Moment-Based Probabilistic Prediction of Bike Availability for Bike-Sharing Systems

We study the problem of future bike availability prediction of a bike station through the moment analysis of a PCTMC model with time-dependent rates. Given a target station for prediction, the moments of the number of available bikes in the station at a future time can be derived by a set of moment equations with an initial set-up given by the snapshot of the current state of all stations in the system. A directed contribution graph with contribution propagation method is proposed to prune the PCTMC to make it only contain stations which have significant contribution to the journey flows to the target station. The underlying probability distribution of the available number of bikes is reconstructed through the maximum entropy approach based on the derived moments. The model is parametrized using historical data from Santander Cycles, the bike-sharing system in London. In the experiments, we show our model outperforms the classic time-inhomogeneous queueing model on several performance metrics for bike availability prediction.