| Author: | Li, Xinyu |
| Title: | Deep learning for short-term bike sharing demand forecast : spatial temporal usage characterization and feature fusion strategies |
| Advisors: | Xu, Yang (LSGI) Shi, Wenzhong (LSGI) |
| Degree: | Ph.D. |
| Year: | 2022 |
| Subject: | Bicycle sharing programs Bicycle sharing programs -- Forecasting Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Land Surveying and Geo-Informatics |
| Pages: | xv, 136 pages : color illustrations |
| Language: | English |
| Abstract: | With more than half a century of development, bike-sharing services have gradually become an important travel mode in many cities worldwide. With the characteristics of flexibility and convenience, shared bicycles are increasingly utilized by citizens for short-distance trips in cities. With the growing usage of shared bicycles, bike-sharing services encounter several daily operation issues, such as a spatial mismatch between users' demand and bicycle supply when operating. Accurate demand prediction facilitates the effective rebalancing and management of shared bicycles in cities. Short-term bike sharing demand prediction is the process used to estimate the possible travel demand for shared bicycles in one or few hours. Many studies have concentrated on developing deep learning architectures to accurately estimate the bicycle demand throughout the city. Based on the deep learning technology, this thesis investigates the impacts of the model architecture design, the strategies for capturing and fusing spatial-temporal features of historical observations, and built environment characteristics in urban areas on the performance of deep learning prediction models. First, this study compares the performance of proposed hybrid deep learning architectures with a collection of classic time-series prediction models across dockless and station-based bike-sharing systems in four cities. This study also investigates the effect of strategies for training data selection on the model performance. Second, this thesis proposes irregular convolution to leverage the hidden linkage between semantic neighbors that refer to non-spatially adjacent urban areas exhibiting similar bicycle usage patterns. Third, based on the concept of semantic neighbors, the similarity of BE characteristics is further incorporated with bicycle usage similarity to guide the semantic neighbor selection. At last, all proposed strategies and prediction models in this thesis are evaluated across two types of bike-sharing systems and over different cities to assess the robustness and reliability of the proposed models. The main contributions of this thesis include: 1) the proposed hybrid deep learning models outperform the classic time-series prediction models over different bike-sharing systems and cities. In addition, the strategies for training data selection significantly affect the performance of deep learning prediction models; 2) the prediction accuracy is further improved by adopting irregular convolution that leverages the spatial relationships between non-spatially adjacent urban areas with similar bicycle usage patterns; 3) the similarity of BE characteristics is a significant indicator to guide the semantic neighbor selection and further affect the model performance; 4) all proposed strategies and prediction models achieve robust and reliable performance over different bike-sharing systems and across different cities, which facilitates the reproducibility and replicability of hybrid deep learning prediction models. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/12196