Optimizing Sustainable Public Transport: A Graph Convolutional Model with Skip Connections for Passenger Volume Prediction

Accurately predicting passenger flow at bus stops is essential for optimizing public transport operations, reducing energy consumption, and enhancing urban mobility. This paper presents a novel approach to forecasting passenger activity—specifically tap-ins and tap-outs—at various bus stops in Singapore. While traditional models such as Multi-Layer Perceptrons (MLPs) and Graph Convolutional Networks (GCN) yielded suboptimal results, our proposed model combining GCN layers with skip connections and linear layers significantly improved prediction accuracy. This model captures the complex patterns in passenger flow more effectively by integrating both graph-based learning and feature transformations. Extensive experiments were conducted, analyzing the model's performance across various time periods and features. The results demonstrate the model's ability to outperform traditional models, hence supporting smarter resource allocation, potentially reducing fuel consumption through better route planning and passenger demand management. Lasso regression was also used to analyze the importance of the different features of the bus stops. The results can be used for city planning and further highlight that our proposed model is better suited for the passenger volume prediction task.