基于MC-ABiLSTM的储层单砂体智能识别方法

Intelligent identification method of reservoir single sand body based on MC-ABiLSTM

油气田开发中后期急需储层单砂体精细识别,为后期剩余油气开发提供地质依据.目前,传统方法依靠研究人员对测井、岩心等资料进行综合对比分析,主观性强、效率低下;而在人工智能领域,现有的基于测井数据的储层识别方法,多研究测井曲线的波形特征,或对单个采样点进行独立分析,忽略了测井数据的纵向相关性,缺少全局认识,识别结果存在一定的局限性.针对上述问题,提出了一种基于注意力机制的双向长短期记忆神经网络(Multi-Scale Convolutional Layers Cascaded Attention-Based Bidirectional Long Short-Term Memory,MC-ABiLSTM)模型的储层单砂体智能识别方法.该模型主体分为MC与ABiLSTM两个模块.MC模块由不同尺度卷积核的并联卷积层组成,用于提取测井数据多尺度的空间特征;ABiLSTM模块采用基于LSTM的编解码框架,用于从MC模块输出的特征中提取上下层信息,获取测井数据的时序特征;最后,采用softmax分类器对提取后的测井曲线数据空间和时序特征进行识别分类.以靖安油田大路沟二区Chang6_(1)^(2)含油小层为研究对象,选取研究区岩心井的4种测井曲线(自然伽马、自然电位、电阻率、声波时差)以及3种物性参数(孔隙度、渗透率、含水饱和度)作为样本,训练MC-ABiLSTM单砂体识别模型,并与其他四种智能模型进行识别效果对比.结果表明:MC-ABiLSTM的单砂体识别精度最高,达到91.2%,可作为利用测井资料进行单砂体识别的有效手段.

In the middle and late period of oil and gas field development,precise identification of reservoir single sand body is urgently needed to provide geological basis for remaining oil and gas development in the later period.At present,traditional methods rely on researchers to make comprehensive comparative analysis of logging and core data,which is highly subjective and inefficient.In the field of artificial intelligence,the existing reservoir identification methods based on logging data,which mainly study the waveform characteristics of logging curves or independently analyze a single sampling point,ignore the longitudinal correlation of logging data,lack of global understanding,and the identification results have certain limitations.Regarding the issue above,the paper proposes an intelligent identification method of reservoir single sand body based on Multi-Scale Convolutional Layers Cascade Attention-Based Bidirectional Long Short-Term Memory Neural Network(MC-ABiLSTM).The main body of the model is divided into two modules,MC and ABiLSTM.MC module is composed of parallel convolutional layers of different scale convolution cores,which is used to extract multiscale spatial features of logging data;ABiLSTM module adopts LSTM based encoder-decoder framework,which is used to extract upper and lower reservoir information from the output features of MC module,and obtain the time series features of logging data.Finally,softmax classifier is used to identify and classify the spatial and temporal features.Taking the Chang6_(1)^(2)oil-bearing layer in the second area of Dalugou,Jing’an Oilfield as the research object,four kinds of logging curves(GR,SP,RT,AC)and three kinds of physical property parameters(porosity,permeability,water saturation)of core Wells in the research area were selected as samples,trained the MC-ABiLSTM single sand body identification model,and compared the identification effect with the other four intelligent models.The results show that:MC-ABiLSTM has the highest single sand body identification accuracy,reaching 91.2%,which can be used as an effective method for single sand body identification using logging data.