Data pre-integration in graph-transformer networks for enhanced spatiotemporal ridership prediction

Accurate ridership prediction is vital for optimizing urban rail transit operations, yet it remains a complex challenge due to the spatiotemporal dynamics influenced by factors like weather, holidays, and station attributes. This study proposes an advanced prediction model that integrates Graph Convolutional Networks (GCNs) with Transformer-based architectures to address these complexities. Our model incorporates a priori knowledge and comprehensive data pre-processing to capture both spatial dependencies and temporal dynamics effectively. The dataset comprises hourly ridership records, weather data, and ticket type distributions from 329 stations across Beijing’s urban rail network in 2019. We employed periodic clustering and autocorrelation analyses to pre-process the data, which were then used to train our GraphTransformer model. The proposed model outperforms traditional statistical and machine learning models, demonstrating superior performance. Furthermore, our model is resilient to fluctuations during holidays, outperforming baseline models at stations with varying attributes. This study underscores the importance of integrating domain-specific knowledge into deep learning models and highlights the potential of Transformer networks for enhancing the accuracy and robustness of ridership predictions in complex urban transit environments.