An explainable experience-driven hybrid model for TOC prediction in shale reservoirs based on data augmentation

Abstract

Total Organic Carbon (TOC) content is a measure of the carbon content in organic compounds, commonly used as a critical indicator for assessing unconventional shale resources. Therefore, an accurate TOC prediction model can help evaluate the reservoir's hydrocarbon potential at a low cost and improve the development efficiency. However, the sparsity of experimental data and the high heterogeneity of reservoirs present challenges for TOC prediction. This study proposes combining data enhancement techniques and expert experience-driven machine learning models for accurate TOC prediction in complex shale reservoirs. Firstly, we propose a set of data enhancement methods to address the problems of weak logging response and insufficient TOC experimental data. We enrich the training dataset by introducing reconstruction curves to visualize the response and designing Generative Adversarial Network (GAN) simulations to generate high-quality data. In the experience-driven model construction, we optimized the traditional ΔlogR method by integrating expert knowledge and a detailed analysis of the physical properties of shale reservoirs. We proposed a density-gamma modified ΔlogR method as the core of the experience-driven approach. Furthermore, we integrated the empirical formula into the fitness function of the Grey Wolf Optimizer (GWO). We combined it with a Support Vector Regression (SVR) model to build a hybrid model. The hybrid method was tested in the Dongying Depression. The R² values for wells A and B were 0.95 and 0.97, with Root Mean Square Error (RMSE) values of 0.31 and 0.29, and Mean Absolute Error (MAE) values below 0.3. The prediction results demonstrated significant improvement over any single method. We also analyzed the correlation between well logging curves and prediction results using the SHapley Additive exPlanations (SHAP) method. By revealing the decision-making mechanism within the model, we verified the reasonableness of the experience-driven and enhanced the model's credibility.