Abstract: Channel sand is one of the important oil and gas reservoirs. The quantitative prediction of sand body thickness is the key to improve the efficiency of oil and gas development. With the enhancement of non-homogeneity of target reservoir, the relationship between seismic attribute and reservoir lithology, physical property and pore fluid becomes more complicated. How to achieve efficient and intelligent quantitative prediction of complex reservoirs under the condition of limited geological information is currently a hot and difficult topic in the field of reservoir prediction. To achieve high accuracy and intelligent prediction of the tight sandstone reservoir, a reservoir effective sand thickness prediction method based on deep fully connected neural network is proposed in this paper. The method constructs a multilayer stacked fully connected neural network to optimize the seismic attributes predicted for the effective sand thickness of the reservoir layer by layer, and maps the optimized attributes directly to the sand thickness. We first analyze the influence of training samples on fully connected neural network modeling, and then compare the performance of the deep and shallow morphology of this network in the case of small samples when the model size is larger than the number of training samples and the model size is smaller than the number of training samples, and find that the deep network outperforms the shallow one when the training samples are small, provided that the number of training samples is larger than the model size. Finally, we apply the deep fully connected neural network to the effective sand body thickness prediction from real data of Shengli oilfield, and the application results show that the method achieves effective identification of sand bodies around 4 m in a tight sandstone reservoir, reflecting the ability of this end-to-end intelligent modeling method to mine latent geological information from seismic attributes, thus confirming its effectiveness in quantitative reservoir prediction.

Key words: deep fully connected neural network, tight sandstone, reservoir parameters, seismic attributes, effective sand body thickness, small sample