Use of Deep Learning to Predict Daily Usage of Bike Sharing Systems

https://doi.org/10.1177/03611981188013542018PDFJournal ArticleTRB18-03033Bicycle travelBicyclesMachine learningMathematical predictionVehicle sharingbike sharepredictive modelingdeep learningconvolutional neural networkNew York (New York)United StatesConnected Cities for Smart Mobility toward Accessible and Resilient Transportation Center (C2SMART)Yang, HongXie, KunOzbay, KaanMa, YifangWang, ZhenyuConnected Cities for Smart Mobility toward Accessible and Resilient Transportation Center (C2SMART)US Transportation CollectionBike sharing systems are increasingly deployed in many cities worldwide. These systems provide a prominent solution for the first/last-mile problem with their cost-effectiveness and eco-friendliness. However, the use of bikes among stations is often spatiotemporally imbalanced, causing many problems in daily operations (e.g., rebalancing challenges). Thus, predictively knowing how the system demand evolves in advance helps improve the preparedness of operational schemes. This paper aims to present a predictive modeling approach to analyze the use of bicycles in bike sharing systems. Specifically, a deep learning approach using the convolutional neural networks (CNN) was proposed to predict the daily bicycle pickups at both city and station levels. A numerical study using data from the Citi Bike System in New York City was performed to assess the performance of the proposed approach. Other than the historical records, relevant information like weather was also incorporated in the modeling process. The modeling results show that the proposed approach can achieve better predictive performance in both city-and station-level analyses, confirming the merits of the proposed method against other baseline approaches. In addition, including information of neighboring stations into the models can help improve the performance in station-level prediction. The predictive performance of the CNN was also found to be related to parameters such as temporal window, number of neighboring stations, learning ratio, patch size, and inclusion of addition data such as drop-offs. Thus, the implementation of the proposed models requires necessary calibration to determine appropriate parameters for a given bike sharing system.82