Abstract: Due to the strong heterogeneity of tight sandstone reservoirs, conventional log interpretation models usually overlook the difference in the longitudinal flow properties of the reservoirs, leading to low accuracy in the calculation and interpretation of permeability. Therefore, the flow units (FUs) of reservoirs were utilized to describe and quantify the heterogeneity characteristics of tight sandstone reservoirs. Permeability interpretation models of different FUs were constructed to improve the accuracy of permeability predictions. First, the FU classification standard was established by combining the frequency distribution and cumulative distribution probability of the experimental FU indicator (I_FZ) from cores, and the optimal classification number of FUs was selected to establish permeability models in different categories. Additionally, deep neural networks (DNN) were introduced to predict I_FZ by combining conventional logging and nuclear magnetic resonance (NMR) logging data. Finally, reservoir permeability was calculated based on the permeability interpretation models in different categories. This permeability calculation method was applied to the Paleogene Enping Formation in Huizhou Sag, Pearl River Mouth Basin, and five FU categories were selected for optimal classification. The identification and classification of FUs at the log scale were in good agreement with sedimentary facies, and the accuracy of permeability calculation was significantly improved compared to the NMR model. The research results provide a new calculation method for evaluating the permeability of deep tight sandstone reservoirs.