Análisis del Servicio BiciMAD mediante Minería de Datos
Loading...
Official URL
Full text at PDC
Publication date
2026
Authors
Advisors (or tutors)
Editors
Journal Title
Journal ISSN
Volume Title
Publisher
License

Citation
Abstract
El presente Trabajo Fin de Máster aborda el análisis del sistema público de bicicletas eléctricas compartidas BiciMAD, gestionado por la Empresa Municipal de Transportes de Madrid. El trabajo se enmarca en el ámbito de la movilidad urbana sostenible y, en particular, en el problema operativo del rebalanceo de estaciones, entendido como la gestión predictiva de la disponibilidad de bicicletas y anclajes a lo largo del día para que el servicio resulte fiable para el usuario en cualquier momento y ubicación. Partiendo de los datos públicos disponibles a través de la API GBFS de la EMT, se construye un pipeline integrado de minería de datos que combina varias técnicas complementarias: el análisis exploratorio de los patrones temporales y espaciales del sistema, la descomposición de series temporales para caracterizar tendencia y
estacionalidad, el clustering con métrica Dynamic Time Warping para segmentar las estaciones según su perfil funcional a lo largo del día, la detección de anomalías operacionales mediante consenso entre Isolation Forest y Local Outlier Factor, el modelado predictivo con LightGBM acompañado de explicabilidad SHAP, y una referencia univariada SARIMA que permite contrastar empíricamente el valor del enfoque multivariado. El objetivo último del trabajo no es solo demostrar el funcionamiento técnico de cada uno de estos métodos sobre datos reales, sino traducir sus resultados en conocimiento accionable para el operador del sistema: identificar las estaciones más críticas, las franjas horarias en las que el rebalanceo es más efectivo, y las variables que mejor anticipan los estados de vaciado o saturación.
This Master’s Thesis analyzes the BiciMAD public electric bike-sharing system, managed by the Madrid Municipal Transportation Company. The thesis falls within the field of sustainable urban mobility and, in particular, addresses the operational challenge of station rebalancingdefined as the predictive management of bicycle and docking station availability throughout the day to ensure the service remains reliable for users at any time and location. Using public data available through the EMT’s GBFS API, an integrated data mining pipeline is constructed that combines several complementary techniques: exploratory analysis of the system’s temporal and spatial patterns, time series decomposition to characterize trends and seasonality, clustering using the Dynamic Time Warping metric to segment stations according to their functional profile throughout the day, detection of operational anomalies through consensus between Isolation Forest and Local Outlier Factor, predictive modeling with LightGBM accompanied by SHAP explainability, and a univariate SARIMA benchmark that allows for empirical validation of the value of the multivariate approach. The ultimate goal of this study is not only to demonstrate the technical performance of each of these methods using real-world data, but also to translate their results into actionable insights for the system operator: identifying the most critical stations, the time slots during which rebalancing is most effective, and the variables that best predict states of underload or overload. In doing so, the study aims to provide a reproducible methodological foundation for evolving rebalancing from a reactive approachtriggered by an incident that has already occurredtoward a preventive approach based on short-term prediction.
This Master’s Thesis analyzes the BiciMAD public electric bike-sharing system, managed by the Madrid Municipal Transportation Company. The thesis falls within the field of sustainable urban mobility and, in particular, addresses the operational challenge of station rebalancingdefined as the predictive management of bicycle and docking station availability throughout the day to ensure the service remains reliable for users at any time and location. Using public data available through the EMT’s GBFS API, an integrated data mining pipeline is constructed that combines several complementary techniques: exploratory analysis of the system’s temporal and spatial patterns, time series decomposition to characterize trends and seasonality, clustering using the Dynamic Time Warping metric to segment stations according to their functional profile throughout the day, detection of operational anomalies through consensus between Isolation Forest and Local Outlier Factor, predictive modeling with LightGBM accompanied by SHAP explainability, and a univariate SARIMA benchmark that allows for empirical validation of the value of the multivariate approach. The ultimate goal of this study is not only to demonstrate the technical performance of each of these methods using real-world data, but also to translate their results into actionable insights for the system operator: identifying the most critical stations, the time slots during which rebalancing is most effective, and the variables that best predict states of underload or overload. In doing so, the study aims to provide a reproducible methodological foundation for evolving rebalancing from a reactive approachtriggered by an incident that has already occurredtoward a preventive approach based on short-term prediction.







