A three-in-one dynamic shared bicycle demand forecasting model under non-classical conditions

Bicycle sharing systems are becoming highly popular and there have accumulated a large volume of users’ trajectory data. In a bicycle sharing system, most users’ borrowing and returning behaviors are arbitrary. Importantly, the bicycle sharing system is affected by weather, time as well as other dynamic factors under non-classical conditions. These factors make the scheduling of shared bicycles imbalanced, which is a typical machine learning problem under non-classical conditions. In addition, imbalanced bicycle scheduling impacts user experiences and causes huge economic losses. In this study, we propose a three-in-one dynamic shared bicycle demand forecasting model under non-classical conditions. The proposed model contains three essential and specific techniques: (1) Clustering stations based on the idea of data field. The activity degree of a station is calculated by constructing a bicycle transition network. The stations’ geographical locations and the bicycle transition patterns are taken into consideration. Then, nearby stations with similar transition patterns are aggregated into a cluster based on data field. (2) Modeling dynamic factors under non-classical conditions. The Pearson correlation coefficient is used to choose the most relevant dynamic weather features from real data. These features are transformed into a three-dimensional vector by taking into account the historical demand for bicycles in the station cluster. (3) Predicting bicycle demand using a LSTM (Long Short-Term Memory) network with multiple features that learns the features from the three-dimensional vector. The bicycle demand in each cluster is forecasted in a real-time and dynamic fashion every ten minutes. These three techniques collectively achieve one goal: accuracy prediction with low errors in real-time response. Extensive experiments are conducted on real Citi Bicycle data to compare the proposed model with traditional machine learning algorithms as well as the state-of-the-art approaches. The results demonstrate the superior performance of the proposed model in different evaluation measurements.