基于NRBO-CNN-LSTM模型的陆相浅水湖盆总有机碳测井预测优选及应用

打开文本图片集

Abstract： Total organic carbon content （TOC） is the main parameter for evaluating the hydrocarbon generation potential of source rocks. Commonly used TOC logging models are greatly affected by differing geological conditions in practical applications，and theirstability is not high，which restricts their abilityto producecomprehensive evaluation results.In this paper，the second member of the Lianggaoshan Formation in northern Sichuan Basin is selected as theresearch object.Basedon TOC coring data and conventional logging data，the random forest algorithm is used to evaluate the importance of selected logging curves; eliminate the influence of measurement error and redundant data between logging curves；and select four types of logging curves as model input parameters.The resultingconvolutional neural network （CNN），which has high accuracy and strong stability，is combined with the long short-term memory （LSTM） and Newton-Raphson optimization （NRBO） algorithms to optimize the resulting neural network， termed NRBO-CNN-LSTM，and determine the optimal hyperparameters relevant for TOC prediction. The model predictions show that the determinationcoefcient ofNRBO-CNN-LSTM is as highas 0.9763，the mean square error and mean absolute error of the prediction results are O.107O and O.240 3，respectively，and the overall error is .0521.Insedimentary environments where sandand mudare frequentlyinterbedded，and the logging curves fluctuate greatly with lithology changes， NRBO-CNN-LSTM makes up for the shortcomings of conventional neural network prediction algorithms and effectively improves the accuracy of TOC prediction.

Key words： total organic carbon （TOC）； machine learning；terrestrial shale；Sichuan Basin；Lianggaoshan Formation

总有机碳含量（TOC是评价烃源岩生烃潜力的重要参数，是定量评价烃源岩有机质丰度最有效、最常用的指标之一（黄东等，2017；徐仕琨等，2020）。（剩余34764字）