Sensing and Predicting the Pulse of the City through Shared Bicycling

Jon Froehlich, Joachim Neumann, Nuria OliverIJCAI 2009 · OpenReview

Abstract

City-wide urban infrastructures are increasingly reliant on network technology to improve and expand their services. As a side effect of this digitalization, large amounts of data can be sensed and analyzed to uncover patterns of human behavior. In this paper, we focus on the digital footprints from one type of emerging urban infrastructure: shared bicycling systems. We provide a spatiotemporal analysis of 13 weeks of bicycle station usage from Barcelona's shared bicycling system, called Bicing. We apply clustering techniques to identify shared behaviors across stations and show how these behaviors relate to location, neighborhood, and time of day. We then compare experimental results from four predictive models of near-term station usage. Finally, we analyze the impact of factors such as time of day and station activity in the prediction capabilities of the algorithms.

Introduction

Urban planners and social scientists have long struggled to obtain high-resolution, real-world data on how people move through a city. Traditional methods are expensive, slow, and often clash with how streets are actually used. As cities digitize, "digital footprints" from infrastructure—like subways, buses, and shared bikes—offer a new way to sense urban dynamics. For the average user, the problem is more practical: "Will there be a bike available when I leave my house?" or "Will there be an empty slot when I get to work?" This paper tackles both the macro-level understanding of city behavior and the micro-level challenge of predicting station occupancy.

Methods

The authors employed a two-pronged approach:

  1. Spatiotemporal Clustering: To identify shared behaviors across 390 stations, they used hierarchical dendrogram clustering. Crucially, they used Dynamic Time Warping (DTW) as a distance metric instead of standard Euclidean distance. This allowed the model to group stations that have similar usage shapes even if their peak activities are shifted by an hour or two.
  2. Predictive Modeling: They compared four models to predict station availability: Last Value (LV), Historic Mean (HM), Historic Trend (HT), and a Bayesian Network (BN). The BN is a probabilistic model that takes the current time, current bike count, and prediction window as inputs to estimate the "delta" (change in bikes).

Datasets

The study used a dataset scraped from the Barcelona Bicing website over 13 weeks (August 27 to December 1, 2008). The scale included 390 stations and 26.1 million individual observations. After removing noisy data from technical failures or station closures, 20.2 million observations remained. Since stations vary in size (15 to 36 slots), data was normalized to "Normalized Available Bicycles" (NAB), representing the percentage of station occupancy.

Spatial distribution of Bicing stations across Barcelona.

Key Insights

  • Cultural Signatures: The data revealed a distinct "three-pronged spike" in activity during weekdays: morning commute, evening commute, and a unique "lunch spike" at 2 PM, reflecting Spain’s late lunch culture.
  • Topography as a Barrier: Barcelona’s incline significantly impacts bike flow. Stations at higher altitudes (80-110m above sea level) were consistently "starving" for bikes because users prefer riding downhill into the city center but avoid the strenuous ride back up.
  • Predictability vs. Activity: There is an inverse relationship between how busy a station is and how easy it is to predict. The most active "downtown" stations are the hardest to model, while stations with little movement are trivial to predict.
Average bike availability (NAB) at Station 37, comparing weekday and weekend patterns.

Key Results

The Bayesian Network (BN) was the top performer, achieving an average error of just 0.08 NAB. For a typical 25-slot station, this means the model is only off by about two bicycles. When predicting whether a station would be "Full" or "Empty" two hours into the future, the BN and HT models achieved approximately 80% accuracy. For short windows (under 60 minutes), the "Last Value" model is surprisingly robust, but for predictions beyond an hour, the Bayesian Network significantly outperforms simpler methods.

Limitations and Future Work

The current models are "blind" to external factors such as weather (rain or extreme heat) and contextual events (concerts or soccer matches) that create non-standard spikes in demand. Furthermore, the models don't yet integrate data from other transit modes like subways or buses. Future work aims to incorporate these contextual features into a Dynamic Bayesian Network to help users find bikes in real-time via mobile applications.

Summary updated

Abstract

City-wide urban infrastructures are increasingly reliant on network technology to improve and expand their services. As a side effect of this digitalization, large amounts of data can be sensed and analyzed to uncover patterns of human behavior. In this paper, we focus on the digital footprints from one type of emerging urban infrastructure: shared bicycling systems. We provide a spatiotemporal analysis of 13 weeks of bicycle station usage from Barcelona's shared bicycling system, called Bicing. We apply clustering techniques to identify shared behaviors across stations and show how these behaviors relate to location, neighborhood, and time of day. We then compare experimental results from four predictive models of near-term station usage. Finally, we analyze the impact of factors such as time of day and station activity in the prediction capabilities of the algorithms.
BibTeX
@inproceedings{froehlich2009sensing,
  author = {Jon Froehlich and Joachim Neumann and Nuria Oliver},
  title = {{Sensing and Predicting the Pulse of the City through Shared Bicycling}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2009},
  url = {https://openreview.net/forum?id=ByWz3QzuWS},
}
Figure