Missing log synthesis based on stacking ensemble learning with invariable features

Abstract

Geophysical logging is a vital measurement technique for exploring underground resources such as oil, natural gas, minerals, and groundwater. However, the missing log problem caused by different reasons will hinder the progress of practical applications. Due to the structure of geological reservoirs being so complex, many prediction methods for revealing the complex nonlinear relationships between different well logs have been proposed in the literature. However, many of these methods are insufficient to explore the reliable ccorrelations and complementary information among different well logs. In addition, the adaptability of these prediction models across different wells is poor, especially on small datasets. To address these problems in certain extent, several novel invariant features are proposed to be extracted at first for robustly estimating missing logs. And then, with the extracted invariant features, three heterogeneous machine learning models are integrated in the stacking ensemble manner for predicting missing logs in small datasets. Multiple experiments were conducted to validate the performance of the proposed method. Experimental results illustrated that the proposed method can efficiently restore missing logs and has stable robustness among different wells. Specifically, the quantitative Pearson Coefficient (PCC) between the estimated logs and the corresponding truth logs can achieve 95.2%.