A Dual Attention-Based Recurrent Neural Network for Short-Term Bike Sharing Usage Demand Prediction

Improved network technology and data openness have generated a large number of trajectory-based data, and trajectory-based knowledge graphs help to provide consecutive and effective development and optimization for intelligent transportation systems (ITS). Bike-sharing systems (BSS) are increasingly important in the traffic planning, management, and deployment of ITS. Therefore, it is crucial to determine how trajectory data are best used to accurately predict the demand for short-term bike sharing usage. This study proposed a recurrent neural network (RNN) model based on the dual attention mechanism to extract spatial and temporal features. The attention mechanism is able to determine and weight all location features of the data in time series to learn mutual correlations. The method applied in this paper can effectively conduct an adaptive combination of local- and global-feature dependencies for trajectory data, in order to effectively predict the trend of short-term bike sharing usage demand. In addition, this study adopted the random walk mechanism to maintain local relations between bike stations in the preprocessing of time series data, which makes it more adaptive to the local location changes of different stations. Finally, the experimental results show that the model architecture in this study combined the attention and random walk mechanisms to achieve better prediction performance.