Analysis of urban rail transit traffic flow based on SARIMA and LSTM models

Urbanization has driven urban rail transit to become a pivotal solution for alleviating traffic congestion. As a critical infrastructure in China's national central city, Xi'an Metro's passenger flow forecasting holds significant implications for operational optimization. This study develops a hybrid forecasting model by integrating the time-series SARIMA model and the neural network LSTM model, analyzing daily total passenger flow data of Xi'an Metro from February 1, 2023, to December 1, 2024.Key findings reveal that: SARIMA effectively captures data trends and periodicity through differencing and a 7-day seasonal component (optimal parameters: SARIMA(1,1,1)(0,1,7)). LSTM demonstrates proficiency in handling long-term dependencies via its gating mechanisms and cell states (showcasing robust generalization after normalization and early stopping optimizations), though its contribution to nonlinear pattern mining is relatively limited within this dataset. Given the inherent limitations of individual models—where SARIMA excels in linear relationships while LSTM specializes in nonlinear dynamics with constrained efficacy—the study employs linear regression weighting to fuse the two models. The optimized hybrid model assigns weights of 0.8004 to SARIMA and 0.1956 to LSTM, yielding superior performance. The combined model achieves lower RMSE and MAE metrics on the training set, exhibiting enhanced prediction stability and accuracy compared to standalone models. These advancements provide scientific support for passenger flow management, scheduling optimization, and the development of intelligent metro systems. Future research could further refine the model by incorporating real-time data.