Tight sandstone reservoir classification based on 1DCNN-BLSTM with conventional logging data

Machine learning-based reservoir classification method is the development trend of intelligent exploration. In this study, a classification model, one-dimensional convolutional neural network-bidirectional long short-term memory (1DCNN-BLSTM), was developed based on conventional logging data to classify tight sandstone reservoirs. The model utilizes 1DCNN to extract spatial features from conventional logging curves, and it can effectively capture the pattern and regularity of the logging data. In additional, BLSTM neural network was employed to learn the temporal features of logging curves from the upper and lower layers. This network is capable of capturing dependencies within sequences and considers the trends of the logging curves in the lower and upper layers for reservoir type prediction. The efficacy of the proposed method was verified by comparing the classification outcomes of the BLSTM method, in tight sandstone reservoirs in the Shaximiao Formation, Sichuan Basin, as an example. Firstly, a fuzzy C-means clustering algorithm was applied to establish reservoir classification criteria based on high-pressure mercury intrusion data. Reservoir classification labels for sampling points of conventional logging data were generated, and the initial dataset was formed. To avoid the imbalanced classification in the initial dataset, a comprehensive sampling method, named as SMOTE-Tomek, was employed to preprocess the dataset and to ensure a relatively balanced quantity of samples in each class. A training set was randomly generated from 90% preprocessed dataset, and the rest was a test set. The model was trained several times, and the accuracy of the test set classification results was averaged to evaluate its performance. Finally, the field logging data were processed with the BLSTM model and the 1DCNN-BLSTM model to classify reservoirs. The results indicate that the reservoir classification results are more accurate with the 1DCNN-BLSTM model than with the BLSTM model.