基于深度残差收缩网络的油气柱高度预测

Oil and Gas Column Height Prediction Based on Deep Residual Shrinkage Network

油藏的含油气柱高度在很大程度上反映了圈闭中油气层的丰富程度。石油含量高度的估计,对于钻前储量评价、井位优化部署等都有着一定重要性。为了提升油气柱高度预测精度,展开基于神经网络模型的油气柱高度预测方法的研究,并侧重于一维残差收缩网络的研究,因为一维的卷积核侧重对每一维特征的提取,更符合本实验数据的特性;其次模型使用了残差块,该模块使用链接跳跃方法来绕过输入信息直接输出来保护信息完整性,进而缓解梯度损失和网络退化问题;软阈值作为非线性变换层插入到深层结构中,以消除不重要的特征,来提高从高噪声数据中学习特征的能力。同时,为了验证模型的有效性,对目前应用较为广泛的模型,如CNN、1DCNN、GoogLeNet、DenseNet、1DRSN在圈闭数据上的应用进行了比较和分析。1DRSN预测准确率达到84.0%,优于其他模型,表明该模型对油气柱高度预测有更加准确的结果。

The height of the hydrocarbon-bearing column of the reservoir reflects the richness of the hydrocarbon layer in the trap to a large extent.Estimation of oil content height is of great importance for pre-drilling reserve evaluation and well location optimization deployment.In order to improve the prediction accuracy of the oil and gas column height,the research on the oil and gas column height prediction method based on the neural network model is carried out,and the research on the one-dimensional residual shrinkage network is focused.Because the one-dimensional convolution kernel focuses on the extraction of each dimension feature,which is more in line with the characteristics of the experimental data.Secondly,the model uses a residual block,which uses the link skip method to bypass the input information and directly output to protect the information integrity,thereby alleviating the problem of gradient loss and network degradation.Soft threshold as a nonlinear transformation Layers are inserted into deep structures to eliminate unimportant features to improve the ability to learn features from noisy data.At the same time,in order to verify the validity of the model,the applications of widely used models,such as CNN,1DCNN,GoogLeNet,DenseNet,and 1DRSN,on trapped data are compared and analyzed.The experimental results show that the prediction accuracy of 1DRSN reaches 84.0%,which is better than that of other models,indicating that the model has more accurate results for the prediction of oil and gas column height.