The sandy conglomerate reservoirs of the W oil-bearing formation in A sag are characterized by medium-to-low porosity, low-to-ultra-low permeability, and strong heterogeneity. Their complex pore-throat structures pose significant challenges to accurate permeability prediction. To establish a refined prediction model suitable for the study area, this paper integrated core data, well logs, and mercury injection capillary pressure tests systematically analyze petrological and pore structure characteristics. On this basis, the application effects of six modeling approaches were comparatively evaluated: conventional porosity-permeability regression, reservoir grain size, sedimentary microfacies, pore-throat difference constraint, Flow Zone Index (FZI) flow unit, and Particle Swarm Optimization-Extreme Gradient Boosting (PSO-XGBoost) machine learning models. The results demonstrate that the conventional porosity-permeability regression model fails to effectively characterize the complex nonlinear features of the reservoir. The reservoir grain size model provides limited prediction accuracy because extensive diagenetic modifications, such as late-stage dissolution and cementation, have significantly weakened the correlation between current pore-throat structures and original sedimentary grain sizes. Although the sedimentary microfacies and pore-throat difference-constrained models improve accuracy to a certain extent, they are hindered by the qualitative nature of microfacies boundary definition and the difficulty of acquiring high-cost experimental data, respectively. Among the data-driven methods, the PSO-XGBoost model achieves the highest statistical precision; however, its application is restricted by ambiguous geological mechanisms and a heavy reliance on large datasets. In contrast, the FZI flow unit model effectively quantifies the strong heterogeneity by utilizing the Flow Zone Index to divide the macroscopically heterogeneous reservoir into evaluation units with uniform internal physical properties. Overall, the FZI flow unit model attains the optimal balance among predictive accuracy, data requirements, and geological mechanisms. It is confirmed as the most applicable permeability prediction method for the study area, providing reliable technical support for the exploration and development of low-permeability reservoirs in the Pearl River Mouth Basin and similar offshore regions.