An improved integration strategy for prediction of shear wave velocity using petrophysical logs: Integration of spatiotemporal and small sample nonlinear feature

Abstract

Shear wave velocity (Vs) plays an important role in seismic data inversion and accurately determining petrophysical parameters. However, owing to the high acquisition cost of S-wave logging, Vs-related data cannot be obtained from all wells, especially the older ones. Therefore, developing an efficient and fast intelligent prediction method for Vs using existing petrophysical logging data is necessary. Owing to the advantages of conventional machine learning (CML) and deep learning (DL) methods, this study developed a data-driven integrated learning (IL) method for Vs prediction, which uses sensitive logging data as input, builds multiple sample sets, analyzes the performance differences of the CML and DL methods for predicting Vs in each dataset, and optimizes the support vector machine (SVM) in the CML method and convolutional neural network-long short-term memory (CNN-LSTM) in the DL method to construct an integrated CNN-LSTM-SVM neural network (ICLSNN). The spatiotemporal feature extraction capability of CNN-LSTM was used to effectively extract the deposition information of the formation from the logging data, and the small sample nonlinear feature extraction capability of SVM was used to alleviate the limitations of overfitting in DL for data-driven intelligent Vs prediction. The ICLSNN showed a higher Vs prediction accuracy than SVM or CNN-LSTM alone. The Vs prediction results of the single well in the discovery without training were better than those of the other IL and empirical models, proving the strong generalization applicability of the proposed model. This method compensates for the limitations of individual machine learning or DL models for Vs prediction and provides a novel research idea. The effective implementation of this scheme will facilitate the efficient prediction of Vs in areas with few wells.