基于卷积神经网络和MPGA-LM算法的阵列侧向测井快速反演方法

A fast inversion method for array laterolog based on convolutional neural network and hybrid MPGA-LM algorithm

在大斜度井/水平井环境下,阵列侧向测井受钻井液侵入、地层倾角和各向异性等多种因素影响,导致测井响应复杂,需借助反演手段提取地层真实电阻率.然而,阵列侧向测井三维正演效率低,难以满足测井资料快速反演和油气藏快速评价的需求.为此,本文基于深度学习并联合混合多种群遗传(MPGA)与列文伯格马奎特(LM)算法建立了一种快速反演方法.首先从地层参数敏感性出发,基于严格的三维有限元正演算法,依次确定侵入、各向异性和地层倾角的敏感性大小;其次,引入深度学习和模型可视化技术,实现斜井各向异性地层阵列侧向测井响应的快速正演;最后,基于数据集分解技术和混合MPGA-LM算法,实现斜井各向异性地层电阻率剖面快速精确重构.数值模拟结果表明:斜井各向异性地层中,阵列侧向测井响应对侵入深度敏感性最高,各向异性和地层倾角次之;相较于反向传播神经网络(BPNN),二维卷积神经网络(2D-CNN)能够实现阵列侧向测井响应的快速精确计算,计算一个测井点仅需0.36 ms,精度可达99%左右;基于三层反演模型的MPGA-LM混合算法稳定性强,电阻率参数反演精度高的优点,可用于阵列侧向测井资料的快速处理.

Array laterolog responses are affected by drilling fluid invasion,dipping angle and formation anisotropy in high angle(HA)and horizontal(HZ)wells,which results in complex logging responses.Therefore,it is necessary to extract formation resistivity from array laterolog data by inversion method.However,the 3D forward modeling is time-consuming,which can′t support fast inversion and fine evaluation for oil and gas reservoirs.Therefore,a fast inversion scheme is presented in this paper based on deep learning,hybrid multi-population genetic algorithm(MPGA)and Lenvenberg Marquardt(LM)algorithm.The sensitivities of responses to invasion,anisotropy and dipping angle are analyzed firstly based on 3D finite element method(3D-FEM).And then,deep learning and model visualization schemes are introduced to realize the fast forward modeling of array laterolog in anisotropic formation with HA/HZ wells.Finally,based on the data set decomposition technology,the MPGA and L-M algorithms are combined to reconstruct resistivity profile of anisotropic formation rapidly and accurately.The numerical simulation results show:array laterolog responses are more sensitive to invasion than anisotropy and dipping angle.Compared with back propagation neural networks(BPNN),2D convolution neural network(2D-CNN)is more efficient,and it only takes 0.36 ms to calculate a logging point with an accuracy of about 99%.Hybrid MPGA-LM inversion algorithm based on three-layer inversion model is robust,which can be applied to array laterolog data processing.