Forecasting tourism demand with pre-holiday attribute

Tourism demand forecasting is critical for decision-making in emergency resource allocation, and personnel management at tourist destinations. However, accurate predictions require the integration of multiple factors, such as network data of scenic areas, weather information, date-related data, and historical trends. To build a precise, robust, and generalizable forecasting model, this study introduces a novel “pre-holiday” feature for date information and employs the Transformer deep learning architecture as the core framework. The model integrates convolutional padding and serialization techniques to extract interval-based data features. Ultimately, it is designed to be applicable across diverse tourist destinations and varying forecasting time horizons, delivering end-to-end intelligent predictions—from data input to final output. Extensive ablation and comparative experiments show that the model is not only adaptable to single time-span forecasting but also achieves high-accuracy predictions for 15-day medium range forecasts. Compared with the existing mainstream research, the prediction error of the model is only 67% of the existing optimal model. Furthermore, the model demonstrates strong transferability across datasets from different scenic areas.