Origin-destination demand prediction for shared mobility service using fully convolutional neural network

Emerging on-demand shared mobility services face significant challenges in balancing demand effectively. The rapid expansion of these services necessitates precise origin-destination demand prediction to optimize fleet management, operational efficiency, and seamless multimodal integration within evolving urban transportation systems. Our previous work addressed this issue using a Masked Fully Convolutional Network (MFCN) model for short-term pick-up/drop-off demand forecasting. In this study, we introduce a predictive modeling framework for short-term origin-destination demand prediction, leveraging Convolutional Neural Networks (CNNs) and integrating our MFCN model. We further enhance the framework with novel prediction fusion and scaling methodologies to improve accuracy. Additionally, we propose a new loss function designed to effectively train the model by incorporating both demand volume and spatial location information. To evaluate the framework, we applied it to shared e-scooter trip data from Calgary, Canada, testing two prediction scenarios: next-hour and next-24-h forecasts. The model's performance was benchmarked against baseline approaches, including a naïve predictor, linear regression, Graph Convolutional Networks (GCN), and other variant models. Our results demonstrate that the proposed model outperforms all baselines in terms of true positive and F1-score values, highlighting its effectiveness in predicting demand. Furthermore, the high degree of regularity in daily mobility patterns suggests that shared e-scooter demand is predictable over a 24-h period. However, when considering spatial error, the performance difference between the two prediction schemes is reduced. This study not only enhances predictive modeling for shared mobility services but also contributes to the broader discourse on transformative mobility patterns. By integrating predictive analytics into Mobility-as-a-Service (MaaS) ecosystems, this approach can facilitate dynamic fleet rebalancing, inform data-driven policy decisions, and support the transition toward more sustainable and adaptive urban transportation systems.