Predicting Station-Level Bike Demand in Bike-Sharing Systems Using a Hybrid Neural Network

Supply-demand imbalances are a prevalent issue in bike-sharing systems with docking stations, hindering effective operation and use of the system. To improve efficiency for both operator and user, accurate short-term predictions are required. This study focuses on Montreal’s bike-sharing system BIXI. Demand for bikes is predicted 15 minutes ahead, using 24 hours of previous data. Predictions are made on a station level and a community level. The community level consists of clustered and aggregated station data, based on the community detection technique Louvain. Three models are implemented: an LSTM baseline model and two interpretations of the hybrid CNN-LSTM architecture. Bike demand is represented in three different ways: number of pick-ups, number of returns, and number of pick-ups subtracted from the number of returns (i.e. change of supply at the station). The results include an advantage of community-level predictions over station-level predictions. The latter suffers from outliers and sparsity of the data, while the former is more robust. The three models perform highly similarly, with no improvement of the predictions from using a hybrid neural network compared to a conventional LSTM. Finally,no difference is found between the prediction of bike pick-ups and bike returns. However, representing the data as the change of the supply at a station generates more accurate predictions than representing pick-ups and returns separately.