Short-term inbound passenger flow forecasting for urban rail transit based on phase space reconstruction and deep learning

Abstract

Accurate and real-time passenger flow forecasting is of significant importance for the operation of urban rail transit (URT). However, the chaotic and nonlinear nature of short-term passenger flow limits the predictive performance of traditional time series and deep learning models. Integrating chaos theory and deep learning, this paper proposes an efficient short-term passenger flow forecasting model to address these challenges. First, the Lyapunov exponent of the passenger flow time series is computed to quantify its chaotic nature. Then, Phase Space Reconstruction (PSR) is used to map the onedimensional passenger flow time series to a higher-dimensional space, uncovering its inherent chaotic properties. Using the reconstructed highdimensional data, a Convolutional Neural Network (CNN) abstract features of passenger flow across different dimensions, while a Long Short-Term Memory (LSTM) network captures temporal features. The combined CNN and LSTM architecture, termed PSR-CNN-LSTM enhances predictive efficiency and stability, with the Grey Wolf Optimizer (GWO) optimizing the model's hyperparameters. Experiments on real-world AFC datasets from five representative metro stations in Shanghai (China) validate the model's generalization capability across diverse station types and passenger flow patterns. Compared with five benchmark models, the PSR-CNN-LSTM model achieves higher predictive accuracy, faster convergence speed, and improved computational efficiency. Ablation studies confirm that each component plays a critical role in enhancing forecasting performance. This research provides subway operators with realtime, reliable insights into short-term passenger flow, optimizing passenger flow management and scheduling.