Predicting short-Term bike-Sharing demand at station level: A multi-Task dynamic graph-based spatiotemporal approach

Bike-sharing systems face challenges with supply-demand imbalances, causing user dissatisfaction and inefficiency. Accurate prediction of demand is crucial for optimizing these services. While deep learning methods have explored spatiotemporal dynamics in bike-sharing demand, most rely on predefined spatial correlations and focus solely on bike check-out demand, neglecting the relationship with check-in demand. To address these gaps, this paper introduces the Multi-Task Dynamic Graph-based Neural Network (MTDG) for predicting hourly bike-sharing demand across city stations. Initially, we analyze historical data and identify three key historical features for each time interval: closeness, period, and trend. Subsequently, we design three separate streams, each targeting one historical feature, with components to capture spatial and temporal dependencies in each. Spatial information is extracted using a graph convolution operator combined with a time-varying semantic adjacency graph based on historical demand similarities. We employ dual-input Long Short-Term Memory (di-LSTM) recurrent block to learn temporal dependencies and facilitate multi-task learning. This component enables the extraction of hidden pairwise demand correlations by treating the prediction tasks of bike check-in and check-out demands as interrelated. We also incorporate global features, like meteorological data, to capture broader-scale changes. The short-term bike-sharing check-in and check-out demands are jointly predicted by integrating spatiotemporal representations from three streams with global features. Using data from Montreal and New York City’s bike-sharing services, our model outperforms state-of-the-art methods, including fully adaptive graph models and Large Language Models (LLM)-based forecasting models. Variants of MTDG, such as single-task methods and alternative semantic adjacency graph configurations, also show superior performance over most baseline models.