A spatial–Temporal Large Language Model with Denoising Diffusion Implicit for predictions in centralized multimodal transport systems

Centralized multimodal transport systems face significant challenges due to data isolation, missing values, and heterogeneous spatial–temporal features, which hinder accurate prediction in traffic flow and travel demand. To address these challenges, we propose Spatial–Temporal Large Language Model with Denoising Diffusion Implicit (STLLM-DF), an innovative which integrates a Spatial–Temporal Denoising Diffusion Implicit Model (ST-DDIM) with a Spatial–Temporal Large Language Model (ST-LLM) to improve the predictions in traffic flow and travel demand in multimodal transport systems. The ST-DDIM effectively learns data distributions to recover noisy and incomplete data, while the ST-LLM captures complex spatial–temporal dependencies across multimodal networks, eliminating manual feature engineering. Extensive experiments conducted on ten real-world datasets from Sydney demonstrate that STLLM-DF consistently outperforms baseline models in both single-task and multi-task predictions (e.g., ), while consistently excelling in short-term and long-term predictions. On average, STLLM-DF achieves improvements in Mean Absolute Error (MAE) by 2.40%, Root Mean Square Error (RMSE) by 4.50%, and Mean Absolute Percentage Error (MAPE) by 1.51%. Furthermore, we evaluate the noise tolerance of STLLM-DF, demonstrating its robust performance under data imperfections. This paper presents a scalable, data-driven solution for managing multimodal transport systems, offering actionable insights for transport regulators.