基于机器学习的稠油油藏注蒸汽过程中汽窜识别研究

稠油油藏蒸汽吞吐过程中汽窜的产生与油藏地质和开发工程等因素有关，目前识别汽窜的方式局限于油藏工程、数值模拟等，此类方法无法准确判别各因素的不确定性和相关性，机器学习方法可识别海量数据间的隐含关系，准确度高且模型易维护。分析了汽窜的影响因素，构建基础数据集后对数据进行特征工程处理，包括数据重构、缺失值处理、维度转换及相似性分析，建立了汽窜预测特征属性集；采取Wrapper方法、Embedded方法、主成成分分析法对数据集进行降维处理，形成3套不同的特征组合方案；分别采用随机森林、SVM、神经网络和XGBoost算法建立汽窜预测模型，给出不同模型的预测准确率和预测汽窜通道分布。研究结果表明:注汽强度、层位渗透率极值和邻井距离对汽窜的影响程度最大，表现最好的组合模型是:PCA数据集+XGBoost模型，该方案在训练集上的准确率为97.20%，在验证集上的准确率为96.11%，实现了对汽窜的精准预警。

Machine learning-based steam channeling identification for steam injection of heavy oil reservoirs

Occurring of steam channeling during cyclic steam injection of heavy oil reservoirs is attributed to both geological and engineering factors. The current methods for identifying steam channeling are limited to the reservoir engineering approach and numerical simulation, which fail to capture the uncertainty and correlation between factors. Nevertheless, machine learning can recognize implicit correlations among massive data and has high accuracy and low maintenance. This research investigated the factors affecting steam channeling and performed the feature engineering processing after building the base dataset, including data reconstruction, dealing with missing values, dimension transformation and similarity analysis to build the feature attribute set for steam channeling. Subsequently, the dimensionality reduction of the dataset was carried out via the Wrapper method, Embedded method and principal component analysis to deliver three schemes of feature combinations. The steam channeling prediction models were built using the random forest, support vector machine (SVM), neural network, and XGBoost algorithms, respectively, of which the prediction accuracies and predicted steam channeling pathway distributions were presented. The research showed that the steam injection intensity, permeability extreme value of layers and well spacing have the largest influences on steam channeling. The data-algorithm combination with the best performance is the PCA dataset with the XGBoost model, which precisely predicts steam channeling with an accuracy of 97.20% for the training set and 96.11% for the validation set.