One of the most important characteristic of naturally fractured rock for simulating the flow in the hydrocarbon reservoir and nuclear waste disposal is permeability. The two major approaches in modelling are: DFN method that consider the fractures permeable and omit the rock matrix conductivity, and the second method is dual porosity (DP) models. In the latter method the two overlaid media are considered: rock matrix and fractures. Till now just the numerical analysis are used to study dual porous media. In numerical analysis usually all parameters kept constant to observe the influence of changing the only one parameter. Using the artificial neural network and multivariate regression analysis can study the effect of all parameters simultaneously. According to the literature the simultaneous effect of the all geometrical parameters such as density, the macro fracture aperture, grain size and the micro aperture on permeability magnitude have not studied. In this study the 2D flow field in the fractured and permeable rock matrix is calculated using a distinct element code and the rock matrix is simulated by voronoi tessellation. 860 models of the synthetic fracture networks were generated based on different combinations of density and three types of correlation between fracture length and aperture take into account and voronoi size and micro fracture aperture. In order to propose the prediction model their different statistical and fractal characteristics were measured such as fractal dimension of intersection point, fluid flow channel (the area of macro fracture), the projection length of fractures in the direction of pressure gradient and the mean area of voronoi, and finally the flow channel between them. The result shows that using the principle component analysis the model is as strong as the linear regression model and also the collinearity problem was solved. The correlation between measured and predicted data is %80. Also the correlation coefficient between measured and predicted data of nonparametric regression using raw data is %79, and with normal data we can see the improvement of result. The correlation coefficient between measured and predicted data is %87. Compare to the both method of parametric and nonparametric regression the prediction capability of neural network is much better. The correlation coefficient of neural network is equal to %96 which shows the artificial neural network is more powerful in the case of distinguish of complicated relation between variables and also in prediction of target parameter.