Spatiotemporal Demand Prediction for Bike Sharing Based on WT-ConvLSTM

To accurately predict the spatiotemporal demand for bike sharing, a hybrid model integrating wavelet transform (WT), long short-term memory (LSTM), and convolutional neural network (CNN) was developed, referred to as the WT-convolutional LSTM (ConvLSTM) model. In this model, Spearman's rank correlation coefficient was employed to identify the factors influencing demand within the target grid from a spatiotemporal perspective. After processing historical bike-sharing order data from Shanghai, the proposed model was applied to predict bike-sharing demand on both working and nonworking days in downtown Shanghai. The prediction accuracy was assessed using mean square error (MSE), root MSE (RMSE), and mean absolute error (MAE) under both 10-fold cross-validation (CV) method and the regular validation method. The findings indicate that the prediction accuracy of the WT-ConvLSTM model is influenced by the intensity of spatiotemporal demand, with better performance when demand is concentrated. Compared to predictions generated by LSTM and WT-LSTM models, the proposed WT-ConvLSTM demonstrated superior accuracy. The MSE values for the proposed model under both the 10-fold CV and regular validation methods were 0.002 and 0.003, respectively. Overall, the WT-ConvLSTM model enhances spatiotemporal prediction accuracy for bike-sharing demand, offering valuable insights for resource allocation and management strategies in bike-sharing systems.