An automatic methodology for lithology identification in a tight sandstone reservoir using a bidirectional long short-term memory network combined with Borderline-SMOTE

The increasing difficulty in conventional oil and gas exploration and development has brought tight sandstone reservoirs into focus. These reservoirs have a variety of lithofacies types and strong heterogeneity. Accurately identifying the complex lithology of tight sandstone reservoirs is crucial for locating favorable reservoirs and guiding oil and gas exploration and development. In this study, we propose a hybrid method for imbalanced lithology identification. This method combines a bidirectional long short-term memory (BiLSTM) deep neural network model with Borderline-SMOTE. The Borderline-SMOTE oversampling method handles class-imbalanced data by region, generating new minority class samples in a targeted manner, balancing the data and reducing noise interference. The BiLSTM model integrates information from the upper and lower surrounding rock strata, capturing complex features of continuous sequence data and accurately identifying small interbedded lithologies within long sections of mudstone or main sandstone. We compared our model with support vector machines and neural network models using actual logging data and the public dataset. We used accuracy, precision, recall, and F_β score as evaluation metrics. The results show that our method performs well on imbalanced data, with higher recall and F_β score than other models. It also demonstrates superior robustness on public datasets, showing potential for practical application.