Bike-Sharing System: Demand Prediction and System Expansion

Abstract

The goal of this thesis is to first, predict bike demand of Citi Bike’s passengers in the New York City, and second, find the optimal locations to place additional bike stations for different purposes. We divide this work into four main steps: data decensoring, demand prediction, search for the spots to put bike stations and experiments. Regarding data decensoring, we apply the method developed earlier by other works and modify it so that it works well with our data. In the second step, we build several polynomial regression models to predict bike demand throughout Manhattan. We apply the concept of Voronoi diagram to spread demand observed at stations to their neighborhoods. It involves many methodologies that we come up with to make the regressions suitable for the nature of bike-sharing system and the habits of bike passengers. In order to select the locations to build bike stations, we propose a greedy method to add bike stations into the system one by one. Finally, we perform simulations of this bike-sharing system after we add bike stations at the locations suggested from earlier parts, assuming that arrival processes at bike stations follow Poisson processes. We count the number of out-of-stock events as well as the length of time periods when bike stations run out of spaces or bikes.