A Fast Inversion Method For Array Laterolog Based On Convolutional Neural Network And Hybrid Mpga-Lm Algorithm

Wu Yizhi,Fan Yiren,Wu Zhenguan,Deng Shaogui,Zhang Pan,Chen Shiyu,Yin Zhongxu

DOI: https://doi.org/10.6038/cjg2021O0499

2021-01-01

Abstract: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.