Fluid Identification in Tight Sandstone Reservoirs Using a Bayesian-Optimized CNN–BiLSTM Model

Abstract

Fluid identification in tight sandstone reservoirs is the core link in the development of tight gas reservoirs, and its accuracy is directly related to the scientific nature of reservoir evaluation and development plan. Due to the complex lithology of the reservoir, the highly overlapping logging response and the uneven distribution of sample categories, the traditional identification methods based on the plate method and conventional machine learning have limited performance in the discrimination of complex types such as "gas and water in the same layer" and "water layer", which is difficult to meet the needs of fine oil and gas identification and efficient development. Therefore, a CNN-BiLSTM deep learning model based on Bayesian optimization is proposed for fluid type identification in tight sandstone reservoirs. Firstly, the composite parameters are constructed based on the original LAS logging data, the input feature dimension is expanded, and a sliding window sequence is formed to express the temporal change trend of fluid response. Secondly, convolutional neural network (CNN) and bidirectional long short-term memory network (BiLSTM) are combined to model local spatial and temporal features. At the same time, Focal Loss is introduced to improve the discrimination ability of the model for minority classes. Finally, Bayesian hyperparameter optimization is carried out by using the Optuna framework to obtain the optimal model structure and learning rate configuration. The results show that the optimized CNN-BiLSTM model has an accuracy of 93.5%, which has good identification ability and application prospect.

Article Type: Research Article

Cited as:

Zhang ZY, Fang SN, Li SM, et al. 2026. Fluid Identification in Tight Sandstone Reservoirs Using a Bayesian-Optimized CNN–BiLSTM Model. GeoStorage, 2(2), 156-171.