合肥工业大学校徽

合肥工业大学学报自科版

合肥工业大学学报(自然科学版)

JOURNAL OF HEFEI UNIVERSITY OF TECHNOLOGY (NATURAL SCIENCE)

主管：中华人民共和国教育部

主办：合肥工业大学

ISSN 1003-5060, CN 34-1083/N

基于卷积神经网络的压裂停泵数据滤波方法

A filtering method of fracturing pump-stop data based on convolutional neural network

期刊信息

合肥工业大学（自然科学版），2023年12月，第46卷第12期：1717-1721

DOI: 10.3969/j.issn.1003-5060.2023.12.020

作者信息

徐靓，李道伦，查文舒

（合肥工业大学数学学院，安徽合肥230601）

摘要和关键词

摘要: 水力压裂停泵时, 由于水击效应的影响, 井口处压力会产生波动, 此时测量得到的压力数据不能反映真实的渗流情况, 不利于试井解释, 需要对压裂停泵后的压力数据进行数据滤波。针对该问题, 文章提出一种基于卷积神经网络的滤波方法: 基于方差思想, 构造可表征数据离散程度的损失函数, 用以约束滤波后数据的离散程度; 基于停泵压力数据的物理特征, 给出可保留停泵压力最高点的损失函数表征方法; 在此基础上, 构造一个9层的卷积神经网络模型, 使得神经网络能够合理完成数据滤波。对于水平井多段压裂停泵压力数据的样本, 只需进行1次合适的参数调整, 同一网络框架可以完成对以上不同段压力数据的处理, 且实验效果好; 以均值和相对误差为参照标准, 对比滤波前后的压力数据, 滤波后的数据与原始数据相比, 离散程度变化较小, 且与原有数据相比仍能保持较小的相对误差, 基本稳定在0.2%左右。该滤波方法为压裂效果评价提供了可靠数据, 具有广阔的应用前景。

关键词: 数据滤波；神经网络；自定义损失函数；一维卷积；水击效应；水力压裂

Authors

XU Liang, LI Daolun, ZHA Wenshu

(School of Mathematics, Hefei University of Technology, Hefei 230009, China)

Abstract and Keywords

Abstract: With the hydraulic fracturing pump stop, the pressure at the wellhead will fluctuate due to the water hammer effect. At this time, the measured pressure data cannot reflect the real seepage situation, which is not conducive to well test interpretation. Therefore, it is necessary to filter the pressure data after the fracturing pump stop. Aiming at this phenomenon, this paper proposes a convolutional neural network-based filtering method. Based on the idea of variance, a loss function is constructed to characterize the dispersion of the data to constrain the dispersion of the filtered data; based on the physical characteristics of the stopping pressure data, a loss function characterization method is given to retain the highest point of the stopping pressure. Based on this, a nine-layer convolutional neural network model is constructed to enable the neural network to reasonably complete the data filtering. For the pressure data samples of multi-stage fracturing in horizontal wells, only one appropriate parameter adjustment is needed, and the same network framework can complete the processing of the above pressure data of different stages, and the experimental effect is good. Taking the mean value and relative error as the reference standard, and comparing the pressure data before and after filtering, it is found that compared with the original data, the dispersion degree of the filtered data has less variation, and the relative error can still be kept at about 0.2%. This filtering method provides reliable data for fracturing effect evaluation and has a broad application prospect.

Keywords: data filtering; neural network; custom loss function; one-dimensional convolution; water hammer effect; hydraulic fracturing

基金信息

国家自然科学基金资助项目（12172115）

前一篇后一篇

返回期刊目录