选择性集成学习模型在岩性-孔隙度预测中的应用

Application of Selective Ensemble Learning Model in Lithology-porosity Prediction

储层是油藏地质建模的主要对象,储层属性参数的预测是建模的重要基础和主要难点之一。利用机器学习方法建立预测模型是目前研究的一个热点。针对单一机器学习方法在孔隙度预测方面存在的容错率低、过拟合等缺点,提出了融合岩性分类进行选择性集成学习建立预测模型的方法。该方法首先使用支持向量机进行岩性分类,并将岩性分类结果作为孔隙度选择性集成预测模型的输入。然后在研究分析典型机器学习方法的基础上,通过主成分方法分析法从支持向量回归、径向基(radial basis function,RBF)神经网络、随机森林、岭回归和K近邻回归等经典模型中选择出一组表现优异的个体学习模型组成集成学习模型,个体在集成模型中的权重由“主成分权重平均”法获得,最终采用加权平均法得到集成学习模型的输出。该方法考虑了岩性对孔隙度的影响,克服了单一模型存在的不足,模型的泛化能力强。研究结果表明,该方法的预测精度明显优于其他单一机器学习方法,适应性好。

Reservoir is the main object of reservoir geological modeling.The prediction of reservoir attribute parameters is one of the important foundations and main difficulties of modeling.Using machine learning methods to build predictive models is a hot topic at present.The problem of low probability of fault tolerance and over-fitting in the prediction of porosity in a single machine learning method,a method of constructing a predictive model by ensemble selection learning with lithological classification is proposed.First,the method used support vector machine for lithology classification and used the lithological classification results as input to the porosity ensemble selection prediction model.Then,based on the research and analysis of typical machine learning methods,a group of excellent individual learning models were selected from the classic models to form an ensemble learning model by the“principal component method analysis”.The classic models included support vector regression,radial basis function(RBF)neural network,random forest,ridge regression and K-nearest neighbor regression The weight of the individual in the ensemble model was obtained by the“principal component weighted average”.Finally,the output of the ensemble learning model was obtained by the“principal component weights average”.The method considers the influence of lithology on porosity,and overcomes the shortcomings of single model.The model s generalization ability is strong.The research results show that the prediction accuracy of this method is obviously better than other single machine learning methods,and the adaptability is good.