing for the generation of a water encroachment map that reflect both water cut trend and the influence of key factors. The generated maps offer deeper insights into water management and surveillance within each cluster area of the reservoir. This methodology provides reservoir engineers with a powerful tool to better predict and manage water encroachment in complex, heterogeneous oil reservoirs.