基于CNN-GRU神经网络的测井曲线预测方法

目前许多测井曲线预测模型存在预测结果不稳定、精度不高的问题。为此,将深度学习中特征表达能力较强的卷积神经网络(CNN)和记忆能力较强的门控循环单元(GRU)相结合,设计并实现了一种通过卷积门控循环单元(CNN-GRU)神经网络进行缺失井曲线预测的方法。以测井数据序列作为输入,首先通过CNN网络提取测井数据的特征,形成时序序列的特征向量,再利用GRU网络进行训练,最后输出测井曲线预测值。该方法综合了卷积神经网络局部特性感知和门控循环单元网络长期记忆的特性,考虑了测井曲线的深度趋势和局部形状,具有较高的预测精度。将该方法应用于四川某地区A、B两个井区3口井的测井曲线预测,并将预测结果与其它3种人工智能预测方法的预测结果进行对比分析,结果显示,基于CNN-GRU神经网络的测井曲线预测方法应用效果显著,能有效提取数据特征,为测井曲线预测提供了一种新思路。

Most conventional well logging prediction models are unstable,and the prediction of missing well curves is inaccurate.In this study,a method for missing well curve prediction through a convolution-gated cycle unit network was proposed.The method takes logging data sequence as the input.It first extracts logging data features through a convolutional neural network (CNN) to form the feature vectors of the time series,then trains them using a gated unit cycle network (GRU),and finally outputs the predicted values of logging curves.In other words,the proposed method uses a CNN-GRU integrated deep learning model,which integrates the perception of local characteristics typical of CNN and the long-term memory characteristics of the GRU.In this way,it can take into account both the trend with depth and the local shape of the logging curve.The method was applied to the prediction of logging curves of three wells in two exploration areas (A and B) in a certain area of Sichuan.The results were compared with those of other three artificial intelligence prediction methods,demonstrating the superiority of the proposed method.

国家自然科学基金(41774142)和国家科技重大专项(2016ZX05002-004-013)共同资助。