基于LSTM循环神经网络的横波预测方法

Shear wave prediction method based on LSTM recurrent neural network

周恒,武中原,张欣,张春雷,马乔雨

中国地质大学（北京）数理学院，北京 100083 北京师范大学统计学院，北京 100875 北京中地润德石油科技有限公司，北京 100083

School of Science, China University of Geosciences, Beijing 100083, China School of Statistics, Beijing Normal University, Beijing 100875, China Beijing Zhongdirunde Petroleum Technology Co. Ltd., Beijing 100083, China

针对碳酸盐岩储层岩性多样、孔隙结构复杂导致传统横波预测方法受限的问题，文中提出利用长短时记忆神经网络（LSTM）预测复杂碳酸盐岩储层的横波时差。相对于传统的简单点对点学习模式，LSTM通过复用神经元结构，有效学习测井参数的序列信息。以苏里格气田苏东地区碳酸盐岩储层为例，选择声波时差、密度、自然伽马等16个对横波速度较为敏感的测井参数，构建了基于LSTM的横波预测模型。和机器学习方法（Bayes，BP，DT，KNN，LR，SVM）以及Xu-Payne岩石物理模型方法相比，基于LSTM的预测模型均方根误差仅为3.36 μs/m，决定系数达到0.967，表明基于LSTM的横波预测模型更加符合实际地质情况，在复杂碳酸盐岩储层的研究中具有广阔的应用前景。

Aiming at limitations of traditional shear wave velocity prediction method due to the diversified lithology and complex pore structure of carbonate reservoirs,  using long-short-term memory neural network （LSTM） to predict shear wave time difference of complex carbonate reservoirs was proposed. Compared with the traditional simple point-to-point learning mode, the LSTM method can fully learn the sequence information of the logging parameters by reusing the neuron structures. Taking the carbonate reservoir in Sudong area of Sulige gas field as an example, 16 logging parameters sensitive to shear wave velocity, such as AC, DEN, and GR, were selected to construct shear wave prediction model based on LSTM. Compared with traditional machine learning methods （Bayes, BP, DT, KNN, LR, SVM） and Xu-Payne rock physics model, the root mean square error of the prediction model based on LSTM is only 3.36 μs/m, and the coefficient of determination reaches 0.967. It shows that the shear wave prediction model of LSTM is more in line with the actual geological distribution, which has broad application prospects in the research of complex carbonate reservoirs.

10.6056/dkyqt202106020