Deep Learning for Short-Term Spatiotemporal Prediction of Shared Dockless E-Scooter Demand: Grid-Based Representation, Temporal Input Design, and Day-Type-Aware Model Benchmarking

Shared electric scooter (e-scooter) systems exhibit dynamic spatiotemporal demand patterns, making accurate short-term prediction critical for fleet rebalancing and operations. Despite growing interest in shared micromobility demand prediction, literature still lacks a coherent image-based deep learning framework. This is important because e-scooter demand is naturally distributed over urban space, and image-based models can learn spatial dependencies directly from demand maps. Moreover, prior studies often under-document how raw trip records are transformed into deep-learning-ready spatiotemporal inputs, rely on heuristic historical input structures, give limited attention to day-type variation, and rarely use statistical testing in model comparisons. In this thesis, an image-based deep learning framework for short-term spatiotemporal prediction of shared e-scooter pick-up and drop-off demand is developed in three stages. First, from a relatively small geographical area over a limited temporal span, UNET and UNETR are benchmarked against the literature baseline under matched experimental conditions. Second, the framework is extended to a larger spatiotemporal setting, where a processing pipeline is established to transform raw trip records into hourly grid-based pick-up and drop-off demand images and support reproducible experimentation. Third, building on this, the thesis develops a statistically grounded methodology for temporal input design and conducts a day-type-aware benchmarking study across multiple deep learning models using masked metrics across demand-types, demand-levels, and day-types, with multi-level statistical testing. The results show that image-based encoder-decoder models outperform the benchmark literature model, while correlation- and error-based temporal input design identifies compact time-lag configurations that outperform baseline structures. They also show that the use of transformer attention in the UNET bottleneck yields the most consistent performance across prediction horizons, yet prediction remains less consistent across day-types, particularly on weekends. This persistent performance gap shows that historical demand and architectural enhancements alone are insufficient for modeling across temporal regimes, highlighting the need for contextual information. Overall, this work provides a foundation for reproducible data preparation, temporal input design, and rigorous deep-learning-based prediction of shared e-scooter demand.

Name thesis:degree_name

Doctor of Philosophy (PhD)

Discipline thesis:degree_discipline

Engineering – Electrical & Computer

Grantor

Schulich School of Engineering

Year dc:date.issued

2026

Author dc:creator

• Sahnoon, Mohammad

Advisors dc:contributor.advisor

• Souza, Roberto
• Demissie, Merkebe Getachew

Committee members dc:contributor.committeemember

• Far, Behrouz H.
• Saidi, Saeid
• Saeedi, Sara
• Antunes, Francisco José Nibau
• Yanushkevich, Svetlana

• Deep learning
• E-scooter demand prediction
• Micromobility and sharing mobility
• Spatiotemporal prediction
• Spatiotemporal data processing
• Statistical testing

Statement dc:rights

• Unless otherwise indicated, this material is protected by copyright and has been made available with authorization from the copyright owner. You may use this material in any way that is permitted by the Copyright Act or through licensing that has been assigned to the document. For uses that are not allowable under copyright legislation or licensing, you are required to seek permission.

Language dc:language.iso

en

DOI

https://dx.doi.org/10.11575/PRISM/51291

OAI identifier oai:identifier

oai:ucalgary.scholaris.ca:1880/124561