Fast prediction of reservoir permeability based on embedded feature selection and LightGBM using direct logging data

Permeability estimation plays an important role in reservoir evaluation and hydrocarbon development, etc. Traditional physical model-based methods have problems with being time consuming and high cost. The applications of machine learning are currently becoming more and more extensive, however, there are still several limitations to previous machine learning-based permeability estimation methods, such as a limited number of samples, a requirement of prior knowledge, and some parameters needing to be calculated indirectly. In this paper, a hybrid reservoir permeability prediction approach, which is based on a certain scale of permeability dataset, embedded feature selection (EFS) and a light gradient boosting machine (LightGBM), is proposed. First, EFS is used to select features from the raw dataset. Then a LightGBM is adopted to predict the permeability. The influence of feature selection threshold, the base learners' number and dataset size on prediction results is also investigated. In addition, different feature selections and prediction models are compared. The proposed hybrid approach is also verified on other datasets. The experimental results show that the proposed approach can effectively predict the reservoir permeability based on limited direct logging data.