BiciMAD: Bike-sharing system dataset creation and preliminary hourly demand prediction using machine and deep learning approaches

This project consists on the construction of a database for the BiciMAD bike sharing system putting together several relevant data sources containing information such as weather, festivities, and bike station demand among others, as well as realizing a preliminary prediction study with 2 state of th...

Descripción completa

Detalles Bibliográficos
Autor: Roset Cardona, Javier
Tipo de recurso: tesis de maestría
Fecha de publicación:2024
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/419563
Acceso en línea:https://hdl.handle.net/2117/419563
Access Level:acceso abierto
Palabra clave:Machine learning
Databases
Bicycle sharing programs
Supply and demand
Bike sharing system
Graph Convolutional Neural networks (GCNN)
XGBOOST
Data science
Data processing
BiciMAD
Aprenentatge automàtic
Bases de dades
Serveis de bicicletes públiques
Oferta i demanda
Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial::Aprenentatge automàtic
Descripción
Sumario:This project consists on the construction of a database for the BiciMAD bike sharing system putting together several relevant data sources containing information such as weather, festivities, and bike station demand among others, as well as realizing a preliminary prediction study with 2 state of the art models. This database is tailored so that it is easy to explore different predictive techniques in order to optimize the prediction of future bikes demand in all the network. To build the database, the following sources have been used; BiciMAD database, containing the status of each bike docking station for every day and hour as well as the recorded bike trip transactions from January 2019 to December 2022 and DatosAbiertos Madrid, specifically Sistema Integral de Calidad del Aire del Ayuntamiento de Madrid, containing several weather variables such as temperature, precipitations, or wind speed in different locations of Madrid (weather stations) at each hour. Also, from DatosAbiertos we obtained information about the working days and holidays. The final dataset consists on a table where each row represents the state of a specific station in a specific hour, being the number of rows of the table stations × hours even though the number of stations was dynamic over time. For the preliminary prediction study over the dataset we have used XGBOOST optimized with Optuna and Graphic Convolutional Neural Networks in a static graph configuration, achieving higher performance for XGBOOST even though in the literature usually GCNN has better performance. As it was a preliminary study, we have not exploited the complete power of each solution. We have analyzed the model performance in several environments such as areas with different characteristics (office areas, residential areas and parks), or time spans with different weather (winter, spring, summer, and rainy days) and also special events such as holidays and work days, observing how the models react to those environments; in most cases capturing the effects present in each test scenario.