Abstract：

Horizontal in-situ stress is the key basic parameter of wellbore stability analysis and hydraulic fracturing, but the geological environment of deep formations is complicated and hidden, which makes it difficult to predict the horizontal in-situstress accurately and quickly. Considering that the traditional logging interpretation and the neural network model cannot describe the spatial correlation between logging data and in-situ stress, a horizontal in-situ stress prediction method based on a Bidirectional Long Short-Term Memory neural network (BiLSTM) was proposed. Taking two vertical wells in the CL gas field in the Sichuan Basin as an example, two vertical wells were taken as the training well and test well respectively, and the nonlinear mapping relationship between logging parameters and in-situ stress was established through the training well, so as to realize the prediction of horizontal in-situ stress of the test well. Combined with the correlation of logging parameters and the actual geological meaning, the prediction effect of horizontal in-situ stress under different combination modes of logging parameters was investigated. The results indicated that: (1) Comparing the logging interpretation and core differential strain testing results, it is found that the logging interpretation error of vertical stress is 0.39%, the logging interpretation error of maximum horizontal in-situ stress is 0.18%~0.64%, and the logging interpretation error of minimum horizontal in-situ stress is 0.29%, which indicated that the logging interpretation is in good agreement with the actual in-situ stress. (2) The order of in-situ stress in the working area is vertical stress > maximum horizontal in-situ stress > minimum horizontal in-situ stress, which belongs to potential normal fault stress state. (3) There is a strong positive correlation between horizontal in-situ stress and true vertical depth (TVD), density (DEN), and natural gamma ray (GR), and a negative correlation between horizontal in-situ stress and interval transit time of P-wave (DTC), borehole diameter (CAL), compensated neutron (CNL) and interval transit time of S-wave (DTS). (4) Different combination modes of logging parameters have different prediction effects on horizontal in-situ stress, the optimal combination of logging parameters is TVD, CAL, DEN, CNL, GR, and DTC. (5) Orthogonal experiments are designed to optimize hyper parameters, and the average absolute percentage errors of maximum and minimum horizontal in-situ stress are 0.48‰ and 0.50‰, respectively. It is concluded that the BiLSTM model can effectively capture the variation trend of logging parameters with depth and the correlation information of logging parameters, and it can realize the accurate prediction of horizontal in-situ stress.

Key words：in-situ stress; horizontal in-situ stress; long short-term memory; BiLSTM; well logging

Received： 2022-05-17

Corresponding Authors： matianshou@126.com

Cite this article：马天寿, 向国富, 石榆帆, 桂俊川, 张东洋. 基于双向长短期记忆神经网络的水平地应力预测方法. 石油科学通报, 2022, 04: 487-504 MA Tianshou, XIANG Guofu, SHI Yufan, GUI Junchuan, ZHANG Dongyang. Horizontal in-situ stress prediction method based on the bidirectional long short-term memory neural network. Petroleum Science Bulletin, 2022, 04: 487-504.

URL：