Establishment and Comparison of Tight Oil Resource Abundance Prediction Models

Tight oil, as a typical unconventional oil and gas resource, has become a new exploration and development hotspot following shale gas. With the ongoing progress in tight oil exploration and development, efficient, convenient, and quantitative assessment of its resource volume is increasingly critical. This study takes 24 scaled regions worldwide as research objects and selects three major geological factors influencing tight oil resource abundance: oil generation intensity, area, and porosity. Multiple modeling approaches were applied, including a multiple nonlinear regression model (functional model), a multilayer neural network model (deep learning), a random forest model (supervised learning), and a gradient boosting tree model (ensemble learning). By comparing prediction accuracy, convergence, and generalization capability among the models, the gradient boosting tree model was found to perform best across all metrics, followed by the random forest model (which exhibited better generalization ability) and the multiple nonlinear regression model (which showed better convergence). The multilayer neural network model performed the worst overall. When applied to an example region, the results demonstrated that predictions from all models were close to the results of the China National Petroleum Corporation’s fourth oil and gas resource assessment, confirming their predictive value. Among them, the gradient boosting tree model yielded the best predictive performance. Based on combined prediction results and bias analysis, the gradient boosting tree model demonstrates high prediction accuracy and strong generalization capability, making it highly suitable for early quantitative estimation of tight oil resources in regions with low exploration maturity and insufficient geological data, thus providing valuable guidance for subsequent exploration and development.