Abstract: In view of the lack of research on fault prediction method, low prediction accuracy and instability of dockless sharing bikes system, and on the basis of mining the information features between the predictive factors and the faulty bikes, an improved entropy weight method is proposed to divide the weights of the predictive factors. And three single prediction models, BP neural network, radial basis function and Elman neural network, are used to establish a combined prediction model based on critical weight, to solve an example on Matlab. The results show that: compared with the single prediction method, the combined prediction method improves the prediction accuracy by about 5%, and can reduce the prediction risk and system error and has good practical value.

Key words: sharing bikes, combination prediction, critical weighted, improved entropy weight method