利用条件生成对抗网络建立曲流河地质模型

Building a Geologic Model of a Meandering River Using Conditional Generative Adversarial Networks

【目的】在传统的河道建模方法中,基于目标的方法难以刻画曲流河点坝且条件化困难,多点地质统计学则难以再现河道的连续形态。条件生成对抗网络可以生成满足一定条件的复杂图形,可解决曲流河地质模型建立过程中点坝、河流形态刻画困难及难以条件化的问题。【方法】以鄂尔多斯盆地苏里格气田南部地区某气田为例,开展了基于条件生成对抗网络的曲流河三维建模方法研究。在建模过程中,首先采用Alluvsim建模方法根据工区曲流河特征建立了200个曲流河模型;再通过卷积神经网络对200个模型进行深度学习,提取模型的特征矩阵,利用条件生成对抗网络建立可以生成曲流河模型的生成器;最后以工区井点数据作为输入数据,利用生成器建立满足曲流河复杂形态和井点数据的三维模型。【结果与结论】所建立的模型可以很好地展现曲流河中河道与点坝的三维形态及对应关系。为明确影响模型结果的关键因素,通过对比训练次数与输入数据发现,适当的训练次数(160次)与大量地输入样本(200个)是建立满足工区条件模型的前提。另外,通过对比传统地质建模方法,条件生成对抗网络建模方法可以很好地再现河道沉积体的空间形态,克服传统曲流河建模方法在条件化方面的困难,为曲流河沉积环境的河道砂体建模提供了新的解决思路,建立的曲流河模型可为油田开发阶段提供参考。

[Objective]In traditional channel modeling methods,object-based methods are difficult to characterize meandering river point bars,and it is difficult to condition.Multi-point geostatistics makes it difficult to simulate the continuous morphology of channel.Conditional Generative Adversarial Networks can generate complex graphics that meet certain conditions,which can solve the difficulties in characterizing point bars and channel morphology during the establishment of geological models for meandering rivers.Moreover,the generated models can meet the given well point conditions.[Methods]Taking a gas field in the southern part of the Sulige gas field in the Ordos Basin as an ex-ample,a three-dimensional modeling method for meandering rivers based on conditional generative adversarial net-works was studied.In the modeling process,firstly,200 meandering river models were established using the Allu-vsim modeling method based on the characteristics of the meandering rivers in the work area.Then,deep learning was performed on 200 models using convolutional neural networks to extract the feature matrix of the models.A gener-ator capable of generating meandering river models was established using Conditional Generative Adversarial Net-works.Finally,taking the well data of the work area as input data,a 3D model that satisfies the complex shape of the well data and the meandering river is established using a generator.[Results and Conclusions]The results indicate that the established model can well demonstrate the three-dimensional morphology and corresponding relationship be-tween the channel and point bar in meandering rivers.To clarify the key factors that affect the model results,it was found through comparing the training times with the input data that an appropriate training times(160 times)and a large number of input samples(200 samples)are prerequisites for establishing a model that meets the working area conditions.In addition,by comparing traditional geological modeling methods,the conditional generative adversarial network modeling method can effectively reproduce the spatial morphology of channel sediment bodies,overcome the shortcomings of traditional meandering river modeling methods in terms of conditioning difficulties,and provide a new solution for modeling channel sand bodies in meandering river sedimentary environments.The established mean-dering river model can provide reference for the oilfield development stage.