Reservoir Stochastic Simulation Based on Octave Convolution and Multistage Generative Adversarial Network

Abstract

For finely representation of complex reservoir units, higher computing overburden and lower spatial resolution are limited to traditional stochastic simulation. Therefore, based on Generative Adversarial Networks (GANs), spatial distribution patterns of regional variables can be reproduced through high-order statistical fitting. However, parameters of GANs cannot be optimized under insufficient training samples. Also, a higher computing consumption and overfitting issue easily occurred by stacking Convolutional Neural Networks (CNNs). Therefore, a hybrid framework combined with octave convolution and multi-stage GAN (OctSinGAN) is proposed to perform reservoir simulation. Specifically, a pyramid structure is introduced to perform multiscale representation based on single Training Image (TI), with feature information being captured under different scales. Then, the octave convolution is used to perform multi-frequency feature representation on different feature maps. Finally, a joint loss function is defined to optimize network parameters to improve simulation qualities. Three different kinds of TIs are used to verify the simulation performance of OctSinGAN. Results show that different simulations are similar to the corresponding TIs in terms of spatial variability, channel connectivity and spatial structures, with a relatively high simulation performance overall.