ScholarWorks@중앙대학교

초록

This study predicts hourly tourist flows in a recurring festival context using high-resolution grid-level mobility data collected within a 300 m radius of an event site. A spatiotemporal graph neural network (GCN+GRU) is systematically compared with SARIMAX and other nonlinear baselines under an identical train-test framework to ensure methodological fairness. The STGNN demonstrates the highest out-of-sample performance (RMSE=99.77), substantially outperforming SARIMAX (RMSE=669.60), and the superiority is statistically confirmed through the Diebold-Mariano test. Comparative results also show that nonlinear tree-based models perform competitively, whereas conventional univariate deep learning models fail to capture peak dynamics adequately. Spatial diagnostics reveal significant clustering of node-level prediction errors (Moran’s I=0.4601, p=0.0010), indicating structured spatial dependence rather than random dispersion. These findings underscore the importance of explicitly modeling spatial interactions in recurrent event tourism forecasting and provide practical implications for crowd management and event-based urban operations.

키워드