Demand forecasting of micro mobility using a gated recurrent unit

Abstract

Mobility as a service (MaaS) aims to integrate moving, which has been segmented by the combination of various means of transportation, and payment systems as a service and provide door-to-door services to users. Providing complete door-to-door MaaS services to users requires advances in the level of transportation services to solve the first-last mile problem. Micro mobility, such as public bikes and electric scooters, is drawing attention as a means of transportation that can be used for first-last mile travels. Providing complete door-to-door services to MaaS users requires preemptive responses to the demand for micro mobility as well as the ability to track that demand in real time. This essentially requires the ability to forecast the demand for micro mobility. Unlike existing public transportation, micro mobility does not have a fixed travel route or travel time, which causes large variations in demand within short time frames. Therefore, demand forecasting methods different from those used for existing public transportation are needed. In this study, a hybrid forecasting model was designed to forecast the demand for micro mobility that fluctuates sharply and is irregular, and its accuracy was verified. The hybrid forecasting model was based on a genetic algorithm and deep learning algorithm, and the Seoul public bike service, Ttareungyi, was selected as a representative example of micro mobility. To test the model, the rental and return data for Ttareungyi were predicted at the station with the highest eigenvector centrality. The results show higher forecasting accuracy with the hybrid forecasting model than with single forecasting models. The rental and return data forecast by the hybrid forecasting model could be provided as 'expected rental availability' through a MaaS application, and operators could use the data to efficiently relocate bikes to meet demand.