Mixture Survival Classification Neural Network for Power Failure Risk Prediction in Two-Wheel Shared Mobility

Shared mobility is becoming increasingly popular in the construction of smart cities. However, there is no relevant research on battery swapping task of shared electric bikes (e-bikes) at present. In this article, the battery swapping problem is transformed into a power failure probability prediction problem. Unlike traditional machine learning models, the goal is to simultaneously estimate when and whether a power failure will occur, in other words, to estimate the distribution function of power failure time for each e-bike. To solve this problem, an innovative mixture survival classification neural network (MSCN) model is proposed, which expands the traditional survival model and integrates the characteristics of Neural Network. Furthermore, for a more accurate prediction, we mine the temporal and spatial correlation of e-bikes flow between sites and model this correlation using graph networks and Sequeeze-and-Excitation network (SENet). Experiments are conducted on a real shared e-bikes riding data set from hellobike, one of the largest sharing mobility service provider in China. The results show that MSCN has excellent predictive performance and can provide data support for the formulation of battery swapping strategy.