A prediction of bike flow in bike renting systems with the tensor model and deep learning

Abstract

Rental bikes are popular in many urban areas to help people expand their mobility. It is important to make the rental bicycle usable and available to the general public at the appropriate time and place. Inevitably, providing the city with a steady supply of rental bicycles becomes a major concern. The most important aspect is the estimation of the number of bicycles required in each bicycle sharing station at any given hour. This paper gives an examination of human mobility as indicated by bicycle renting information of the bike sharing system. In this paper, we proposed a new approach for forecasting the bike inflow and outflow from one station to another during certain time slots. Our method analyses human mobility pattern by two steps: (1) Using Tuckers tensor decomposition to create a 3D tensor to model human mobility and extract latent temporal and spatial characteristics of various stations and time slots. (2) to use a Long-Short Term Memory Neural Network to model the relationship between mobility patterns and the derived latent spatial and temporal features in order to predict bike flow between stations. The main contribution of this study that with the extracted latent characteristics through Tuckers factorization we improve the accuracy of prediction by 16% and decrease the amount of training data that used in prediction. Also, a root mean squared error of prediction is 1,5 bike. We compare our model with baseline models as historical average, ARMA, the feed-forward neural network, and KNN. The proposed method showed the best results.