Demand Forecasting with Machine-Learning Models for Bike-Sharing-Systems based on Open Data

Abstract

The utilization of climate-friendly and cost-efficient bike-sharing systems (BSS) is gaining worldwide acceptance. For operators, rental demand constitutes a pivotal factor in operational and strategic decision-making, generating high-quality forecasts of this demand remains challenging. This study evaluates novel machine learning (ML) architectures, which have not yet been applied in the BSS domain, using open data. The study’s findings indicate that Transformer and Long Short-Term Memory (LSTM) models demonstrate superiority in terms of forecast accuracy when compared to other models, including DLinear, Temporal Convolutional Network (TCN), and Timeseries Dense Encoder (TiDe). Additionally, the study underscores the utility of open historical data sources in deriving pertinent features associated with BSS demand.