Yongliang Tang ¹ Dong Chen ¹ Hucheng Deng ² Fenglai Yang ¹ Haiyan Ding ¹ Yuyong Yang ^2* Cuili Wang ¹ Xiaofei Hu ² Naidong Chen ¹ Chuan Luo ² MING TANG ² Yu Du ²

• ¹ PetroChina Tarim Oilfield Company, Korla, Xinjiang, China
• ² Chengdu University of Technology, Chengdu, China

Fractures play a crucial role in tight sandstone gas reservoirs with low permeability and low effective porosity. If open, they not only significantly increase the permeability of the reservoir but also serve as channels connecting the storage space. Among numerous fracture identification methods, seismic data provide unique advantages for fracture identification owing to the provision of three-dimensional information between wells. How to accurately identify the development of fractures in geological bodies between wells using seismic data is a major challenge. In this study, a tight sandstone gas reservoir in the Kuqa Basin (China) was used as an example for identifying reservoir fractures using deep-learning-based method. First, a feasibility analysis is necessary. Intersection analysis between the fracture density and seismic attributes (the characteristics of frequency, amplitude, phase, and other aspects of seismic signals) indicates that there is a correlation between the two when the fracture density exceeds a certain degree. The development of fractures is closely related to the lithology and structure, indirectly affecting differences in seismic attributes. This indicates that the use of seismic attributes for fracture identification is feasible and reasonable. Subsequently, the effective fracture density data obtained from imaging logging were used as label data, and the optimized seismic attribute near the well data were used as feature data to construct a fracture identification sample dataset. Based on a feed-forward neural network algorithm combined with natural fracture density and effectiveness control factor constraints, a trained identification model was obtained. The identification model was applied to seismic multi-attribute data for the entire work area. Finally, the accuracy of the results from the training, testing, and validation datasets were used to determine the effectiveness of the method. The relationship between the fracture identification results and the location of the fractures in the target reservoir was used to determine the reasonableness of the results. The results indicate that there is a certain relationship between multiple seismic attributes and fracture development, which can be established using deep learning models. Furthermore, the deep-learning-based seismic data fracture identification method can effectively identify fractures in the three-dimensional space of reservoirs.