基于集成学习的致密气藏产能预测新方法

A novel approach for predicting production capacity of tight gas reservoirs based on ensemble learning

致密气藏产能预测是准确认识气藏、合理编制开发方案的重要基础。由于致密气藏储层物性差、非均质性强及压裂施工导致渗流规律复杂,其产能预测一直是项重大的挑战。利用集成学习思想,提出了将支持向量回归和AdaBoost回归通过堆叠模式耦合来进行产能预测的新方法,并以鄂尔多斯盆地M致密气区块为例进行了应用。通过最大互信息系数法筛选出M气藏产能主要影响因素为气层厚度、渗透率、孔隙度、含气饱和度、实际砂量、液氮量;将主要影响因素数据及相应无阻流量构成学习样本,利用孤立森林算法进行异常值检测,并通过K近邻填补缺失值;利用本文产能预测新方法,并通过贝叶斯优化算法进行了参数优化。随机选取20%数据进行盲测,并以决定系数为评价标准。测试结果表明,本文所提方法的决定系数R2值为0.936,预测精度相较单一算法有了明显提升。该方法有效结合了不同算法的优点,为致密气产能预测提供了一种新的思路。

Predicting the production capacity of tight gas reservoirs is a crucial foundation for the accurate understanding of the gas reservoir and the rational formulation of development plans.Due to the poor physical properties,strong heterogeneity of tight gas reservoirs,and complex seepage laws caused by fracturing operations,predicting the production capacity of tight gas reservoirs has always been a significant challenge.By utilizing the concept of ensemble learning,a novel approach was proposed to predict production capacity by coupling support vector regression and AdaBoost regression through a stacking model.This method was applied to the M tight gas area in the Ordos Basin.The main factors affecting the production capacity of the M gas reservoir were selected by using the maximal information coefficient method,including gas layer thickness,permeability,porosity,gas saturation,actual sand volume,liquid nitrogen volume.The data of the main influencing factors and the corresponding unobstructed flow rate formed the learning samples.The isolated forest algorithm was used for outlier detection,and missing values were filled using the K-nearest neighbor method.The proposed novel approach for predicting production capacity was used,and parameter optimization was performed by using the Bayesian optimization algorithm.A blind test was conducted on20%of the randomly selected data,which was evaluated based on the coefficient of determination.The test results show that the coefficient of determination R2 of the proposed method is 0.936,indicating a significant improvement in prediction accuracy compared to single algorithms.This method effectively combines the advantages of different algorithms,providing a new approach for predicting the production capacity of tight gas reservoirs.