This study proposes a novel method for predicting reservoir parameters in carbonate reservoirs, named Fusion Neighborhood Multi - layer Long Short-Term Memory Neural Network (FN-MLSTM). The approach first utilizes Principal Component Analysis (PCA) to extract independent features from well logging data, followed by employing the K-Means algorithm for unsupervised clustering of well groups to optimize their division. Subsequently, by integrating neighborhood information, a Multi-layer Long Short-Term Memory Neural Network (MLSTM) is utilized to accurately predict reservoir parameters. Experimental results demonstrate that on a test set comprising 22 wells in a certain region of southern Sichuan, China, the proposed model outperforms traditional prediction methods such as Long Short-Term Memory Networks (LSTM), Deep Neural Networks (DNN), Extreme Gradient Boosting Trees (XGBoost), and Random Forests (RF) in terms of Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Coefficient of Determination (R2). Additionally, sensitivity analysis results confirm that integrating neighborhood information effectively enhances prediction accuracy.