基于深度置信网络的总有机碳含量预测方法

Prediction method of total organic carbon content based on deep belief nets

页岩总有机碳含量(TOC)是页岩气层的重要评价参数.前人已经发展出多种TOC预测方法,其中BP神经网络方法的应用较多且效果较优,但是BP神经网络的权值随机初始化使误差收敛容易陷入局部最小或产生过拟合.本文针对页岩连续总有机碳含量的获取,构建了基于测井参数的深度置信网络(Deep Belief Nets,DBN)模型,应用于川东南丁山地区HF1井和HF3井上奥陶统五峰组—下志留统龙马溪组下部页岩气层段TOC含量的预测中,并与BP神经网络模型的预测结果进行对比分析.结果表明,基于测井数据的深度置信网络模型相较于BP神经网络模型而言,有效的减少了陷入局部最小值与产生过拟合等现象,总体样本的实测值与预测值之间的相关系数提高了0.0282,均方根误差降低了0.2142,具有更高的预测准确率和更好的泛化能力.

The Total Organic Carbon content (TOC)of shale is an important evaluation parameter of shale gas.Predecessors have developed a variety of TOC prediction methods.Among them,the BP neural network method has more applications and better results. However,BP neural network random initialization is easy to fall into the local minimum or produce over-fitting.In this paper,Deep Belief Nets model based on logging data is built,which is applied to the prediction of TOC content in the shale gas section of Upper Ordovician Wufeng and the Lower Silurian Longmaxi in the HFI and HF3 wells in the Dingshan area of southeastern Sichuan,and compared with the prediction results of BP neural network model. The results show that the total organic carbon content prediction results of the deep belief network based on well logging data can effectively reduce the phenomenon of falling into local minimum and over-fitting compared with the prediction results of total organic carbon content based on BP neural network,the correlation coefficient between the measured value and the predicted value of the overall sample is increased by 0.0282,the root mean square error is reduced by 0.2142,with higher predictive accuracy and better generalization.