基于门控循环单元神经网络的测井曲线预测方法

Logging curve prediction method based on GRU

为了减少泥浆侵入对测井曲线的影响,许多油田采用随钻测井技术,需先预测未钻地层测井曲线,这对随钻测井具有非常重要的指导作用。为此,提出一种基于门控循环单元神经网络(GRU)预测未钻地层测井曲线的方法,该模型将长短期记忆神经网络(LSTM)的输入门和遗忘门合并成更新门,输出门变成重置门,使模型结构简单,不易出现过拟合现象,保留LSTM模型的长时记忆功能,且能有效缓解梯度消失或梯度爆炸问题。以新疆油田直井和南海西部油田随钻测井的实际测井数据为例,选取已钻地层以及邻井的自然伽马、深感应电阻率、声波时差、密度和井径5条测井曲线数据作为训练样本输入到LSTM和GRU模型中进行学习训练,将训练好的模型用于预测未钻地层的测井曲线。应用结果表明,GRU比LSTM模型在新疆油田和南海西部油田预测测井曲线的平均相关系数分别提高13.78%和12.13%,平均均方根误差分别下降27.08%和42.17%,GRU模型能够准确地预测未钻地层测井曲线的变化趋势。

Logging while drilling(LWD)technologies are employed in many oilfields to reduce the impact of mud intrusion on logging curves,which require the prediction of logging curves for undrilled formations as it is of great guiding significance to LWD. Therefore,a method based on the gated recurrent unit(GRU)neural network was applied to predict the logging curves of undrilled formations. The model combines the input gate and forget gate of long short-term memory(LSTM)into an update gate and turns the input gate into a reset gate,which makes its structure simple and not prone to overfitting.Meanwhile,it retains the long-term memory function of the LSTM model and can effectively alleviate the problem of gradient vanishing or explosion. Taking real logging data from vertical wells in Xinjiang Oilfield and LWD data in western South China Sea Oilfield as examples,this study selected the five logging curves of drilled formations and adjoining wells,namely,the curves of the natural gamma ray,deep induction resistivity,acoustic time difference,density,and well diameter,as training samples and input into the LSTM model and GRU model for learning training. The trained models were then used to predict the logging curves of undrilled formations. The application results indicate that the average correlation coefficients of predicted logging curves for Xinjiang Oilfield and western South China Sea Oilfield by the GRU model are 13.78%and 12.13% higher than that of the LSTM model,and the mean root mean square errors are decreased by 27.08% and42.17%,respectively. The GRU model can accurately predict the variation trend of logging curves for undrilled formations.