A New Tool for Searching Sweet Spots by Using Gradient Boosting Decision Trees and Generative Adversarial Networks

Abstract

High-density completions prevail in shale oil formation in China due to the difficulty of identifying the sweet spot with high accuracy. Knowing the location of sweet spots benefits the horizontal well drilling and the selection of perforation clusters. Generally, field engineers determine sweet spots from the well logging interpretation. However, a group of prevalent classifiers based on gradient boosting decision trees were introduced to automatically determine sweet spots according to datasets from the well logging. Based on boosted tree algorithms, Extreme Gradient Boosting (XGBoost), Unbiased boosting with categorical features (CatBoost) and Light Gradient Boosting Machine (LightGBM) are utilized to control the over-fitting issues. Compared with linear support vector machines (SVMs) or kernel machine, these robust algorithms can deal with comparative scales of the features and learn non-linear decision boundaries via boosting. Moreover, they are less influenced by the presence of outliers. Another prevailing approach, named Generative Adversarial Networks (GANs), was implemented to augment the training dataset by using a small number of training samples. In terms of the training purpose, we randomly selected 60 horizontal wells. In each well, tens to hundreds of datasets of different formation intervals were collected. Features, such as resistivity, interval transit time, layer thickness, shale content, porosity, permeability, oil saturation, and coordinates in three dimensions, were extracted from well-logging datasets and regarded as inputs for classifiers. Datasets of remaining wells were used for testing. Compared with conventional SVMs, the prediction accuracies of sweet spots by XGBoost and CatBoost were significantly improved to 81.61% and 82.5%, respectively. Additionally, GANs, as an unsupervised machine learning tool, have been attempted to augment the dataset by utilizing a relatively small number of training samples. A generative model is used for capturing the data distribution, and a discriminative model aims at predicting a label to which that data created by the generative model. Without special pre-processing of the input datasets and fine tuning CTGANs model, the fake dataset could still bring 68.58% accuracy for all detections and 59.01% of the label corresponding with oil formation that showing its potential in data augmentation. This paper illustrates a new tool for categorizing the reservoir quality by using gradient boosting decision trees and GANs methods, which further helps search and identify sweet spots. An extensive application has been built for the field cases in a certain oilfield. This tool provides a guideline for covering more sweet spots during the drilling and completion treatment, which immensely decreases the exploration cost.