Sparse trip demand prediction for shared E-scooter using spatio-temporal graph neural networks

Abstract

The shared electric scooter (E-scooter) is an emerging micro-mobility mode in sustainable cities. Accurate hourly trip demand prediction is critical for effective service maintenance, but it poses a challenge due to the dynamic distribution influenced by urban complexity. We propose a model, the Sparse Diffusion Convolutional Gated Recurrent Unit (SpDCGRU), which incorporates diffusion convolution layers into the gated recurrent unit (GRU) model, enabling the simultaneous capture of spatio-temporal dependencies. Tackling the data in Louisville, Kentucky, USA, we demonstrate that spatial data reclustering and fusion loss training strategies contribute to the prediction performance. Moreover, the periodic and weather features positively impact predicting the low and high trip demand levels, respectively. Our model outperforms others in terms of overall performance and each trip demand level, with a 4.75% improvement in the mean absolute error (MAE) compared to the graph convolutional recurrent network (GCRN).