Unconventional oil resources are essential for global energy security and low-carbon transition, particularly before renewable energy technologies fully mature. Sandwich-type unconventional oil accumulation (SUOA) systems, a common hydrocarbon accumulation type in petroliferous basins, hold substantial resource potential to meet rising energy demand. However, SUOA exploration faces challenges from complex geological heterogeneity and technical uncertainties, necessitating reliable predictive methods to mitigate costly drilling risks. Using a representative SUOA system in China’s Ordos Basin, this study integrates systematic main controlling factor screening via multi-criteria ranking and quantitative evaluation to develop a novel heterogeneous ensemble prediction framework. The framework adopts a two-step “classificationthen-regression” strategy, incorporating a heterogeneous ensemble classification model (HECM) for oil charging identification and a heterogeneous ensemble regression model (HERM) for predicting cumulative oil-layer thickness (COLT), a proxy for resource volume. Results demonstrate that the proposed model significantly outperforms homogeneous ensemble methods and single-type machine learning models in both charging identification and COLT prediction, with predictions closely matching actual exploration outcomes. Feature importance and sensitivity analyses indicate that tight reservoir properties, source rock quality, and excess pressure jointly control oil charging, whereas COLT is primarily governed by macroscopic tight reservoir attributes (e.g., sandstone percentage, thickness, and fracture density). Notably, uncertainties from main controlling factor quantification/screening and ensemble component selection can affect model accuracy. This prediction framework provides a robust tool to reduce costly drilling errors in SUOA exploration and offers valuable insights for digital-intelligent oilfield technologies in the context of sustainable energy development.