深度学习与Eaton法联合驱动的地层孔隙压力预测方法

A novel prediction method of formation pore pressure driven by deep learning and Eaton method

海上深部复杂地层孔隙压力的精确预测一直以来是钻井工程面临的难题之一。针对传统Eaton法的局限性和现有数据驱动法的不足,通过构建地层压力实测点的扩充方法,构建卷积神经网络(convolutional neural network, CNN)和长短期记忆网络(long short term memory, LSTM)组合模型,充分挖掘钻测录震多源数据与Eaton指数之间的复杂非线性关系,可基于区块内已钻井的有限实测地层压力数据,实现全井Eaton指数的精细预测,为新探区地层压力实测点较少且分布不均等条件下地层孔隙压力的准确预测提供有效手段。结果表明:建立的方法预测深部复杂地层孔隙压力的平均相对误差为2.70%,而传统Eaton和LSTM方法的平均相对误差分别为7.60%和5.12%;通过深度学习与Eaton法联合驱动,不但提高了深部复杂地层孔隙压力的预测精度,也为传统方法融入了多源数据响应特征,为数据驱动方法提供了理论支撑。

The accurate prediction of formation pore pressure in deep complex formations has been one of the challenges in drilling engineering.In this paper,the limitations of the traditional Eaton methods and the shortcomings of the existing data-driven methods were discussed,and a convolutional neural network(CNN)and short-term memory network(LSTM)combi-nation model was constructed to fully explore the complex nonlinear relationship between drilling and recorded multi-source data and Eaton index by constructing an extension method for measuring formation pressure points.Based on the limited for-mation pressure data measured in drilled formations,a precise prediction of the Eaton index of the entire well can be a-chieved,which can provide an effective means for accurate prediction of the formation pore pressure in new wells with few measured points and uneven distribution of formation pressure.Field case studies show that the average relative error of the method established for predicting pore pressure in deep complex formations is 2.70%,while the average relative error of the traditional Eaton and LSTM methods is 7.60%and 5.12%,respectively.The combination model of deep learning with Eaton method,not only can improve the prediction accuracy of deep complex formation pore pressure,but it can also integrate multi-source data response features into the traditional methods,providing a theoretical support for the data-driven methods.