基于Attention机制的LSTM测井曲线预测方法

LSTM Logging Curve Prediction Method Based on Attention Mechanism

测井曲线预测是解决因井径变化、仪器故障等因素造成曲线测量效果不佳或某段曲线缺失的有效手段。从记忆学习的角度入手,对传统长短时记忆(LSTM)循环神经网络在测井曲线预测过程中特征提取、权值匹配及预测误差等问题进行分析,阐释Attention机制(以下简称Att机制)在解决此类问题中的重要性。基于TensorFlow平台搭建了Att-LSTM预测模型,并利用该模型实现了测井曲线的预测。结果表明:在LSTM神经网络中引入Att机制增强了不同时深下的测井特征关联性,优化了不同时间段的测井特征权重分配问题。将Att-LSTM模型和LSTM模型实际预测结果进行对比,Att-LSTM模型预测的测井曲线误差有所下降,精度较传统LSTM模型提高约8%,证实了在LSTM模型中引入Att机制的合理性及实用性,为测井曲线预测提供了一种新思路。

Logging curve prediction is an effective method to solve the problem of poor curve measurement results or missing of a certain section of the curve caused by factors such as well diameter changes and instrument failures.Starting from the perspective of memory learning,this paper first analyzes the feature extraction,weight matching and prediction error of traditional long short-term memory(LSTM)recurrent neural network in the process of logging curve prediction,and then explains the Attention mechanism(hereinafter referred to as the Att mechanism)plays an important role in solving such problems.Finally,the Att-LSTM prediction model is built based on the TensorFlow platform,and the prediction of the logging curve is realized by using this model.The prediction results show that:the introduction of the Att mechanism into the LSTM neural network enhances the correlation of logging features at different depths and optimizes the weight distribution of logging features in different time periods.Comparing the actual prediction results of the Att-LSTM model and the LSTM model,the error of the logging curve predicted by the Att-LSTM model has decreased,and the accuracy is about 8%higher than that of the traditional LSTM model,which confirms the rationality and practicability of introducing the Att mechanism into the LSTM model.This study provides a new idea for the prediction of well logging curve.