深度Transformer迁移学习的页岩气储层核心参数预测案例

地层纵横向非均质性强, 工区间数据分布存在差异.这导致基于已有工区数据构建的机器学习储层参数预测模型, 推广到新工区会存在较大预测误差.常规地质方法是在岩心与测井响应特征分析基础上建模, 利用测井资料计算储层参数, 流程复杂.该方法需要岩心校准模型, 同样难以快速推广到新的工区.考虑地层纵横向非均质性, 本文设计了一种深度Transformer迁移学习网络, 通过已有工区的测井与岩心资料构建预测模型, 实现未取心新工区储层参数快速准确预测.首先利用无监督学习算法-孤立森林剔除测井数据中存在的异常噪声数据.然后设计Transformer特征提取网络, 提高网络特征提取能力, 以此深入挖掘测井数据与储层参数的内在联系.最后设计深度迁移学习网络, 构建网络损失函数, 利用随机梯度下降算法优化网络参数, 实现储层参数准确预测.本方案应用于四川南部地区五峰组—龙马溪组页岩储层参数孔隙度、总有机碳含量和总含气量预测.实验结果与工区校正后计算结果、主流机器学习模型预测结果对比, 本方案结果与岩心数据具有更高的一致性.应用结果表明: 本文方案具有实用性、有效性和可推广性.

Prediction case of core parameters of shale gas reservoirs through deep Transformer transfer learning

Due to the strong heterogeneity of the reservoir, it is difficult to extend the classical reservoir parameter prediction model to the new work area. Traditional geological methods rely on experience, while machine learning algorithms lack consideration of reservoir heterogeneity. This makes it impossible for these methods to achieve good prediction performance in the new work area. In this paper, we design a deep Transformer transfer learning network, which builds a learning model adopting log and core data from the old work area to accurately predict reservoir parameters in the new work area. First, we adopt the isolation forest to remove abnormal noise data from log data. Second, we design a Transformer feature extraction network to deeply mine the intrinsic relationship between logging data and reservoir parameters. Finally, we design the transfer network and define the loss function. We optimize model parameters through the stochastic gradient descent algorithm to achieve an accurate prediction of reservoir parameters. The method was applied to the prediction of shale gas porosity, total organic carbon content, and total gas content in southern Sichuan. Compared to the actual results and the predicted results of classical machine learning, our method achieves the best performance in the new work area. The experimental results show that the scheme is practical, effective, and scalable.