Porosity is one of the key indices to characterize the pore structure of dense sandstone reservoirs, and accurate prediction of porosity is important in reservoir evaluation. The accurate determination of porosity in tight sandstone needs to be based on coring well data, how to accurately predict the porosity of the whole well based on a very small amount of coring well data is a meaningful problem. To solve the problem of accurate prediction of porosity, a deep neural network (DNN) model is proposed to predict porosity accurately by using only a small amount of logging and coring data. The effective porosity curve obtained by logging interpretation in the dense sandstone reservoir of JY oilfield, Ordos Basin is taken as the training target. A mapping of complex and nonlinear relationship between logging curve and porosity was constructed based on a deep feed-forward neural network. After the neural network model was trained, it was used to the entire research target area, and porosity was obtained by means of inversion. This model can effectively improve the accuracy and reliability of reservoir porosity prediction. So a solid technical reserve is provided for petroleum geological exploration.