Accurate identification of deep carbonate microfacies is crucial for reservoir characterization and sweet spot prediction. Deep carbonate reservoirs usually exhibit complex compositions and numerous microfacies types, leading to dramatic challenges and low accuracy in microfacies identification. This study employs conventional logging curves, elemental mud logging curves, and processed mineral interpretation logs from deep carbonate reservoirs as input. A Residual Long Short-Term Memory (ResLSTM) network-based supervised model is developed to establish nonlinear mapping relationships between logging data and carbonate microfacies for intelligent reservoir microfacies identification. The test results demonstrate that: (1) The residual structure incorporated in the ResLSTM network effectively mitigates gradient vanishing and explosion issues during network training. Compared with traditional LSTM networks, the proposed ResLSTM achieves over 10% improvement in prediction accuracy for deep carbonate microfacies. (2) For thin interbedded layers within the reservoir, the ResLSTM model achieves 92.4% microfacies prediction accuracy, demonstrating its strong robustness. These findings highlight the ResLSTM's superior capability in handling the heterogeneity and complex patterns inherent in deep carbonate reservoirs. Furthermore, the tests also demonstrate that the distribution of training data exerts a significant influence on prediction accuracy of ResLSTM. Specifically, in scenarios of input data imbalance, the ResLSTM tends to develop a pronounced predictive bias toward the majority lithofacies categories due to their numerical dominance in the training set. The systemic bias introduced by imbalanced lithofacies distributions presents a critical challenge in petrophysical machine learning applications, demanding urgent methodological innovations to enhance model generalizability across minority facies classes.