基于注意力机制和双向长短时记忆网络的横波速度预测方法及应用

Shear wave velocity prediction based on bidirectional long short-term memory networks with attention mechanism

横波速度信息对油气勘探而言至关重要,但实际测井资料中常常缺失横波速度资料。横波速度与测井参数之间存在非线性相关性,二者关系复杂难以用解析解表征。为此,提出了一种基于注意力机制和双向长短时记忆网络的横波速度预测方法(AT-BLSTM)。该方法首先利用注意力机制为测井参数分配权重,自动聚焦对横波速度预测贡献大的测井参数,然后利用双向长短时记忆网络以及横波速度曲线纵向上的时序特征,挖掘各种测井参数与横波速度之间的相关关系,获得各种测井参数与横波速度之间的学习模型,再输入优选测井参数,最终可直接获得横波速度的预测结果。将上述方法应用于挪威北海Volve油田和我国西南某工区的实际测井资料进行横波速度预测,并将预测结果与常规双向长短时记忆网络、门控循环神经网络以及基于经验公式的传统方法的预测结果进行对比。结果表明,利用基于注意力机制和双向长短时记忆网络的横波速度预测方法得到的测井参数权重分配合理,横波速度预测结果与实测横波速度误差较小、相关系数较高,有效提高了横波速度预测精度,预测结果具有良好的稳定性。

Shear wave information plays an important role in oil and gas reservoir exploration and development,but is often lacking in actual logging data.There is a relationship between reservoir parameters and shear wave velocity,but it is too complex to obtain analytical solutions.Considering the correlation between the shear wave velocity and other reservoir parameters,shear wave velocity prediction can be realized by deep learning.In this study,we propose a shear wave velocity prediction method based on the attention mechanism and bidirectional long short-term memory network(AT-BLSTM).First,the weight of logs is automatically assigned by the attention mechanism,and the logs that contribute the most to shear wave velocity prediction become the focus,whereas the logs with low sensitivity are ignored.In other words,human supervision can be avoided by using the proposed method when selecting features for shear wave prediction.Next,the influence of the upper and lower formation parameters on the shear velocity and the sequence characteristics of the longitudinal log data were fully considered to obtain the learning model that relates the shear velocity to the other parameters.Lastly,the optimal reservoir parameters can serve as the input in the learning model,and the prediction results of the shear wave velocity can thus be obtained.The proposed method was applied to the actual well logging data of the Volve oilfield on the Norwegian continental shelf and a land area in southwest China,and the results were compared with those obtained using the conventional gate recurrent unit neural network,bidirectional long short-term memory network(LSTN),and the conventional prediction method based on empirical formulas.The results showed that the error between the results predicted by the model and the measured shear wave velocity was smaller and the correlation coefficient was higher,demonstrating that the proposed method can effectively reduce the influence of manual characteristic curve selection and improve the shear wave prediction accuracy.