<p>Accurate prediction of passenger flow and origin-destination (OD) relationships in metro networks is crucial for optimizing transit operations, improving service quality, and enhancing passenger experience. This study evaluates the performance of five machine learning models— Adaptive Feature Fusion Network (AFFN), Convolutional Neural Network (CNN), Support Vector Machine (SVM), Random Forest, and Relational Graph Convolutional Network (RGCN)—in forecasting metro passenger flows. A comprehensive dataset consisting of historical ridership records, temporal attributes, and spatial characteristics was utilized for model training and evaluation. Advanced preprocessing techniques, including outlier detection via Z-score and Isolation Forest, as well as data normalization, were implemented to improve model robustness. The temporal dependencies and seasonality in passenger demand were captured through time-series decomposition and deep learning architectures. Performance was assessed using multiple evaluation metrics, including accuracy, root mean square error (RMSE), and mean absolute error (MAE), to ensure a holistic comparison of model effectiveness. While deep learning models demonstrated superior predictive accuracy, traditional machine learning methods exhibited advantages in computational efficiency and interpretability. A comparative analysis with statistical forecasting models such as ARIMA and SARIMA was also conducted to benchmark the machine learning models against conventional approaches. The findings provide valuable insights for transportation planners and metro authorities in deploying AI- driven solutions for demand forecasting and operational planning.</p>