Prediction of microscopic pore structure of tight reservoirs using convolutional neural network model
LIAO Guangzhi, LI Yuanzheng, XIAO Lizhi, QIN Zhijun, HU Xiangyang, HU Falong
1 State Key Laboratory of Petroleum Resources and Prospecting, China University of Petroleum-Beijing, Beijing 102249, China 2 International Joint Laboratory of Unconventional Oil and gas of Ministry of Education, China University of Petroleum-Beijing,Beijing 102249, China 3 Key Laboratory of Earth Prospecting and Information Technology, Beijing 102249, China 4 Exploration and Development Research Institute of Xinjiang Oilfield Company, PetroChina, Karamay 834000, China 5 Research Institute of Zhanjiang Company, CNOOC Limited, Zhanjiang 524057, China 6 PetroChina Research Institute of Petroleum Exploration & Development, Beijing 100083, China
Obtaining accurate information about the microscopic pore structure is very important for the exploration and development of low-porosity and low-permeability oil and gas reservoirs. This information is the basis for productivity prediction and dynamic reservoir simulation. At present, the commonly used pore structure characterization methods are all models based on physical experiments, such as mercury intrusion, thin section, nitrogen adsorption, and nuclear magnetic resonance. These methods are quite different and differ in terms of characterization mechanism, effective resolution and response range, which making it difficult to apply in downhole. Deep learning algorithms have great potential for application in the modeling and prediction of small sample data. In this paper, data mining algorithms such as grey relation analysis, principal component analysis, factor analysis,and intelligent clustering are used to perform deep mining of mercury injection capillary pressure data. The pore structure types of the study block are divided into five categories. Then, using conventional logging data as the input layer, a single-layer convolutional neural network and a double-layer convolutional neural network were used to predict the reservoir's micropore structure, and the training model was applied to the test well. The results show that the convolutional neural network can be used to predict the micro-pore structure of the reservoir. The double-layer convolutional neural network is better than the single-layer neural network model. Furthermore, deeper and more abstract reservoir features can be extracted through convolution operations.The prediction results are compared with the pore permeability characteristics reflected in the log interpretation, and the two are in high agreement. The double-layer convolutional neural network model can achieve a prediction accuracy of more than 80% on the test set. The research method provides a new idea for reservoir pore structure evaluation using core analysis data and logging data, and has important guiding significance for complex oil and gas exploration and development.
Key words：
pore structure; data mining; deep learning; convolutional neural network
