Beyond Trade-Offs: Advancing Spatial-Temporal Forecasting In Transportation With Deep Learning

Abstract

Accurate and efficient traffic flow prediction is essential for modern IntelligentTransportation Systems (ITSs). As urban road networks and sensor infrastructuresexpand, forecasting methods must capture complex spatial–temporal dependencies whileremaining computationally tractable for real-time use. Traditional statistical methodsstruggle with non-linear patterns and large-scale data, whereas deep learning approaches—such as CNNs, RNNs, and GNNs—face challenges in modelling long-rangedependencies or incur high computational costs. This thesis advances spatial–temporal deep learning for traffic flow and passengerdemand prediction, addressing the trade-off between accuracy and efficiency. First,CCDSReFormer introduces a criss-cross dual-stream Transformer with enhanced rectifiedself-attention and geographic–semantic masking, delivering improved accuracy with lower complexity. Second, ST-MambaSync fuses selective state-space (Mamba) andTransformer mechanisms, providing theoretical and empirical evidence of theircomplementarity, achieving notable reductions in prediction error, FLOPs, and inferencetime. Third, STDAtt-Mamba employs dynamic attention and state-space modelling tocapture heterogeneous passenger behaviours in multimodal public transit, outperformingexisting methods in multi-type demand forecasting. Together, these innovations offer theoretical insights, empirical validation, and practicalsolutions for scalable, efficient, and equitable transportation management. Experiments onreal-world datasets confirm consistent gains in both accuracy and speed, culminating in acohesive framework that enhances the scalability, robustness, and responsiveness ofspatial–temporal traffic forecasting in ITSs.