Abstract
Emerging on-demand shared mobility services face the difficulty of effectively balancing demand. Influx of these mobility services urges for more precise prediction of origin-destination demand becomes essential and urgent. Our previous work addressed this issue with a Masked Fully Convolutional Network (MFCN) model for short-term pick-up/drop-off demand prediction. In this study, we present a predictive modeling framework designed for short-term origin-destination demand prediction. This framework harnesses the capabilities of Convolutional Neural Networks (CNNs), integrates our previously developed MFCN model, and introduces novel prediction fusion and scaling methodologies. Furthermore, a new loss function is developed and designed to effectively train the model with demand and location information. We evaluated the proposed framework using shared e-scooter trip data from Calgary, Canada. Our evaluation encompasses two prediction scenarios: next-hour and next-24-hour predictions. The performance of our framework is benchmarked against baseline models including the naïve predictor, linear regression, GCN, and variant models. Our model shows the best performance regarding the true positive and F1-score values. The results suggest a high degree of regularity in the daily demand as the next-24-hour predictor performs better than the other scheme. Nonetheless, when a spatial error is considered, the performances of the two prediction schemes are comparable.