Convolutional Neural Network Deep-Learning Models for Prediction of Shared Bicycle Demand

Abstract

Digital technologies over the past decade have enabled the growth of the Sharing Economy, through fast, cheap, connected mobile devices, coordinating massive databases and machine learning/artificial intelligence services. We applied TensorFlow, the leading industry AI tool, to analyze 55 million contracts from Mobike, the world’s largest bike-sharing company. We elicited demand behavioral cycles, finding demand peaks at periodicities of 7, 12, 24 h and 7-days. Bicycle demand showed wide variance in frequency of use, thus time-series models would strongly overfit the data yielding unreliable models. We applied deep-learning TensorFlow analysis to the time-series axis of our data using a 1D convolutional network, and the 2D location and 1D environmental variables with fully connected network layers. Rebalancing to place shared bikes where demand was predicted to be greatest was predicted to reduce rebalancing costs by between 21.4% and 28.9%.