Ablation Study of Spatiotemporal Features and Bidirectional LSTM in Bicycle Demand Forecasting

The penetration rate of shared bikes in urban transportation continues to rise, yet scheduling challenges caused by temporal and spatial supply-demand mismatch hinder industry development. Accurate demand forecasting is key to optimizing operational strategies and improving service efficiency, with significant research value. This paper focuses on bike-sharing demand forecasting, proposing an enhanced spatio-temporal attention model to boost prediction accuracy. It uses Mobike riding order data from Shanghai, extracting temporal features like hours and weekdays, and constructs a matrix with spatial features (average neighborhood demand after latitude-longitude network division). Synthetic Minority Over-Sampling Technique (SMOTE) is adopted to oversample peak data for data imbalance. Processed data is input into a bidirectional Long Short-Term Memory (LSTM) to capture temporal correlations in the sequence, combined with a spatio-temporal attention mechanism to focus on key features. Model performance is evaluated via ablation experiments (removing spatial features, bidirectional LSTM, or attention mechanism) using indicators such as RMSE, MAE, MAPE for non-zero demand, zero-value prediction accuracy, and peak hit rate. Results show the model effectively captures spatio-temporal features, providing strong support for bike-sharing scheduling management.