Dynamic linear models to predict bike availability in a bike sharing system

A significant increase in the use of a bike sharing systems (BSSs) can cause imbalances in the distribution of bikes, creating logistical challenges. Moreover, imbalanced BSSs discourage bike riders, who may find it difficult to pick up or drop off a bike at their desired location. The first step to solve this logistical problem is forecasting the availability of bikes at each station in the BSS. Forecasting the bike count is a challenging task because the developed models have to consider the change in the demand patterns at each station caused by the interaction between users and the BSS. This paper adopted dynamic linear models (DLMs) to predict bike counts in a BSS. We used first- and second-order polynomial models because of their simplicity. Both DLMs were applied to a BSS in the San Francisco Bay Area. Different prediction horizons were used: 15, 30, 45, 60, and 120 minutes. The adopted models successfully captured the temporal evolution in the underlying demand pattern and hence the bike count at BSS stations. The results show that the first- and second-order DLMs are able to predict the bike counts with a prediction error of only 0.37 bikes/station for a 15-minute prediction horizon (corresponding to a percentage error of 2% accounting to the capacity of the station) and 1.1 bikes/station (6%) for a 2-hour prediction horizon. Comparison results show that the DLMs outperformed the least-squares boosting algorithm and were comparable to the performance of a random forest model for 15- and 30-minute prediction horizons.