基于空间近似概率约束的混合密度网络砂体厚度预测

利用地震数据估计储层参数具有不确定性。传统神经网络等方法可视为一个复杂函数,不适用于不确定性问题。针对这一问题,以砂体厚度预测为例,提出了一种基于空间近似概率约束的混合密度网络储层参数预测方法。首先,通过混合密度网络获得地震道储层参数“观测”概率分布;然后,根据地下介质的空间横向渐变性假设,获得地震道储层参数“估计”概率分布;最后,将两种概率分布融合,选取融合后概率分布的期望作为储层参数最优估计。建立了理论模型,测试了BP神经网络、混合密度网络以及所提方法在不同数量训练样本条件下的预测效果,结果表明,随着训练样本的减少,3种方法准确度均下降,但是,在相同数量训练样本下本方法的预测效果更好。X工区的实际应用结果与模型测试结果相一致,综合判断基于空间近似概率约束的混合度网络储层参数预测方法具有良好的应用潜力。

Prediction of sand body thickness based on a mixed density network constrained by a spatially approximated probability

Estimating the reservoir parameter from seismic data is complex due to the uncertainty associated with these data.The traditional neural network method,which could be regarded as a complex non-linear function,cannot solve the uncertainty problem.To address this problem,a method was proposed to predict the reservoir parameter based on a mixed density network(MDN)constrained by a spatially approximate probability.First,the“observed”probability distribution of the reservoir parameter was obtained through the MDN.Then,with the assumption that the subsurface medium has spatial continuity,the“estimated”probability distribution of the reservoir parameter was obtained.Finally,the above two probability distributions were fused,and the expectation of the fused probability distribution was selected as the optimal estimation of reservoir parameters.This method was compared with the BP neural network and the MDN,through tests on a theoretical model and real data in the X area.The results showed that with a decrease of the number of training samples,the accuracy of the three methods all declined.However,with the number of training samples being the same,the proposed method yielded the best prediction.The proposed method could therefore provide an effective mean for reservoir parameter prediction with a high application potential.