Description

Bike sharing systems have become increasingly popular around the world for its benefits in public health, environment, and urban mobility. Managing the operations of such systems, however, requires a good understanding of bike usages. In fact, the prediction of bike demands is a critical component of many operational problems, including downstream bike re-balancing efforts. In recent years, there is a growing interest in adopting the cluster-then-predict approach, whereby bike stations are grouped into clusters first and bike demand predictions are made later for each cluster. This thesis project has two main objectives: first, examining how various station-to-cluster assignments may impact the bike check-out and check-in predictions, and second, understanding how transition information may influence the bike check-in predictions. We present a five-step cluster-then-predict framework that generates station clusters using a state-of-the-art clustering algorithm, makes check-out predictions, computes cluster-to-cluster transition matrices, makes check-in predictions, and evaluates the predictions.