基于机器学习方法的多采样点储层粒度剖面预测

Reservoir grain size profile prediction of multiple sampling points based on a machine learning method

地层砂的粒度特征值d50(筛析曲线累计质量分数50%对应的粒径值,μm)是防砂设计中的关键参数,为获得粒度纵向分布剖面,开展了基于机器学习方法的储层粒度与测井曲线响应关系研究。经典机器学习往往缺少模型内部的特征提取过程,而且采用单一采样点作为输入,缺失相邻数据关联关系反映层位信息。考虑到储层的地质连续性,利用测井曲线趋势和背景信息,将深度相邻数据点作为机器学习特征值,提出了一种基于多采样点的粒度剖面预测方法,构造和训练了基于随机森林(Random Forest)、支持向量机(Support Vector Machine)、Xtreme Gradient Boosting Tree(XGBoost)、人工神经网络(Artificial Neural Network)的预测模型。研究结果表明,与单点映射模型相比,考虑储层纵向地质连续性的各模型预测精度均高于单点预测,其中五点映射的ANN模型(ANN-5)预测效果最好,测试集d50预测相关系数最高为0.819,误差MAE最小为9.59,证实了多个采样点作为输入隐含利用了部分地层信息,有效地提高了预测精度。研究了特征点密度对模型准确率的影响,对训练集二维输入空间中样本的特征点高斯核密度分布以及测试集样本点处的训练集特征点密度进行估算,得出在高密度区域中的测试集样本点的RMSE普遍较低。当增加训练样本数量时,模型预测精度将进一步提高。采用层次分析法确定影响模型选择各因素的权重,通过模糊综合评判法优选机器学习模型,根据优选出的模型对临近区块储层粒度剖面进行预测,预测结果很好地捕捉了粒度变化趋势,模拟了其峰值。

The particle size characteristic(d50, the particle size value corresponding to 50% of the cumulative mass fraction of the sieve analysis curve, μm) of formation sand is a key parameter in sand control design. In order to obtain the vertical distribution profile of particle size, the response relationship between reservoir particle size and logging curve based on a machine learning method is studied. Classical machine learning often lacks a feature extraction process inside the model. Moreover,when a single sampling point is used as the input, the adjacent data association relationship is missing to reflect the horizon information. Considering the geological continuity of reservoirs, using the trend and background information of logging curves,taking the depth adjacent data points as machine learning eigenvalues, a grain size profile prediction method based on multiple sampling points is proposed. A prediction model based on random forest, support vector machine, Xtreme gradient boosting tree and artificial neural networks is constructed and trained. The results show that, compared with the single point mapping model, the prediction accuracy of each model considering the vertical geological continuity of reservoir is higher than that of single point prediction. The five point mapping ANN model(ANN-5) has the best prediction effect, with the highest correlation coefficient 0.819 and the least error measures 9.59 of the testing set. It is proved that multiple sampling points are used as input to implicitly utilize part of the stratum information and effectively improve the prediction accuracy. The influence of feature point density on the accuracy of the model is also studied. The Gaussian kernel density distribution of the feature points of the samples in the two-dimensional input space of the training set and the feature point density of the training set at the sample points of the test set are calculated. It is concluded that the RMSE of the sample points of the test set in the high-density area is generally low. The prediction accuracy of the model will be further improved as the number of training samples increases. AHP is used to determine the weight of each factor affecting the model selection, and fuzzy comprehensive evaluation is used to optimize the machine learning model. According to the optimized model, the grain size profile of the reservoir in adjacent blocks is predicted. The predictions capture well the trend of grain size change and simulate its peak value.