Bike sharing has brought convenience to solve the ‘last mile’ problem. In order to provide better service, the shared bikes should be dynamically scheduled according to the demand. Currently, researches on demand prediction can be divided into two categories. One is based on traditional statistical models, the other applies deep learning models. The former aims at short-term prediction of 5 to 10 minutes, while the scheduling cannot respond in such a short time. The latter always ignores influential factors, leading to a decline in accuracy and reliability. Therefore, this paper aims to develop a demand forecasting model that can significantly improve accuracy for long-term prediction with higher reliability. Taking the region within inner ring of Pudong, Shanghai as study area, an improved LSTM NN model is proposed, based on POI data which is represented land-use information. The improved model takes land-use information as a factor that influences the demand of shared bikes. The result suggests that improved LSTM NN has higher accuracy than statistical models, and the original LSTM NN as well, especially for long-term prediction.

CICTP