Fracture identification in shale reservoir using a deep learning method: Chang 7 reservoirs, Triassic Yanchang formation

Abstract

Natural fractures in shale oil formations play a significant role in fluid flow channels. However, their identification can be difficult due to the problems of subtle log responses of fractures and impact of sedimentary cycles on logging curve. To address these issues, a fracture identification method (FRNN) based on deep learning is proposed. FRNN combines window sliding and bidirectional recurrent neural networks (BiLSTM). A large number of sequence data samples from logs are obtained through window sliding, denoted with fracture labels based on rock core observations or borehole image log interpretation. BiLSTM processes sequences bidirectionally to mitigate the impact of sedimentary cycles on well logging curves and amply log responses caused by fractures. The integration of the sliding process and BiLSTM endeavors to address the challenge of fracture identification, which is often marked by significant uncertainty and subtle log responses. The experiments employ a dataset from Triassic Yanchang Formation's Chang 7 reservoirs in Q Oilfield, China, which contains shales and sandstones formed in braided river delta lacustrine systems that act as oil reservoirs. The statistical analysis and blind well test show that the fracture identification by the proposed FRNN method is consistent with those of core descriptions. Furthermore, examination of wellbore cross-sections reveals a congruence between the anticipated vertical distribution of fractures and the influencing factors thereof, thus validating the efficacy of the proposed methodology. The findings of this study can provide assistance in subsurface fracture modeling and hydraulic fracturing operations, etc.