Abstract: In newly developed or data accumulating scenic spots, passenger flow prediction often faces challenges such as insufficient samples, high data noise, and unstable structure. Traditional deep learning models often have poor performance in such situations due to high requirements for data volume, complex structure, and easy overfitting. In response to this issue, this article proposes a combined prediction method that combines statistical models and machine learning algorithms to improve the accuracy of passenger flow prediction in the Ink Danxia Scenic Area. This method combines the linear modeling advantage of ARIMA model with the nonlinear fitting ability of XGBoost model, and introduces GARCH model to model the volatility of ARIMA residuals. The weight allocation of the two models is dynamically adjusted according to the volatility, thereby enhancing the adaptive ability of the model in unstable passenger flow sequences. The research results indicate that the combined model can achieve more stable and accurate prediction results when facing complex and volatile tourist data, providing more practical prediction tools and scientific decision support for local governments and scenic area managers.