Scalability evaluation of forecasting methods applied to bicycle sharing systems.

Area of Science:

• Urban Mobility
• Data Science
• Transportation Engineering

Background:

• Public Bicycle Sharing Systems (BSS) are increasingly common in urban environments.
• Predictive analysis is crucial for BSS operations, including demand forecasting and bike rebalancing.
• Current BSS algorithm evaluations often lack scalability assessments across different system sizes.

Purpose of the Study:

• To assess the performance of popular prediction algorithms across varying Bicycle Sharing System (BSS) sizes.
• To identify algorithms that provide consistent results regardless of system scale.
• To understand data sufficiency challenges in smaller BSS for predictive modeling.

Main Methods:

• Evaluation of well-established prediction algorithms.
• Testing across three distinct BSS sizes: small (~20 stations), medium (400+ stations), and large (1500+ stations).
• Comparative analysis of algorithm accuracy and reliability based on system scale.

Main Results:

• Prophet and Random Forest demonstrated the most consistent predictive performance across different BSS sizes.
• Smaller BSS (around 20 stations) frequently exhibited insufficient data, hindering robust algorithm performance.
• Algorithm effectiveness is significantly influenced by the volume of historical data available, which correlates with system size.

Conclusions:

• Prophet and Random Forest are recommended for BSS prediction tasks due to their consistent performance.
• Data availability is a critical factor for successful predictive modeling in BSS, particularly for smaller systems.
• Future research should focus on methods to improve predictions in data-scarce BSS environments.