Dynamic bike sharing traffic prediction using spatiotemporal pattern detection

This study proposes a two-step pattern detection methodology for dynamic bike share station traffic prediction using historic traffic and spatiotemporal characteristics. The model is developed on the 15-minute aggregated Washington, D.C. Capital Bikeshare data to predict bike share station traffic for both shortand long-term horizons ranging from 15 min to 4 h. The results show the prediction accuracy equals 100% for 15-minute, 1-hour, and 2-hour horizons and slightly more than 95% for 3-hour and 4-hour horizons at the system level. Not surprisingly, the prediction accuracy drops at the station level. For 15-minute and 1-hour horizons, the prediction accuracy equals 77% and 82%, and it ranges from 24% to 31% for 2-hour, 3-hour, and 4-hour horizons. The results also show that temporal characteristics contribute more than spatial characteristics in the short-time horizons, but the contribution is flipped for long-time horizons. The proposed models have the capacity to estimate bike share traffic for both shortand long-time horizons in less than 20 s of runtime, which illustrates the practicality of the models in dynamic bike sharing traffic prediction, and the potential of the proposed model to be updated in real-time and incorporate the most recent observations into predictions.