基于深度生成模型的河道砂建模方法

在传统的河道砂体建模方法中，基于目标建模的方法难以条件化，多点地质统计学难以再现连续河道砂体的形态，导致建模成果难以直接应用于油田生产。深度生成模型通过深度学习，可以生成足够精确的河道砂体模型，能再现复杂的河道砂体形态，很好地满足井点条件，弥补了传统建模算法的不足。在建模过程中，首先基于目标模拟方法与计算机匹配操作建立了20 000个河道砂体模型与对应的条件集，并结合变分自编码（VAE）与生成对抗网络（GAN）的理论，建立深度生成模型，其中包括分类器、编码器、解码器与判别器。将条件数据与真实模型输入深度生成模型中得到对应的河道砂体模型，通过大量的训练建立了可以生成满足井点条件的河道砂体的生成器，最后将井点数据输入生成器中建立相应的河道砂体模型。研究结果表明，深度生成模型建模算法与传统建模算法相比不仅展现出了连续、清晰的河道砂体，并且在给定的条件下可以生成多个河道砂体模型。该建模方法克服了传统河道砂建模方法的不足，为河道砂体储层建模提供了新的解决思路，建立的河道模型可为油田开发提供参考。 

A modeling method of channel sand based on deep generation models

In traditional channel sand body modeling methods, the method based on target modeling is difficult to condition, and multipoint geostatistics have difficulty reproducing the shape of continuous channel sand bodies, which makes the modeling results difficult to directly apply to oilfield production. Through depth learning, the depth generation model can generate a sufficiently accurate channel sand body model, reproduce the complex channel sand body shape, meet the well point conditions, and compensate for the shortcomings of the traditional modeling algorithm. In the process of modeling, 20 000 channel sand body models and corresponding condition sets are established based on the target simulation method and computer matching operation. Combined with the theory of variational autoencoders (VAE) and generative adversarial networks (GAN), a depth generation model is established, including a classifier, encoder, decoder and discriminator. The condition data and the real model are input into the depth generation model to obtain the corresponding channel sand body model. Through extensive training, a generator that can generate channel sand bodies that meet the well point conditions is established. Finally, the well point data are input into the generator to establish the corresponding channel sand body model. The results show that compared with the traditional modeling algorithm, the depth generation model modeling algorithm not only shows a continuous and clear channel sand body but can also generate multiple channel sand body models under given conditions. This modeling method overcomes the shortcomings of traditional channel sand modeling methods and provides a new solution for channel sand reservoir modeling. The established channel model can provide a reference for the oilfield development stage.