Abstract: Conventional log interpretation models usually overlook the longitudinal flow properties due to the heterogeneity of tight sandstone reservoirs, leading to low accuracy in permeability interpretation. Therefore, the flow units (FU) are utilized to describe the heterogeneity characteristics of tight sandstone reservoirs. Constructing permeability models of different FU are expected to improve the evaluation accuracy of permeability. In this study, the flow unit classification standard is established by combining the frequency distribution histogram and cumulative probability plot of experimental Flow Zone Indicator (I_FZ), and the optimal number of flow unit is selected to establish multiple permeability models. Additionally, the deep neural network (DNN) is introduced to predict I_FZ by combining conventional logging and nuclear magnetic resonance (NMR) logging data. Finally, well permeability is calculated based on multiple permeability interpretation models. This permeability calculation method is applied to Paleogene Enping Formation in HuiZhou depression, Pearl River Mouth Basin, and five FU types are selected to optimal classification. The application result shows that the predicted FU types are well aligned with sedimentary facies and the accuracy of permeability calculations has significantly improved compared to NMR model. Overall, the permeability calculation method based on classification and optimization of FU provides a new insight for accurate evaluation of logging permeability in deep tight sandstone reservoirs.