基于双向循环神经网络的河流相储层预测方法及应用

河流相储层通常具有横向变化快、地震反射特征多解性强的特点,因而河流相储层地震预测难度大。将测井信息与地震多属性相结合实现河流相储层地震预测,传统的方法包括多元线性回归方法、地质统计学方法和BP神经网络等。人工智能深度学习方法为井震信息的融合提供了新的解决思路。通过构建井震学习样本,提出了一种基于双向循环神经网络的井震融合储层预测方法。从储层沉积连续性角度,将地震数据看成具有纵向联系的时序数据,以CD地区100余口井馆上段地层的储层和非储层为学习样本,构建双向循环神经网络储层预测方法,通过训练优选超参数建立井震融合的深度学习储层预测模型。该预测模型应用于CD地区河流相储层预测的效果显著,细小河道形态清楚,预测精度高,有效指导了CD地区的勘探部署。

Predicting fluvial reservoirs using seismic data based on a Bi-recurrent neural network

Fluvial reservoirs are difficult to identify from seismic data because the characteristics of sandbodies—including their response to seismic waves—are highly variable.The general approach for fluvial reservoir prediction entails a combination of logging information with seismic multi-attributes,which can be realized by multiple linear regression,geostatistical analysis,or back-propagation neural networks.Deep learning offers new solutions to the problem of information fusion of well and seismic data.In this paper,a new method of information fusion based on a Bi-recurrent neural network was proposed.The method uses selected samples of seismic and well data.From the perspective of reservoir sedimentary continuity,seismic data are considered as vertical time sequences.The super-parameters of the model were calibrated by training the model with reservoir and non-reservoir data of more than 100 wells in the CD area.The so-trained model succeeded in describing fluvial sandbodies in the CD area.In particular,the shape of small river channels was clearly imaged with high accuracy.Therefore,it can be concluded that the proposed method is effective for reservoir prediction in oil field exploration.