Abstract: The demand prediction of shared bicycles is important for the refined operation of shared bicycles. This serves as the premise of solving the balance of supply and demand in the bicycle area. Aiming at accurately predicting the demand for bike-sharing, this study firstly analyzes the spatiotemporal characteristics of the bike-sharing data of Citi Bike in New York from 2017 to 2019 and explores the distribution patterns of bike-sharing trips. Subsequently, the study integrates the feature extraction ability of convolutional neural networks and the bidirectional temporal data processing ability of bi-directional long-short term memory to construct a convolutional bidirectional long-short-term memory network CNN-BiLSTM model. The input features of the CNN-BiLSTM model are determined using correlation analysis from travel data, meteorological data, and space-time travel characteristics of bicycles, and then a well-built bicycle prediction model is generated. Finally, CNN, LSTM, BiLSTM, and CNN-LSTM are selected as benchmarks to evaluate the performance of the CNN-BiLSTM prediction. Results show that the CNN-BiLSTM model has the smallest evaluation indicators of MAE and RMSE with 0.035 and 0.058, and the largest R² with 0.922, and achieves the best prediction performance. This research provides a reference for the actual scheduling and redistribution of shared bicycles.

Key words: bike-sharing, demand predict, CNN, BiLSTM, CNN-BiLSTM