基于地震属性和深度前馈神经网络的天然气水合物饱和度预测

Estimation of gas hydrate saturation based on seismic attributes and deep feedforward neural network

天然气水合物饱和度是评价资源量的重要参数,常规的地震反演储层预测方法存在精度低、效率不高的问题,且无法解决地震数据与储层参数之间的非线性问题。随着人工智能技术的迅速发展,许多相关技术已经应用于地震勘探领域,其中人工神经网络是人工智能的一个重要分支,其可以通过从大量的样本数据中不断学习,进而拟合复杂非线性函数来实现地下储层特征反演,有着很强的非线性映射和泛化能力。因此,本文在分析了常规线性公式以及岩石物理建模法优缺点的基础上,提出了基于地震属性和深度前馈神经网络预测水合物饱和度的方法。首先,基于测井和地震数据,通过筛选出不同类型与水合物饱和度相关性高的地震属性体,多维度构建样本标签数据;然后采用地震反演与端到端(地震数据-储层物性数据)反演相结合的策略,对全连接神经网络的隐藏层数、神经元数量、迭代次数等参数进行测试训练,最后将训练结果应用于地震数据体获得水合物饱和度预测结果。实际数据应用结果表明:基于地震属性和深度前馈神经网络预测的饱和度结果精度高、多解性低,与测井数据吻合好,证明该方法具有较好的应用价值;同时,预测的水合物空间分布特征表明研究区水合物成藏为平面游离气-水合物过渡成藏模式。

Gas hydrate saturation is an important parameter to evaluate the potential of gas hydrate resources.Conventional seismic inversion methods for reservoir prediction,however,exhibit limitations in accuracy and efficiency,mainly due to their inability to adequately address the inherent nonlinear relationship between seismic data and reservoir parameters.The rapid development of artificial intelligence(AI)has led to the exploration of novel approaches for seismic exploration.Among these,artificial neural networks(ANNs)have emerged as a promising tool.ANNs possess the capability to learn complex patterns from large datasets,enabling them to approximate nonlinear functions with high accuracy.This inherent ability to model complex relationships makes ANNs particularly suitable for inverting subsurface reservoir characteristics,including gas hydrate saturation.Therefore,this paper proposes a novel method for predicting gas hydrate saturation based on seismic attributes and deep feedforward neural networks.This approach consider the advantages and disadvantages of conventional linear formula and petrophysical model methods.The proposed method involves a three-step process.First,different types of seismic attribute bodies with a strong correlation to gas hydrate saturation are identified based on logging and seismic data.These attributes are then used to construct multi-dimensional sample label data.Second,a strategy combining seismic inversion with end-to-end(seismic data-reservoir physical data)inversion is used to test and train the parameters of the fully connected neural network.This includes optimizing the number of hidden layers,the number of neurons,and the number of iterations.Finally,the trained neural network is applied to the seismic data body to obtain predictions of hydrate saturation.Actual data application results show that the saturation predictions based on seismic attributes and deep feedforward neural networks exhibit high accuracy and low multi-solution,showing good agreement with the logging data,which proves that the method has good application value.Furthermore,the predicted spatial distribution characteristics of hydrate indicate that the hydrate accumulation pattern in the study area is a flat-lying transitional free gas to gas hydrate system.