摘要
多元线性回归等传统的储层孔隙度预测方法,通常难以有效捕捉测井数据的时空特征及掌握测井数据与孔隙度间复杂的非线性关系,导致预测结果与实测数据间存在较大误差。为此,引入卷积神经网络(CNN)-长短期记忆网络(LSTM)混合神经网络方法,阐述基于测井数据的CNN-LSTM混合神经网络储层孔隙度预测原理,采用互信息法筛选出与孔隙度相关性较强的声波时差、体积密度、补偿中子和自然伽马4个测井参数作为模型输入特征,构建CNNLSTM混合神经网络模型的预测流程,选取样本数据并划分训练集及测试集,通过缺失值处理、标准化和数据重塑对样本数据进行处理,最终建立了基于测井数据与储层孔隙度间非线性映射关系的孔隙度预测模型。模型的测试和评估结果表明:该混合神经网络模型在同井预测的误差指标MAE和RMSE分别降至0.2和0.25以下,比多元线性回归模型低60%以上,比循环神经网络模型低50%以上,比长短期记忆网络模型低40%以上。在未参与训练的C井的现场应用中,预测精度达到92.3%,应用效果良好。
Traditional reservoir porosity prediction methods,such as multiple linear regression,are usually difficult to capture the spatial-temporal characteristics of logging data and to grasp the complex nonlinear relationship between logging data and porosity,resulting in greater errors between predicted results and measured data.To address this,a hybrid neural network method of Convolutional Neural Network(CNN)—Long Short-Term Memory(LSTM)is introduced to expound the reservoir porosity prediction principle of CNN—LSTM hybrid neural network based on logging data.Four logging parameters with strong correlation to porosity,namely interval transit time,volume density,compensated neutron and natural gamma,were selected by mutual information method screening as the input features of the model to construct the prediction process of the CNN—LSTM hybrid neural network model.Sample data were selected and divided into train and test sets.The sample data were processed through missing value handling,standardization,and data reshaping,ultimately establishing a porosity prediction model based on the nonlinear mapping relationship between logging data and reservoir porosity.The test and assessment results of the model show that the hybrid neural network model reduces the error indices MAE and RMSE in the same well prediction to below 0.2 and 0.25respectively,representing more than 60%lower than multiple linear regression model,more than 50%lower than recurrent neural network model,and more than 40%lower than long short-term memory network model.In the field application of well C that did not participate in training,the prediction accuracy reached 92.3%,and the application effect was good.
作者
范伟
孙红华
杨丹
窦松江
程兴春
马丽杰
FAN Wei;SUN Honghua;YANG Dan;DOU Songjiang;CHENG Xingchun;MA Lijie(No.2 Mud Logging Company,BHDC,CNPC,Renqiu,Hebei 062550,China;Exploration Division/Geophysical Department,PetroChina Dagang Oilfield Company,Tianjin 300280,China;Exploration and Development Research Institute,PetroChina Dagang Oilfield Company,Tianjin 300280,China)
出处
《录井工程》
2025年第4期13-20,28,共9页
Mud Logging Engineering
