论文详情
深度神经网络通过分析由不同地质和地球物理来源获得的数据之间的相关性确定油气储层特征,从而开展地震油气储层分布预测。经过适当的训练,深度神经网络可以通过识别与样本数据相关的复杂非线性关系来预测油气储层性质,然而目前仍缺少深度神经网络超参数的选取对地震油气藏分布预测结果影响的系统性研究。为此,在分析深度神经网络隐含层数目、隐含层节点数及激活函数的基础上,探讨了深度神经网络模型超参数选取对含油气性多波地震响应特征提取结果的影响,并利用多种评价指标对不同网络结构模型的性能进行了对比。结果表明,深度神经网络隐含层数目等超参数的选取会影响地震油气藏分布范围的预测精度;同时,深度神经网络在参数选取满足精度要求(即均方误差MSE小于0.001)的情况下,可以取得良好的预测结果,从而验证了深度神经网络用于含油气性多波地震响应特征提取的有效性和可行性。
The main purpose of applying deep neural networks(DNNs) to seismic oil and gas reservoir distribution prediction is to determine the characteristics of oil and gas reservoirs by analyzing the correlation between data obtained from different geological and geophysical sources.With proper training,a DNN can predict reservoir properties by identifying complex nonlinear relationships associated with the sample data.However,there is still a lack of systematic analysis on the influence of hyperparameter selection of DNNs on reservoir prediction.Therefore,based on the number of hidden layers,number of nodes in hidden layers,and activation function of DNN,the influence of hyperparameter selection of the DNN model on the extraction of oil and gas reservoir characteristics was discussed,and the performances of different network structure models were compared using various evaluation indexes.The results showed that the selection of hyperparameters(such as the number of hidden layers) could affect the prediction accuracy of the oil and gas reservoir boundary distribution.In addition,the DNN model can achieve good prediction results when the parameter selection meets the accuracy requirements(i.e.,MSE is less than 0.001),which verifies the effectiveness and feasibility of DNN in seismic reservoir characteristics extraction.This research result is helpful in creating a better configuration scheme when applying neural networks to other datasets in the future.
山东省自然科学基金项目(ZR202103050722)和国家自然科学基金项目(41174098，41374126)共同资助。