Abstract: Tourism is a pillar industry in many countries, however, due to its “spatial mobility of people”, tourism is one of the industries most affected by the COVID-19 epidemic. Accurately forecasting tourism demand under the impact of the COVID-19 epidemic is very important for the strategic planning of tourism destinations and tourism-related companies. However, the uncertainty brought about by the impact of the COVID-19 epidemic has led to major challenges in forecasting tourism demand. Therefore, it is extremely important to grasp the law of changes in tourism demand as the epidemic changes, and provide guidance and advice for the tourism industry to judge the changes in tourism demand under the influence of the epidemic.
   However, most of the existing tourism demand forecasting models are limited to the selection of input variable types or the application of preprocessing methods such as synthesis or decomposition of input variables, ignoring the analysis and interpretation of the coupling relationship between tourism demand and different influencing factors. Although the existing deep learning algorithms can improve the accuracy of tourism demand forecasting, they still belong to the category of “black box” models. The lack of explanatory power creates some barriers for tourism managers to accept the research information. Therefore, it is urgent to adopt new technologies to build a new model for interpretable tourism demand forecasting under big data, and to achieve rapid decision-making and ensure forecasting accuracy in complex dynamic environments such as epidemic factors or Internet big data environments. Different from previous studies, this study introduces an explainable deep learning model for tourism demand forecasting, which can provide a comprehensive explanation for tourism demand forecasting based on multi-source heterogeneous data.
   This study proposes a novel interpretable tourism demand forecasting framework that considers the impact of the COVID-19 epidemic, by using multi-source heterogeneous data, namely historical tourism volume, new local confirmed cases, Baidu index, and weather data to predict changes in domestic tourism demand under the influence of the epidemic. This study introduces the concept of epidemic-related search engine data for tourism demand forecasting and proposes a new composition leading search index–variational mode decomposition method to process search engine data. To improve the interpretability of existing tourism demand forecasting methods, a new JADE-TFT interpretable tourism demand forecasting model is proposed, which utilizes an adaptive differential evolution algorithm with external archiving (JADE) to optimize the hyperparameters of Temporal Fusion Transformers (TFT) efficiently. TFT is an attention-based deep learning model that combines high-performance forecasting with interpretable temporal dynamics analysis, showing excellent performance in forecasting research. The TFT model produces interpretable tourism demand forecast outputs, including the importance ranking of different input variables and attention analysis at different time steps. The proposed prediction framework is verified by a real case based on Huangshan tourism data. The interpretable experimental results show that epidemic-related search engine data can better reflect tourists’ travel concerns in the post-epidemic era.
   This study aims to construct an interpretable tourism demand forecasting framework considering the impact of the epidemic. In addition, this study introduces the concept of an epidemic-related search index, thereby providing new insights into tourism forecasting under the impact of the epidemic. This study has important practical implications for managers of tourist destinations and attractions. First of all, the repeated occurrence of the epidemic has led to large fluctuations in tourism demand. It is impossible to judge the fluctuation of tourist arrivals under the impact of the epidemic by relying only on the traditional low-season and peak-season information commonly used in the past. The proposed method emphasizes the importance of the epidemic-related search index, thereby improving the accuracy of tourism demand forecasting under the impact of the epidemic. In the long run, tourism authorities can apply tourism demand forecasts to support crowd management and better preparedness against COVID-19. Secondly, tourism operators can judge the impact of the epidemic on tourism demand through search indicators related to the epidemic, rather than the number of newly confirmed local cases, because the former can better reflect tourists’ concerns. Finally, the high-frequency subsequences obtained by the CLSI-VMD method are helpful for identifying the peaks and valleys of the tourism market and provide better support for the decision-making of tourism managers.
   This study also has some limitations. First of all, tourism demand forecasting considering the impact of the epidemic is very complicated, and more influencing factors can be further considered such as the impact of government policies and restrictions on the number of tourists. Second, considering the diversity and complexity of input variables by using multi-source data, other powerful deep learning models can be used. Finally, in addition to CLSI and VMD methods, other efficient composition or decomposition methods can also be used to process search engine data effectively. In the future, we will further study the problem of interpretable tourism demand forecasting under the impact of the epidemic.

Key words: tourism demand forecasting, interpretable forecasting, compound index, deep learning, COVID-19

摘要： 本研究提出了一种新颖的考虑疫情冲击的可解释性旅游需求预测框架,通过使用多源异构数据,即历史旅游量、本土新增确诊病例、百度指数、天气数据来预测新冠疫情影响下国内旅游需求变化。本研究引入疫情相关的搜索引擎数据的概念来进行旅游需求预测,并提出合成领先搜索指数和变分模态分解法结合处理搜索引擎数据的方法。同时,本研究提出了JADE-TFT可解释旅游需求预测模型,该模型利用带有外部归档的自适应差分进化算法智能高效地优化时域融合变换器的超参数。TFT是一种基于注意力的深度学习模型,将高性能预测与可解释的时间动态分析相结合,在预测研究中表现出优异的性能。TFT模型产生可解释的旅游需求预测输出,包括不同输入变量的重要性排序和不同时间步的注意力分析。基于黄山旅游数据的实际案例验证了所提出预测框架的有效性。可解释的实验结果表明,疫情相关的搜索引擎数据可以更好地反映后疫情时代游客对出行的担忧。

关键词: 旅游需求预测, 可解释性预测, 复合指数, 深度学习, 新冠疫情