A Bike-sharing Demand Predicting Model with Integrating Temporal Convolutional Network and Self-Attention

As a green and convenient means of transportation, bike-sharing has received significant attention and made rapid development worldwide. However, due to the differences in urban areas and users' riding habits, the supply and demand of bike-sharing is unbalanced, which affects the full utilization of bike-sharing. Given this, it is necessary to predict the demand for bike-sharing. In this paper, we propose a Bike-sharing Demand Predicting (BDP) Model with integrating Temporal Convolutional Network (TCN) and Self-Attention. BDP model takes multiple inputs to extract the feature information from multi-source data and employs the parallelism of TCN and Self-Attention to improve the training speed. Moreover, we bring the time signals for reflecting the periodicity of bike-sharing demand in the unit of day, week, and year. To evaluate the research work, we select five deep learning models, and two alternated BDP models with different inputs as the baselines. The results show that BDP model outperforms the baseline models with the highest prediction accuracy and the better indicators of MAE, MAPE, SMAPE, and RMSE.