Study on Clustering of Free-Floating Bike-Sharing Parking Time Series in Beijing Subway Stations

Dandan Xu, Yang Bian, Jian Rong, Jiachuan Wang, Baocai YinSustainability 2019 · DOI

TL;DR

This paper tackles the growing problem of free-floating bike-sharing (FFBS) parking congestion around Beijing's subway stations. By clustering FFBS parking time series data using Dynamic Time Warping (DTW), they identify distinct parking patterns and propose a two-level early-warning system to help city managers better regulate this increasingly vital urban transport mode.

What Problem Does This Paper Address?

Free-floating bike-sharing (FFBS) systems have exploded in popularity, offering a convenient "first/last mile" solution for urban commuters. However, this rapid growth has led to significant challenges, particularly concerning parking. Bikes are often concentrated around high-demand areas like subway stations, leading to cluttered sidewalks, pedestrian obstruction, and general urban disorder. Existing management strategies are often reactive and lack a nuanced understanding of the underlying parking dynamics. This paper aims to provide a data-driven approach to understand these dynamics, specifically focusing on the temporal patterns of FFBS parking around subway stations in Beijing.

Methods

The authors employ a multi-stage approach:

  1. Time Series Construction: They create time series data representing the cumulative parking quantity of FFBS at each subway station. This is derived from rental and return data, assuming an initial parking quantity of zero at the start of the day.
  2. Data Preprocessing: To improve clustering accuracy, the raw time series data is filtered to reduce noise and then normalized to bring all series to a comparable scale.
  3. Parking Intensity Classification: Subway stations are first classified into seven levels based on their FFBS parking intensity, calculated as the maximum absolute value of the peak demand or supply within their time series. This provides a spatial overview of parking hotspots.
  4. Hierarchical Clustering with DTW: The core of their methodology involves clustering the preprocessed time series data. They opt for a hierarchical clustering approach using Dynamic Time Warping (DTW) as the similarity measure. DTW is chosen over simpler methods like Euclidean distance because it can effectively handle time series of unequal lengths and is robust to temporal distortions, making it suitable for capturing the nuanced shapes of parking patterns.
  5. Feature Extraction and Analysis: After clustering, they analyze the characteristics of each cluster by examining peak and off-peak periods, identifying trends like continuous descending (demand consistently exceeding supply), concave, convex, and irregular patterns.
  6. Early-Warning Index Development: Finally, they propose a two-level early-warning index (T) that considers both the real-time parking quantity (P) and the estimated parking capacity (Cs) around subway stations. This index is designed to trigger specific management responses.

Datasets

The study utilizes FFBS rental and return data from eight FFBS operators for 297 subway stations in Beijing, collected from June to August 2018. The data is granular, recorded every 5 minutes. The authors selected data from August 1, 2018, for analysis due to observed consistency in parking time series across different working days.

Key Insights and Ideas

  • Parking Patterns are Diverse and Spatially Correlated: The clustering reveals that FFBS parking patterns around subway stations are not uniform. Stations with high intensity are concentrated in important transit hubs and city center locations, while suburban stations show lower intensity, likely due to operational challenges and limited demand.
  • Demand Outstrips Supply in Many Stations: A significant finding is that the majority of stations fall into the "continuous descending" category, indicating that FFBS demand consistently exceeds supply. This suggests that the parking problem is often driven by excessive bike redistribution (scheduling) rather than simply user parking behavior.
  • DTW is Effective for Temporal Parking Pattern Analysis: The use of DTW for time series clustering proves effective, achieving a high purity score and outperforming Euclidean distance in capturing the complex temporal dynamics of FFBS parking.
  • A Proactive Management Approach is Needed: The paper advocates for a shift from reactive to proactive management. The proposed early-warning index and corresponding strategies aim to prevent congestion before it becomes severe.

Key Results

  • 11 Clusters Identified: The hierarchical clustering successfully grouped the 297 subway stations into 11 distinct FFBS parking time series clusters.
  • High Clustering Purity: The clustering achieved a purity of 0.939, indicating a high degree of accuracy in grouping similar parking patterns. This was notably higher than a previously reported Euclidean distance-based method.
  • Dominance of "Continuous Descending" Pattern: Over half of the stations (162 out of 297) exhibited a "continuous descending" parking pattern, highlighting a widespread issue of supply not meeting demand.
  • Two-Level Early-Warning System: The developed index T provides a quantitative measure for parking congestion, with thresholds (T < 1.00, 1.00 ≤ T < 1.67, T > 1.67) triggering different management actions, from no intervention to manual arrangement or dispatching.

Limitations and Future Work

The authors acknowledge several limitations:

  • Data Access: The proposed framework could not be fully verified on a real-time FFBS monitoring platform due to limitations in accessing open data.
  • In-depth Cause Analysis: While clustering results are presented, a detailed exploration of the specific causes behind each cluster's pattern and the implications for management is deferred to future work.

Future research could focus on integrating the proposed framework with actual monitoring platforms, conducting more in-depth qualitative analyses of the clustering results, and exploring the causal relationships between urban planning, user behavior, and FFBS parking dynamics.

Summary updated

Abstract

In recent years, the free-floating bike-sharing (FFBS) system has become a significant mode of travel to satisfy urban residents’ travel demands. However, with the rapid development of FFBS, many problems have arisen, among which the parking problem is the most prominent. To solve the FFBS parking problem around urban subways, firstly, the time series of FFBS parking pattern and subway station classification in Beijing were constructed based on parking intensity, showing a significant spatial distribution of subway stations with different intensity levels. Second, a hierarchical clustering method based on dynamic time warping (DTW) was proposed to cluster the FFBS parking time series. Subway stations in Beijing were grouped into 11 clusters, and the clustering purity reached 0.939, which achieved the expected effect. Then, the peak and off-peak period features of time series were extracted to discuss the clustering results. Finally, a two-level early-warning index for monitoring FFBS was constructed, which took the real-time parking quantity and land use capacity of FFBS into consideration. And FFBS parking management strategies for different early-warning indices were put forward. It is very important for the sustainable development of FFBS and cities.
BibTeX
@inproceedings{xu2019study,
  author = {Dandan Xu and Yang Bian and Jian Rong and Jiachuan Wang and Baocai Yin},
  title = {{Study on Clustering of Free-Floating Bike-Sharing Parking Time Series in Beijing Subway Stations}},
  booktitle = {Sustainability},
  year = {2019},
  doi = {10.3390/su11195439},
}
Figure