关键词: 城市交通, 客流预测, 概率预测, 城轨站点客流, 多源数据, 不确定性量化

Abstract: This paper proposes a short-term probabilistic model for forecasting passenger flow in urban rail transit networks, termed PD-STGCN, which considers the uncertainties of both data and model. It aims to obtain the probabilistic information of network-wide passenger volume at future time steps. The model constructs a collaborative framework consisting of the Spatio Temporal Uncertainty Prediction Module (STUPM) and Probabilistic Quantification Module (PQM). In the STUPM, addressing uncertainties inherent in both the passenger flow data and forecasting model, a novel loss function is developed by integrating Gaussian Negative Log-Likelihood (GNLL) and Monte Carlo Dropout (MC Dropout) techniques to quantify these dual uncertainties. Within the PQM, the discrete sample sets are obtained through the random normal sampling based on prediction results, and continuous probabilistic forecasting outputs are generated using the Gaussian Kernel Density Estimation (KDE). Using the data of passenger flow from a urban rail transit system in a large city as a case study, the model is validated under both weekday and non-weekday scenarios. The results demonstrate that, compared to the baseline forecasting models, PD-STGCN improves the Prediction Interval Coverage Probability (PICP) and Continuous Ranked Probability Score (CRPS) by 8.01% and 20.77%, respectively, which provides a better coverage of actual passenger flow values and has higher forecasting accuracy. Ablation experiments confirm that uncertainty is the most significant factor affecting model performance. The model considering dual uncertainties leads to the improvements of at least 1.91% in PICP and 4.02% in CRPS over that considering only a single type of uncertainty.

Key words: urban traffic, passenger flow prediction, probability prediction, urban rail transit station passenger flow, multi-source data, uncertainty quantification