Improved daily tourism demand prediction model based on deep learning approach

The rapid escalation in tourist visitation poses significant challenges, including traffic congestion, overcapacity at scenic attractions, heightened risks of safety incidents, and diminished visitor satisfaction. To optimize scenic area management through better resource allocation and service delivery, as well as to facilitate informed travel and visitor planning, this study proposes a hybrid predictive model, called Secretary Bird Optimization Algorithm-Convolutional Neural Network-Bidirectional Gated Recurrent Unit-Attention Model (SBOA-CBAM), for forecasting tourist volume within scenic areas. The methodology involves constructing a foundational Convolutional Neural Network-Bidirectional Gated Recurrent Unit-Attention Mechanism (CBA) model, where input data pre-processed via the Maximal Information Coefficient (MIC) algorithm undergoes feature extraction using Convolutional Neural Network (CNN), followed by bidirectional temporal feature mining via BiGRU, and output weighting via an Attention Mechanism to emphasize critical features and generate predictions. Subsequently, the Secretary Bird Optimization Algorithm (SBOA) is employed to autonomously identify the optimal hyperparameter configuration for the CBA model, thereby enhancing its predictive accuracy and computational efficiency. Comparative simulation experiments demonstrate the high applicability of the CBA model for scenic area tourist flow forecasting and reveal that the SBOA-optimized CBAM model achieves statistically significant performance enhancements, namely, a 3.8966% increase in R 2 , alongside reductions of 19.9025% in RMSE, 12.1726% in MAE, 8.3196% in MAPE, and 43.7662% in MSE. Statistical validation via the Wilcoxon signed-rank test confirmed the significance of the improvements (RMSE: p = 0.0001; MAE: p = 0.0007), with substantial effect sizes indicated by Cohen’s d values of 0.8982 (RMSE) and 0.7028 (MAE). These findings corroborate that the SBOA algorithm not only substantially elevates predictive precision but also enhances model stability and robustness against disturbances.