Dublin City University
Learning-based Methods for Optimising Shared Mobility Systems with Multimodal Data
Abstract
dc:description.abstractThis thesis explores the use of learning-based methods in Shared Mobility Systems (SMS), utilising multimodal data to address three key operational challenges: improper parking behaviour, energy consumption prediction, and pollution-aware routing. The overarching goal is to improve the efficiency, sustainability, and user experience of SMS through data-driven, task-specific solutions. The first challenge is addressed by developing U-Park, a user-centric parking recommendation system. It predicts trip destinations and parking availability in real time using multimodal inputs, including partial trip data, GPS trajectories, and environmental features. Combining an attention-based RNN and a contextualised parking model, U-Park improves the chances of finding available parking by up to 29.66%. The second contribution focuses on privacy-aware energy consumption modelling for shared battery electric vehicles. A Federated Learning (FL) framework enables model training across distributed data sources without sharing raw data. FL algorithms and local models are evaluated on multimodal features such as speed, altitude, and derived variables. The proposed FedAvg-LSTM model reduces mean absolute error by up to 67.84% and supports deployment in edge-cloud environments. For the third challenge, a pollution-aware route planning system is introduced. Multimodal data from fixed and mobile air quality sensors is used to construct a high-resolution PM2.5 map, combining temporal imputation with spatial interpolation. Models including IDW, RF, LSTM, and Conv-LSTM are evaluated for short-term forecasting. The resulting pollutant maps inform route selection, reducing average exposure by 25.88% with minimal extra travel distance. These contributions highlight the value of integrating multimodal data and adopting tailored learning approaches. The thesis also discusses challenges such as data sparsity, integration uncertainty, and model explainability, and outlines future directions including ensemble learning, uncertainty-aware modelling, and multi-objective optimisation.
Degree
thesis:*- Name dc:type.qualificationname
- phd
- Level dc:type.qualificationlevel
- doctoral
- Grantor dc:publisher.institution
- Dublin City University
- Year dc:date.issued
- 2025
Author and committee
dc:creator, dc:contributor.*- Author dc:creator
-
- Yan, Sen
Rights
- Language dc:language
- en