Dynamic bicycle scheduling problem based on short-term demand prediction

As a low-cost environmentally-friendly travel mode, public bicycles have been widely applied in many large cities and have greatly facilitated people’s daily lives. However, it is hard to find bicycles to rent or places to return at some stations in peak hours due to the unbalanced distribution of public bicycles. And the traditional scheduling methods have hysteresis, in general, the demands might have changed when the dispatch vehicle arrives the station. To better solve such problems, we propose a dynamic scheduling (DBS) model based on short-term demand prediction. In this paper, we first adopt K-means to cluster the stations and adopt random forest (RF) to predict the check-out number of bikes in each clustering. In addition, the multi-similarity inference model is applied to calculate the check-out probability of each station for check-out prediction, and a probabilistic model is proposed for check-in prediction in the cluster. Based on the prediction results, an enhanced genetic algorithm (E-GA) is applied to optimize the bicycle scheduling route. Finally, we evaluated the performance of the models through a one-year dataset from Chicago’s public bike-sharing system (BSS) with more than 500 stations and over 3.8 million travel records. Compared with other prediction methods and scheduling approaches, the proposed approach has better performance.