Station-level demand prediction in bike-sharing systems through machine learning and deep learning methods

Sammanfattning

Public Bike-Sharing systems have been employed in many cities around the globe. Shared bikes are an efficient and convenient means of transportation in advanced societies. Nonetheless, station planning and local bike-sharing network effectiveness can be challenging. This study aims to address these challenges by predicting the station-level demand in public bicycle-sharing systems through the application of machine learning and deep learning methods. This has been done by outlining the influencing urban built environment factors of bike-sharing systems and comparing the machine and deep learning models’ performance prediction results. The Spatial Regression Graph Convolutional Neural Network (SRGCNN) and SRGCNN-Geographically Weighted methods are employed and compared with traditional machine learning methods, namely Multiple Linear Regression, Multilayer Perceptron Regressor, Support Vector Machine, and Random Forest Regressor. For the SRGCNN method, a graph has been constructed based on the number of trips between the stations. Additionally, factors related to the urban environment have been considered for the prediction task. Also, evaluation metrics have been utilized to assess and facilitate the comparison of the machine and deep learning models’ performance. Concluding this case study in Zurich, the deep learning models and Random Forest Regressor were found to outperform other machine learning methods, showcasing better accuracy in terms of several evaluation metrics. Despite recognizing limitations such as the relatively low number of stations, this study highlights future research opportunities to enhance model accuracy and deepen our understanding of micro-mobility demand dynamics, especially in the context of PBS station planning.