关键词: 地铁客流量预测, Informer模型, 图卷积网络, 膨胀因果卷积, 注意力机制, 频率增强信道

Abstract: Given its advantages of safety, convenience, punctuality and comfort, the subway has become a primary mode of transportation in people’s daily lives. As urban development continues, the pressure on subway passenger flow increases daily. Oversaturation of passenger flow, caused by various factors, leads to safety hazards such as station congestion, reduced operational efficiency, and even crowd crush incidents. These issues significantly affect passenger safety and the operational management of transportation departments. Therefore, accurate passenger flow prediction is crucial for supporting reasonable travel arrangements and informed decision-making by subway operation and management departments.
Passenger flow data exhibits both temporal dependence and strong spatial correlations, along with clear periodic patterns. However, existing prediction models struggle to balance long-term and short-term dependencies when addressing spatial-temporal correlations and periodicity effectively. Consequently, this paper develops a combinatorial prediction model for subway passenger flow, integrating dual-dimensional temporal and spatial information while considering periodic characteristics. Theoretically, this study aims to accurately predict passenger flows at subway stations, enhancing the accuracy of current models. Practically, the findings can support governmental decisions regarding subway construction, assist transportation management in resources allocation, minimize resource wastage, and offer the public travel planning references.
Initially, existing passenger flow forecasting technologies are reviewed on the Internet, their advantages and disadvantages are analyzed, and insights are gained to inform the design of our research plan. Subsequently, the card swiping records of Hangzhou Metro and the adjacency matrix of Hangzhou Metro transportation network, provided by the Hangzhou Municipal Public Security Bureau, are selected as the dataset for this study. Various factors affecting passenger flow are identified. We choose to consider and collect relevant data from three aspects: travel mode, weather conditions, and historical data at corresponding time points, to construct relevant features. Finally, to mitigate any negative impact of the constructed features on the model’s prediction results, LightGBM importance analysis and Pearson correlation analysis are employed to select features, thereby enhancing the model’s generalization ability.
In this paper, we propose a combined model, GCN-DFInformer, for subway passenger flow prediction, which integrates a graph convolutional network (GCN) and an improved Informer model (DCC-FECAM-Informer, DFInformer) in parallel. Firstly, an inflated causal convolutional self-attention mechanism is introduced to compensate for the Informer’s insensitivity to local information and to enhance the model’s ability to capture trend changes and local fluctuations in passenger flow over short periods. Secondly, considering the obvious periodicity and seasonality of subway passenger flow data, the frequency enhanced channel attention mechanism (FECAM) is introduced to improve the model’s ability to identify and utilize inherent features in the data series for better predictions. Finally, the prediction is performed by the parallel-connected graph convolutional network to achieve the fusion of temporal and spatial information. The experiments comparing the model’s predictions with actual values demonstrate that the GCN-DFInformer model exhibits strong predictive performance and robustness. The experimental results from the scenario dataset test demonstrate that, compared to other models, the proposed model yields smaller errors and a higher coefficient of determination (R²). This validates the superior predictive performance of the model and highlights its effectiveness in improving the accuracy of subway passenger flow prediction.
The proposed model significantly enhances prediction accuracy by integrating the Informer, dilated causal convolutional self-attention mechanism, FECAM module, and graph convolutional network. This model leverages the strengths of both GCN and DFInformer, and their parallel structure helps maintain the independence of spatial and temporal information. In the next phase, additional influencing factors related to the station will be incorporated, such as the functional area where the station is located, and other auxiliary information like the subway operation timetable, to further improve the model’s prediction accuracy.

Key words: subway passenger flow forecast, Informer, graph convolutional networks, dilated causal convolution, attention mechanism, frequency enhanced channel