摘要
孔隙度是评价储层质量的关键指标,实验测试获取连续的孔隙度数据十分昂贵,基于测井数据的准确预测对储层刻画至关重要。但传统的神经网络调参复杂且无法充分学习测井曲线与孔隙度之间复杂非线性关系的问题,笔者提出一种基于遗传算法(GA)优化的一维卷积神经网络(1D CNN)模型。首先,利用皮尔逊相关系数分析孔隙度与密度、声波时差、泥质含量、铀、钾等测井参数的相关性,结果表明相关系数分别为-0.80、0.72、-0.36、0.43和-0.34。在此基础上,构建GA-1D CNN模型,以川南X地区Y1井龙马溪组储层孔隙度预测为研究对象,并与传统CNN、GRU、LSTM和BP模型进行对比。结果显示:①GA优化提升了模型全局搜索能力,加速收敛速度,提高预测性能;②GA-1D CNN训练100轮后收敛,在训练集和测试集表现最佳;③Y1井测试集上,R²、MAE和RMSE分别为97.98%、0.1292和0.2948,优于其他模型。该方法降低了过拟合风险,在储层参数预测中展现出良好应用潜力。
Porosity is a key indicator for evaluating reservoir quality,and obtaining continuous porosity data through experimental testing is extremely expensive.Accurate prediction of porosity based on well log data is therefore crucial for reservoir characterization.However,traditional neural networks suffer from complex parameter tuning and are unable to fully learn the complex nonlinear relationship between well log curves and porosity.In this study,we propose a one-dimensional convolutional neural network(1D CNN)model optimized by a genetic algorithm(GA).First,we analyze the correlations between porosity and various well log parameters—density,acoustic time,shale content,uranium,and potassium—using the Pearson correlation coefficient,which are found to be-0.80,0.72,-0.36,0.43,and-0.34,respectively.Based on these results,the GA-1D CNN model is constructed for porosity prediction in the Longmaxi Formation shale reservoir of Well Y1 in the X area of southern Sichuan,and its performance is compared with that of conventional CNN,GRU,LSTM,and BP models.The results show that:(1)GA optimization enhances the model's global search capability,accelerates convergence,and improves prediction accuracy;(2)the GA-1D CNN converges after 100 training iterations,performing best on both the training and test sets;(3)on the test set of Well Y1,the R²,MAE,and RMSE are 97.98%,0.1292,and 0.2948,respectively,outperforming the other models.This method reduces the risk of overfitting and demonstrates great potential for application in reservoir parameter prediction.
作者
刘佳杰
徐川
解馨慧
李勇
曾杨帆
LIU Jiajie;XU Chuan;XIE Xinhui;LI Yong;ZENG Yangfan(College of Geophysics,Chengdu University of Technology,Chengdu 610059,China)
出处
《物探化探计算技术》
2026年第1期36-46,共11页
Computing Techniques For Geophysical and Geochemical Exploration
基金
2023年度国家资助博士后研究人员计划B档资助(GZB20230089)。
关键词
龙马溪组页岩
孔隙度预测
遗传算法
种群进化
1D卷积神经网络
Longmaxi Formation shale
porosity prediction
genetic algorithms
population evolution
1D convolutional neural network
