Research On Shale Simulation Methods Based On Generative Adversarial Networks

With the rapid consumption of conventional energy,the exploration and development of unconventional energy sources is imminent.Shale has the characteristics of good dense,the presence of a large number of natural fractures and wide distribution,and has a vast potential for oil and gas exploration and development.At present,due to the limitation of technology and equipment,the cost of exploration and development of shale oil and gas is high,then,shale simulation based on limited training images will be a significant guide to predict the size of shale oil and gas reservoirs and determine whether they have exploration and development value,etc.In the field of traditional shale simulation,the multi-point geostatistical method(MPS)is widely used,which requires scanning the training images to obtain statistical information,and subsequent complex probability calculations to reproduce this information into the simulated images,which is a large amount of work.At the same time,since it is a stochastic modeling with equal probability information,the results of multiple simulations eventually need to be averaged to obtain the simulated images.Such an approach takes longer to implement,is less efficient,and even requires more conditional data in some cases.Generative adversarial networks,on the other hand,as deep learning techniques,do not require probabilistic statistics of training image information for simulation,but spontaneously learn the data characteristics of training images through unsupervised methods.To address the problems of traditional shale simulation methods,this paper proposes a shale simulation method based on multi-resolution generative adversarial networks,called Multi GAN,in Chapter 3.The method learns the global and local features of the training data in stages by constructing a pyramid structure.The staged learning helps prevent overfitting caused by end-to-end training to generate the input image,while also fully learning the structural features of the training image.The global structural features of the original training image are captured by the generator and discriminator of the five-layer convolutional network structure at a low resolution training image,and as the resolution of the training image increases,the more detailed information is learned by the generator until the maximum resolution network convergence is reached when the generator learns the distribution of the real training data and generates a simulated image.In addition,the Multi GAN method can store the learned feature information locally in the form of parameters without the need to rescan the training image for each simulation as in the MPS method,reducing the time required for multiple simulations.The detailed analysis of porosity,variance function,multi-point connectivity,permeability,and pore network model-related properties confirms that the Multi GAN method can generate higher quality simulation images than the MPS method,has faster simulation speed,reduces the need for training images,and also saves the model for easy reuse;compared to the shale simulation method using the base GAN,the Multi GAN only requires a single shale training image instead of a large shale training dataset,which is of great interest for shales where it is inherently difficult to obtain training images,and the Multi GAN method has a smaller memory footprint,higher GPU utilization,and faster training speed.However,the Multi GAN method trains one phase in one time period,i.e.,training one resolution training image,which results in each phase being sequential and without interaction,which is not conducive to updating the network parameters for better training results.Therefore,this paper also proposes the concurrent multi-resolution generative adversarial network(Con Multi GAN)method in Chapter 4,which improves on the Multi GAN method by setting three resolution images to be trained in one time period,which improves the interaction during the training of different resolution images,and therefore increases the training speed while improving the training quality.In addition,in order to further improve the training efficiency,the formula for calculating the size of the training image size is improved from strictly proportional to having more training images in the low resolution case;the update formula for the learning rate is adjusted from a fixed value to an exponentially varying value,with the lowest learning rate in the training process for the low resolution training images and the largest learning rate value for the training process for the high resolution training images.The experimental analysis results confirm that the Con Multi GAN method has a better training effect compared to the Multi GAN method,and the time is also reduced by more than half.