Bike-Sharing Demand Prediction at Community Level under COVID-19 Using Deep Learning

Area of Science:

• Transportation Science
• Data Science
• Urban Planning

Background:

• Efficient bike-sharing service planning requires accurate short-term demand prediction to ensure optimal bike allocation.
• Montreal's bike-sharing network presents a complex system for demand forecasting.

Purpose of the Study:

• To forecast shared bike demand 15 minutes ahead in Montreal using deep learning.
• To evaluate the effectiveness of various LSTM-based architectures and compare them against a benchmark model.

Main Methods:

• Community detection using the Louvain algorithm to segment the bike-sharing network.
• Development and application of four LSTM-based deep learning architectures for demand prediction.
• Utilizing historical trip data (2017-2021), weather conditions, temporal variables, and engineered features.
• Comparison with a univariate ARIMA model as a benchmark.

Main Results:

• Deep learning models demonstrated superior performance compared to the ARIMA benchmark.
• The hybrid Convolutional Neural Network-LSTM (CNN-LSTM) architecture achieved the highest prediction accuracy.
• Incorporating additional features like weather and temporal data significantly enhanced model performance.

Conclusions:

• Deep learning approaches, especially hybrid CNN-LSTM, are highly effective for short-term bike demand forecasting.
• Enriching models with diverse data inputs provides deeper insights into demand patterns for operational management.
• Accurate forecasting supports efficient bike-sharing system operations and user satisfaction.