High-Frequency Tourist Flow Forecasting with Gated Recurrent Units and Attention Mechanisms

Accurate prediction of tourist flow is essential for effective crowd management and service optimization in popular tourist destinations, especially during peak hours and high-frequency periods. This study proposes a novel deep learning model—BiGRU-Attn, which integrates Gated Recurrent Units (GRU), Attention Mechanisms, and Bidirectional Recurrent Neural Networks (BiRNN) to capture complex temporal dependencies in sequential data. The model is evaluated using real-world, high-frequency passenger flow data from Fuzhou National Forest Park, a major sport tourism site in China. Comparative experiments against five baseline models—including both bidirectional and conventional recurrent architectures—demonstrate the superior predictive performance of BiGRU-Attn. It achieved the highest R 2 (.948) and the lowest RMSE (218.986), MAE (131.914), and MAPE (0.340) in 72-hr forecasts. During peak-hour periods (12:00–18:00), BiGRU-Attn maintained an R 2 above .9, consistently outperforming all other models. Notably, bidirectional models (BiGRU-Attn and DBi-LSTM) showed clear advantages over their conventional counterparts, particularly in capturing short-term fluctuations. These findings underscore the potential of BiGRU-Attn to support smart tourism applications, enabling more accurate real-time visitor flow forecasting, proactive resource planning, and enhanced visitor experience management in dynamic environments.