A deep learning model based on multi-source data for daily tourist volume forecasting

ABSTRACT

Demand forecasting is important for management and decision making in the tourism sector. However, research on deep learning forecasting, which combines multiple data sources, is still in the development phase. This study proposes a bidirectional long short-term memory (BiLSTM) forecasting method incorporating an attention mechanism (ATT-BiLSTM) that can better extract data features from a set of predictor variables consisting of multiple predictor variables (generated from historical tourist volume, search engine data, weather data and day off data). The research experimentally validates the effectiveness of the method using the famous Chinese tourist attraction Jiuzhaigou as a case study. The results show that the proposed model not only has better generalization ability but also significantly outperforms the four benchmark models, convolutional neural network (CNN), SVR, LSTM, and BiLSTM, in terms of prediction accuracy. In addition, we analyse the importance of the different predictor variables in the prediction model characteristics.

KEYWORDS:

• Attention mechanism
• Bidirectional LSTM (BiLSTM)
• deep learning
• long-short-term-memory (LSTM)
• SHAP analysis
• tourism demand forecasting

Disclosure statement

No potential conflict of interest was reported by the author(s).

Data availability statement

The historical tourist volume data, weather data, search engine data and the Official China public holiday schedule, can be collected from the following websites, https://www.jiuzhai.com/news/number-of-tourists, http://www.weather.com.cn, https://index.baidu.com, and http://www.gov.cn/zhengce/xxgk/index.htm respectively.

Notes

1 After the magnitude 7.0 earthquake, the Jiuzhaigou Attraction Management Committee issued an announcement on 8 August 2017 suspending the reception of tourists. For details, refer to the website https://www.jiuzhai.com/news/notice/5945-2017-08-08-15- 45–42.

2 The search volume curve figures for several major keywords can be found in the Online supplementary document.

Additional information

Funding

This work is partially supported by Beijing Natural Science Foundation [Grant number No. 9222005].