Geomechanics Analysis from Horizontal Stress Prediction Using Supervised Machine Learning Algorithm: Case Study of AZI Wells, East Natuna, Indonesia

Abstract

This research goal is to predict Maximum Horizontal Stress (σH) and Minimum Horizontal Stress (σh) using borehole logs and machine learning algorithms in the carbonate reservoir. The machine learning algorithms include Ridge Regression (RR), K-Nearest Neighbor Regression (KNNR), Extra Trees Regression (XTR), and Gradient Boosting Regression (GBR). The purposes of this research are to determine the error of the training data, the worst algorithm with its disadvantages, as well as the best algorithm with its excellencies, and the influence of DEPTH (as a feature among other borehole logs) for the best algorithm. There are several methods plotted for prediction, namely initial preparation, preprocessing, algorithm modeling, development, validation, and best model evaluation. The data for this research consists of AZI-1 and AZI-2 Well as validation data, while AZI-3 to AZI-7 Well (except AZI-5) as training data. Results show that Coefficient of Determination (R2) and Root Mean Square Error (RMSE) of σH prediction for all algorithms are RR (0.69; 1.9E-3), KNNR (0.99; 6.7E-5), XTR (0.97; 4.9E-5), and GBR (0.98; 3.5E-5); while error scores (R2; RMSE) of σh prediction for all algorithms are RR (0.61; 9.6E-4), KNNR (0.98; 5.1E-5), XTR (0.99; 6.5E-5), and GBR (0.99; 3.7E-5). The worst algorithm is RR with disadvantages that include bias and underfitting in recognize the horizontal stress prediction pattern. Hereafter, the best algorithm is GBR which performs accuracy in training data prediction, gives the best performance in validation data prediction, and Absolute Percentage Error (APE) with Mean APE (MAPE) under 10%. This model can also forecast horizontal stress magnitudes from other wells in East Natuna effectively and accurately, compared with the usage of empirical correlation. As the best algorithm for horizontal stress prediction, GBR is developed using the DEPTH as one of the selected features. The new model of GBR without the DEPTH feature is resulting in a good prediction for horizontal stresses in AZI-1 Well but fails to predict the horizontal stresses in AZI-2 Well.