<p>Tight carbonate gas reservoirs are widely distributed in unconventional hydrocarbon resources and exhibit considerable potential for exploration and development. However, these reservoirs commonly display strong heterogeneity in both core observations and well log responses, which makes accurate lithofacies identification difficult and constrains subsequent fine reservoir characterization and evaluation. To address this issue, this study proposes an intelligent lithofacies identification method for tight carbonate gas reservoirs by integrating dynamic clustering with a BiLSTM-Attention network. First, standardized preprocessing is applied to mitigate the interference caused by differences in the magnitudes of well logging parameters during the learning process. Second, a dynamic clustering algorithm is employed to perform unsupervised classification of well log data, enabling automatic lithofacies labeling. Finally, a BiLSTM-Attention model is constructed, in which the attention mechanism is introduced to capture the temporal dependencies among well log curves and achieve intelligent lithofacies identification. Validation experiments are conducted using data from the Southwest Kansas Panoma gas field in North America. The results demonstrate that the proposed method outperforms other comparative models in terms of evaluation metrics, sample labeling performance, single well identification, and model generalization. It significantly improves the overall accuracy of lithofacies identification and enhances the adaptability to lithofacies variations under complex geological conditions as well as cross-well generalization capability. This study provides effective technical support for lithofacies classification and identification in tight carbonate reservoirs and contributes to subsequent reservoir classification evaluation and development optimization of unconventional gas reservoirs.</p>