Multi-granular metro passenger flow forecasting via MMoE and hierarchical multi-source data fusion

Accurate forecasting of metro passenger flows at multiple granularities, such as station-level inflow/outflow (IO) and inter-station Origin-Destination (OD) and Destination-Origin (DO) flows, is vital for efficient metro system management. However, the heterogeneous temporal patterns, asynchronous dependencies, and complex inter-task relationships make unified modeling highly challenging. To address these issues, we propose MG-MMoE, a unified multi-granularity forecasting framework that combines a multi-gate Mixture-of-Experts (MMoE) architecture with hierarchical multi-source data fusion. Specifically, MG-MMoE incorporates MMoE modules into each encoder and decoder layer to facilitate dynamic knowledge sharing across IO, OD, and DO tasks, allowing each to adaptively leverage shared and task-specific information. Meanwhile, a hierarchical fusion strategy integrates spatial (POI similarity), temporal (time-of-day, holiday), and static (station attributes) features into separate layers, enhancing the semantic richness of the learned representations. Extensive experiments on two real-world metro datasets show that MG-MMoE achieves superior performance over the state-of-the-art baselines across multiple prediction horizons and tasks, demonstrating the effectiveness of cross-task interaction and structured data fusion in multi-granular passenger flow forecasting.