基于元学习Transformer-BiLSTM的测井曲线重构

Well-Log Reconstruction Based on Meta-Learning with Transformer-BiLSTM

摘要: 随着页岩气资源勘探和开发工作的不断深入，测井数据在储层评价中所起的作用也越来越大。但是由于设备故障、作业成本等各方面的原因，在实际的测井过程中常常会出现曲线缺失或者异常的情况，从而影响到地质解释的精度以及资源开发的决策。本文设计出一种将元学习、Transformer和双向长短期记忆神经网络(BiLSTM)结合在一起的深度学习模型，简称MTB。元学习模块依靠跨任务元训练机制，使得模型可以从很少的数据样本中迅速学会测井数据的内在规律，进而明显改善小样本情况下的重构性能和泛化水平；Transformer编码器可以捕捉到测井序列里长距离的依赖关系，加强了全局特征的表达能力；BiLSTM会继续对前后向时序特征加以提取，从而加强对于测井曲线动态改变的建模精确度。三者共同作用可以使得模型具有很好的学习和重构能力，提高测井曲线的重构精度和稳定性。在四川绵阳北部地区6口井的实测数据集中进行单目标测井曲线重构实验，结果表明MTB模型在RMSE、MAE、MAPE、R²等各方面都比现有的主流模型有更高的预测精度和更好的泛化能力。

Abstract: With the continuous development of shale gas exploration and production, more and more data from well logs are being used to assess the reservoirs. Due to various reasons such as equipment failure, operating costs, etc., there are often missing or irregular curves in the actual logging operation; as a result, the accuracy of geological interpretation and the reliability of resource development plans are greatly affected. To address the problem of missing or abnormal well logs, a new kind of deep learning model called MTB has been put forward in this paper, which combines meta-learning, Transformer and bidirectional long short-term memory (BiLSTM). A meta-learning module is a cross-task meta-training method that can learn the general characteristics of well-logging data from a small number of samples very quickly to improve reconstruction performance and generalisation ability in a few-shot scenario. The Transformer encoder can collect information from the far-away parts of the well log sequence. BiLSTM is used to obtain the forward and reverse temporal information, and then changes in the well log over time can be modelled. The three components together can achieve good learning and reconstruction by the model to improve the accuracy and stability of well log curve reconstruction. Experiments on the single-target well log reconstruction were carried out based on the actual data set of the six wells in the northern Mianyang area of the Sichuan Basin. From the above results, it can be seen that the MTB model is better than other typical models in terms of RMSE, MAE, MAPE and R², and has relatively high prediction accuracy and generalisation ability.