A Hybrid Bayesian Self-Attention Mechanism Long Short-Term Memory Method for Robust Production Prediction in Tight Sandstone Gas Reservoirs

Abstract Reliable production forecasting is essential for the efficient management of tight sandstone gas reservoirs, where frequent well shut-ins and restart operations create discontinuities in production sequences and undermine the reliability of conventional forecasting methods. This study aims to develop a state-aware framework for daily gas production forecasting and shut-in identification under such operating conditions. The framework combines a long short-term memory (LSTM) self-attention module for production forecasting with an LSTM autoencoder for unsupervised shut-in identification, and Bayesian optimization is applied separately to tune the hyperparameters of the two modules. The framework was ultimately evaluated using daily production data from representative high-, medium-, and low-production wells in the Sulige tight sandstone gas field. The proposed Bayesian-optimized self-attention LSTM (BSL) framework achieved an average root-mean-square error (RMSE) of 0.175 and an average coefficient of determination (R2) of 0.972 across the three test wells. The corresponding RMSE and R2 values were 0.215 and 0.952 for the low-production well, 0.168 and 0.984 for the medium-production well, and 0.141 and 0.981 for the high-production well. In the optimization algorithm ablation experiment, BSL outperformed the random search and grid search variants, achieving shut-in identification accuracies of 85.21%, 90.36%, and 91.19% for the low-, medium-, and high-production wells, respectively. These results demonstrate that the framework can characterize long-term temporal dependencies and abnormal shut-in states within a unified workflow. The proposed framework thus provides a data-driven tool for production forecasting and well operational state identification in tight sandstone gas reservoirs.