Short-term prediction for bike-sharing service using machine learning

Abstract The bike-sharing service has brought many conveniences to citizens and served as an effective supplement to the mass transit system. For docked bike-sharing service, each docking station has the designated location to store bikes and the station could be empty or saturated in different times. Bike-sharing operators generally redistribute bikes between stations by driving trucks according to their experiences which might lead unnecessary human resources It is ineffective for the operators and inconvenient for users to access this service. Therefore, predicting an accurate number of available bikes in the stations is important for both the operators and users. This paper mainly focuses on the short-term forecasting for docking station usage in a case of Suzhou, China. Two latest and highly efficient models here, LSTM and GRU, are adopted to predict the short-term available number of bikes in docking stations with one-month historical data. Random Forest is used to compare as a benchmark. The results show that both RNNs (LSTM and GRU) and Random Forest able to achieve good performance with acceptable error and comparative accuracies. Random forest is more advantageous in terms of training time while LSTM with complex structures can predict better for the long term. The maximum difference between the real data and the predicted value is only 1 or 2 bikes, which supports the developed models are practically ready to use.