基于PCA-BNN的页岩气压裂施工参数优化

A Study on the Optimization of Fracturing Operation Parameters Based on PCA-BNN

檀朝东 贺甲元 周彤 刘健康 宋文容

TANChaodong HEJiayuan ZHOUTong LIUJiankang SONGWenrong

中国石油大学(北京)油气资源与探测国家重点实验室, 北京 昌平 102249 中国石化石油勘探开发研究院, 北京 海淀 100083 北京雅丹石油技术开发有限公司, 北京 昌平 102200

State Key Laboratory of Petroleum Resource and Prospecting, China University of Petroleum, Changping, Beijing 102249, China Petroleum Exploration and Production Research Institute, SINOPEC, Haidian, Beijing 100083, China Beijing Yadan Petroleum Technology Development Co. Ltd., Changping, Beijing 102200, China

国内外学者在已有大量国外页岩压裂样本数据的前提下，开展了基于机器学习的页岩气压裂有效期预测及压裂参数优化的研究。随着近年来中国F气田不断地规模开发，积累了大量的压裂施工、生产动态、解释成果数据。通过利用已有的200口井的压裂施工历史数据及储层物性参数建立贝叶斯神经网络模型来优化压裂施工参数。选取对压裂效果有影响的储层物性参数、完井参数、压裂施工参数，用皮尔逊相关系数法分析11个参数的相关性；用主成分分析法（PCA）进一步降维处理，以降维后的主成分作为贝叶斯神经网络模型的输入参数，以压裂效果评价指标（有效期）为输出参数，引入贝叶斯方法自适应调整正则化系数避免神经网络过拟合，生成三层贝叶斯神经网络预测模型。用200口井中90%的井数据作为训练集，10%的井数据作为测试集，对该模型进行训练，实验结果表明，训练后该模型预测测试集的相对误差均值在5%以内，可以用来优化压裂施工参数。

Domestic and foreign scholars have carried out the research of shale gas fracturing production prediction and fracturing parameter optimization based on machine learning on the premise of a large number of foreign shale fracturing sample data. With the continuous development of F Gas Field in recent years, a large number of data from fracturing operation, production dynamic and interpretation results have been accumulated. In view of the fact that these data are not fully utilized in the design of fracturing operation parameters at present, the Bayesian neural network model is established to optimize the fracturing operation parameters by using history data of the fracturing operation parameters and reservoir physical parameters from 200 wells. The reservoir physical parameters, completion parameters and fracturing operation parameters which have an impact on the fracturing effect are selected. The correlation of 11 parameters is analyzed by Pearson correlation coefficient method. Principal Component Analysis (PCA) is used for further dimension reduction. The principal components are used as input parameters of Bayesian Neural Network model. The validity period is used as output parameter. Bayesian method is introduced to adjust regularization coefficient adaptively to avoid neural network overfitting. And then, a three-layer Bayesian neural network prediction model is generated. The model is trained by using 90% of the data of 200 wells as training set and 10% as test set. The experimental results show that the mean relative error of the model prediction results after training is within 5%, which can be used to optimize the fracturing operation parameters.

10.11885/j.issn.1674-5086.2020.05.12.05