Publications

Machine learning has been widely used in oil and gas industry for exploration and production. This method efficiently increases productivity, reduces risk, minimizes operating costs, and maximizes return on investment (ROI). This paper uses multiple supervised and unsupervised machine learning algorithms to identify infill well locations to enhance oil production. The use of machine learning is proven to be much faster, requires shorter time, and more cost-effective compared to conventional reservoir simulations, which require longer time to generate multiple scenarios, and have high computational costs. This study used 83 well data that contain reservoir and production properties. The reservoir dataset comprises porosity, permeability, lithology proportion, and six geomechanical attributes. The production dataset consists of cumulative oil and water production. The dataset was preprocessed by removing the outliers and imputing null values using the k-nearest neighbor imputer. Using an unsupervised ¬¬k-means clustering algorithm, the wells were clustered based on the cumulative production data. The clusters were ranked based on hydrocarbon and water productivity to determine production quality. Generally, high hydrocarbon and low water production wells have good production quality. After that, for each location in the field, all reservoir properties were interpolated using Gaussian kriging. Lastly, the extra trees classifier was modeled to correlate the reservoir properties with the expected production quality. This classifier model can predict the production quality for each location and the recommended perforation depth. For the well production quality labeling, the unsupervised model could show three optimum numbers of clusters indicated by its relatively high silhouette score (0.432) and an inflection point on the inertia (at 5.733). By analyzing the productivity, the model clustered 24.7%, 17.8%, and 57.5% of the production wells as good, fair, and poor quality, respectively. The supervised model can predict well production quality with precision, accuracy, AUC-ROC (area under the receiving operating characteristic curve), and F1 scores of 0.9231, 0.8421, 0.9458, and 0.8256, respectively. Hence, the overall model is suitable for predicting the sweet infill well locations. In the future, this research can be expanded by adding additional data, such as dynamic reservoir properties, to the built machine learning model. These data are needed for calculating the remaining reserves in unexploited areas to be recommended for the following infill well location.