<p>Accurate and interpretable tourism demand forecasting helps to improve the operation and management of tourism-related businesses and government departments. Although existing studies have proposed a variety of tourism demand forecasting models based on deep learning methods, which significantly improve the accuracy of tourism demand forecasting, there is a lack of a comprehensive interpretable forecasting method that provides a comprehensive, transparent, and reliable explanation of the forecasting process. In this study, a comprehensive method for multidimensional explainable tourism demand portfolio forecasting based on multi-source data is proposed. First, tree shapley additive explanations methods are used to select the feature data most relevant to tourism demand and provide an interpretable analysis. Then, attention - sparse autoencoder method is used to reduce the dimensionality of the selected important features. Meanwhile, an efficient and robust Informer - bidirectional long short-term memory neural network model is proposed in this study and optimized using polar lights optimizer. Finally, a multi-objective optimization algorithm enhanced by reinforcement learning is used to optimize the weights of the ensemble prediction model. In addition, the local interpretable model-agnostic explanations interpretability method is used in this study with the results of Tree shap to form a multidimensional interpretability framework. The experimental results show that the proposed method achieves higher accuracy in forecasting tourism demand, while providing managers with more reliable and comprehensive process and decision support.</p>