基于GRU的汶川地震川西气井压力前兆异常识别研究  

Research on recognition of pressure precursor anomaly of gas well in Western Sichuan before Wenchuan earthquake based on GRU

作　　者：刘杰 曹俊兴[1,2] 蒋旭东 王俊[1,2] 熊玄辰[1,2] 周欣 LIU Jie;CAO JunXing;JIANG XuDong;WANG Jun;XIONG XuanChen;ZHOU Xin(State Key Laboratory of Oil and Gas Reservoir Geology and Exploitation,Chengdu University of Technology,Chengdu 610059,China;Geophysics School,Chengdu University of Technology,Chengdu 610059,China)

机构地区：[1]成都理工大学油气藏地质及开发工程国家重点实验室,成都610059 [2]成都理工大学地球物理学院,成都610059

出　　处：《地球物理学进展》2021年第3期901-907,共7页Progress in Geophysics

基　　金：国家自然科学基金项目(41974160,41774192)资助。

摘　　要：前兆异常出现的时间特征能够体现地震孕育的过程,通过分析前兆信息可以对地震进行短临预测,在地震预测研究中发挥着重要作用.然而传统的计算方法和处理分析模式已经很难快速地从海量观测数据中自动定位异常,识别的精度和适用性也存在不足.因此,本文从井生产数据出发,结合STL时间序列分解、趋势拟合、扩容理论和油气运移等,对汶川地震川西气井压力前兆异常现象和产生机理进行分析,并提出将门控循环神经网络(GRU)模型应用于前兆异常识别中.结果表明,川西地区多口井的气井压力数据在汶川地震前同时出现了异常降低现象,很好地反映出地震前的短临异常.相比于循环换神经网络(RNN),GRU模型充分考虑了井生产数据之间的复杂非线性关系和历史关联程度,能够准确地识别出气井压力的前兆异常现象,模型结果也具有较小的均方根误差(RMSE)和平均绝对误差(MAE),可以作为一种新思路应用于地震前兆异常识别当中.The temporal characteristics of precursory anomalies can reflect the process of earthquake preparation. Through the analysis of precursory information, short-term and imminent earthquake prediction can be carried out, which plays an important role in earthquake prediction research. However, the traditional calculation method and processing analysis mode have been difficult to automatically locate anomalies from massive observation data, and the accuracy and applicability of identification are also insufficient. Therefore, based on the well production data, combined with STL time series decomposition, trend fitting, expansion theory and oil and gas migration, this paper analyzes the pressure precursor anomaly phenomenon and generation mechanism of gas wells in Western Sichuan in Wenchuan earthquake, and proposes to apply the Gated Recurrent Unit network(GRU) model to the precursory anomaly identification. The results show that the abnormal decrease of gas well pressure data of several wells in Western Sichuan before the Wenchuan earthquake, which well reflects the short-term and impending anomalies before the earthquake. Compared with the Recurrent Neural Network(RNN), GRU model fully considers the complex nonlinear relationship and historical correlation between well production data, and can accurately identify the precursory anomalies of gas well pressure. The results of the model also have small Root Mean Square Error(RMSE) and Mean Absolute Error(MAE), which can be used as a new idea in the identification of seismic precursory anomalies.

关 键 词：前兆异常 气井压力 STL时间序列分解 循环神经网络 门控循环单元神经网络 

分 类 号：P315[天文地球—地震学]