|
Abstract
|
This thesis proposes a deep learning framework—ATT-BiLSTM—for forecasting daily tourist volumes using multi-source data such as historical tourist numbers, search engine trends, weather conditions, and holiday schedules. By integrating attention mechanisms with bidirectional LSTM networks, the model prioritizes relevant inputs to improve predictive accuracy. The research demonstrates that this model outperforms traditional and other deep learning models (like CNN and LSTM) across key evaluation metrics. SHAP analysis is employed to enhance interpretability and identify key influencing factors. The framework supports both theoretical advancement and practical applications in tourism planning and sustainable management.
|