Scenic spot crowd flow prediction: Combining large language model semantic understanding and spatio-temporal network

Accurate forecasting of tourist flows is crucial for urban planning, transportation management, and overcrowding prevention. Tourism demand is highly sensitive to external stimuli, often leading to abrupt and irregular fluctuations that are difficult to quantify using existing methods. To address this issue, we propose an event-aware dynamic spatiotemporal network (EAD-STN) that integrates large language model embeddings with dynamic spatiotemporal graph modeling. External factors are encoded into high-dimensional representations to capture their impacts on visitor demand. The model further incorporates both dynamic spatial adjacency and LLM-derived semantic adjacency among attractions, enabling a more comprehensive characterization of inter-attraction relationships. Experiments on real-world datasets from Beijing and Shanghai demonstrate that EAD-STN consistently outperforms state-of-the-art baselines. By leveraging large language models to quantify complex external influences and semantic relationships, this framework provides a new perspective for improving the accuracy and interpretability of tourism flow forecasting.