Oil saturation is an important parameter for reservoir evaluation and hydrocarbon resource estimation. To address the limited prediction accuracy of oil saturation caused by the complex pore structure and strong heterogeneity of conglomerate reservoirs, an oil saturation prediction method based on a heterogeneous ensemble model is proposed. First, electrical parameters were established separately for sandstone, fine conglomerate, and coarse conglomerate using core experimental data. Oil saturation was then calculated using the Archie equation to provide physically meaningful target labels for machine learning model training. Subsequently, conventional well-logging parameters were selected as input features, and LightGBM, Back Propagation Neural Network (BPNN), General Regression Ne[ ]ural Network (GRNN), and Dilateformer were employed as expert models to construct a heterogeneous ensemble framework for collaborative prediction. The results show that the proposed method achieves a coefficient of determination (R2) of 0.92, outperforming both traditional empirical formulas and individual machine learning models. The proposed approach effectively mitigates the influence of reservoir heterogeneity on oil saturation prediction and improves both prediction accuracy and generalization performance. This study provides a reliable approach for oil saturation prediction in strongly heterogeneous conglomerate reservoirs.