An optimized XGBoost method for predicting reservoir porosity using petrophysical logs

Abstract To overcome the deficiencies of current porosity prediction methods, the XGBoost algorithm is introduced to construct a model for porosity prediction, and the obtained model is optimized by the grid search method and genetic algorithm. First, the optimal values of the three integer hyperparameters and the ranges of optimal values for the five floating-point hyperparameters of the XGBoost algorithm are determined by the grid search method. Then, the optimal values of the five floating-point hyperparameters of the XGBoost algorithm are determined with the genetic algorithm based on the determined value ranges. In this way, the model for porosity prediction based on the XGBoost algorithm and optimized by the grid search method and genetic algorithm (GS-GA-XGBoost) is constructed, and it has eight hyperparameters with determined optimal values. Compared with other porosity prediction methods, our method solves the problems such as the strong subjectivity and poor generalizability of conventional logging interpretation methods and the insufficient generalization performance of machine learning methods in previous porosity prediction studies, and the accuracy of the constructed model for porosity prediction is also greatly improved. Specifically, the RMSE, MAE and MAPE generated by GS-GA-XGBoost on the test set are 0.527946, 0.155880 and 0.020500 respectively, while those generated by linear regression (LR), support vector regression (SVR), random forest (RF), the XGBoost algorithm with default parameters (XGBoost) and the XGBoost algorithm optimized only by the grid search method (GS-XGBoost) on the test set are 3.535521, 2.801047, 0.375713, and 2.695310, 2.002280, 0.283582, and 1.015801, 0.638878, 0.085942, and 2.781069, 1.860557, 0.293334, and 1.380065, 0.979419, 0.128486, respectively. Finally, the multithread technology is introduced to improve the computational speed of GS-GA-XGBoost. GS-GA-XGBoost provides some technical references for the construction of porosity prediction models for oil fields in Northern Shaanxi, China, and other regions.