Application of deep learning for porosity prediction in tight sandstone reservoirs: A case study of blocks Su14 and Su36

Porosity is a fundamental parameter for assessing reservoir characteristics, significantly impacting the storage capacity of hydrocarbons, groundwater, and other subsurface resources. Traditional methods for measuring porosity, such as core analysis and well logging, are limited by high costs, low efficiency, and inadequate applicability in heterogeneous reservoirs. To address these limitations, this study proposes a novel hybrid deep learning model, CNN-BiLSTM-Attention, for predicting porosity using well log data from Blocks Su14 and Su36 in the Sulige Gas Field. The model combines the feature extraction capabilities of Convolutional Neural Networks (CNN), the temporal dependency modeling of Bidirectional Long Short-Term Memory (BiLSTM), and the dynamic weighting provided by the Attention mechanism. Leveraging ten key well log parameters and advanced preprocessing techniques, the model achieved an R² of 0.86112 and RMSE of 0.036274 on the training set, and an R² of 0.8591 and RMSE of 0.037009 on the test set. Validation using independent datasets from Blocks Su14 and Su36 yielded an R² of 0.8533, RMSE of 0.015465, and RPD of 2.4641, highlighting the model's robustness and practical applicability. Comparative analysis demonstrated that the CNN-BiLSTM-Attention hybrid model outperforms traditional methods, including BP, CNN, ELM, RF, and SVM. This study offers a reliable, efficient, and cost-effective approach for porosity prediction in complex reservoirs, effectively addressing challenges associated with heterogeneity and data nonlinearity.