An improved algorithm with particle swarm optimization-extreme gradient boosting to predict the contents of pyrolytic hydrocarbons in source rocks

Abstract

Establishing a method to estimate the hydrocarbon-generation potential of source rocks driven by combinations of well logs is of major significance. The pyrolytic hydrocarbon value (S2) is an important pyrolysis parameter for source-rock evaluation; however, has received limited attention in predictive modeling. This study focused on optimizing ML algorithms to improve the accuracy of S2 predictions. A comparative analysis of ML regression algorithms, including extreme gradient boosting (XGBoost), random forest (RF), support vector regression (SVR), elastic net regularization (Elastic Net), convolutional neural network (CNN), and Bayesian ridge regression (BRR), was conducted to identify the most suitable method for predicting S2. Through confusion matrix analysis, a typical dataset for S2 prediction was established, which combined four well logs (spontaneous potential (SP), deep lateral log (LLD), acoustic transit time (AC), and shallow lateral log (LLS)) as the input variables. Data preprocessing steps, namely, normalization, moving average filtering homogenization, reduced dimension method t-distributed stochastic neighbor embedding (t-SNE), and density-based spatial clustering of application with noise (DBSCAN) reduction, were carried out to minimize errors between the abnormal well logs and the corresponding depth of the measured points. Among the evaluated algorithms, XGBoost exhibited excellent performance in predicting S2 values. Particle swarm optimization (PSO) and genetic algorithm (GA) were selected to optimize the hyperparameters of the XGBoost model. The improved algorithm (t-SNE + DBSCAN + PSO_XGBoost) effectively predicted S2 in the source rocks, serving as a powerful tool for hydrocarbon-generation potential evaluation.