A Sweet-spot Prediction Method for Hydrocarbon Reservoirs Based on an Interpretable Bayesian Hyperparameter Optimization-driven Local Cascade Ensemble Algorithm

Abstract Accurate evaluation of sweet-spot in hydrocarbon reservoirs is essential for effective reservoir prediction, conventionally reliant on well logs data or seismic inversion results. Scarce well-log data constrain the applicability of traditional methods in underexplored areas. Machine Learning (ML) offers a promising alternative by uncovering complex nonlinear relationships among diverse datasets, enabling sweet-spot prediction under limited data conditions. In this research, sensitive parameters—regarded as feature parameters in ML—are identified through intersection analysis and correlation analysis. A novel Local Cascade Ensemble algorithm is concerned, which integrates Random Forest and eXtreme Gradient Boosting Trees (XGBoost) into a cohesive predictive framework. Bayesian hyperparameter optimization is employed to enhance model efficiency and accuracy. To ensure interpretability, the SHapley Additive exPlanations (SHAP) method is applied to provide comprehensive explanations and evaluations of both the model and its feature parameters. Seismic inversion yields extensive 3D elastic parameter volumes, establishing a robust data foundation for spatially predicting sweet-spot distribution. Application results demonstrate the feasibility of the proposed approach, achieving a prediction accuracy exceeding 95%. This method provides reliable data support for future drilling deployment in underexplored areas.