The invention belongs to the field of deep learning, and particularly relates to a shared bicycle demand prediction system, method and device. The system comprise a data extraction and conversion unitused for extracting weather characteristics and time characteristics with maximum correlation from the real weather data, and converting the weather characteristics and the time characteristics intoa two-dimensional matrix by combining historical station state data; a neural network unit used for extracting, learning and training the feature information in the two-dimensional matrix. Then, optimization parameter updating operation is carried out; output of bicycle demand predictor. The system has the advantages of high efficiency, accurate prediction and high real-time performance.