Journal of Geo-information Science >
High-temporal-frequency Forecast of Tourist Flow for Tourist Attraction based on LBS and Deep Learning
Received date: 2022-04-27
Revised date: 2022-06-08
Online published: 2023-04-19
Supported by
National Natural Science Foundation of China(52078160)
In order to achieve accurate high-frequency forecasts of tourist flow for tourist attractions, this study proposes a forecasting method based on LBS and deep learning techniques. This method generates spatial-temporally controllable forecasts by converting the LBS data and using the core model — Deep Bidirectional Gated Recurrent Unit (DBi-GRU) model — built based on Bidirectional Recurrent Neural Network and GRU algorithms. To test the performance of our proposed method, we take the Shenzhen Dameisha Waterfront Park as an example, and three analysis methods including fitting curves, error criteria, and DM tests are used to test the forecasting performance of our DBi-GRU model. Additionally, five other deep learning models are set as reference models to compare with our model. The experimental results show that, first, DBi-GRU model proposed in this study has ideal forecasting performance in high-frequency forecast of tourist flow for tourist attractions and yields highly accurate forecasts in peak periods of tourist flow, and its performance is much better than the other deep learning models. Second, Bidirectional Recurrent Neural Network based models, particularly the Bidirectional LSTM based model, generally provide better performance than conventional Recurrent Neural Network based models. Though the forecast accuracy of the Bidirectional LSTM based model is not as high as DBi-GRU model, there is no significant difference between their model capability. Third, using the same network parameters, GRU algorithm has higher forecast accuracy than LSTM and RNN algorithms which are used by previous researchers. This study develops a new method for high-frequency tourist flow forecasting, and the high-frequency information forecasted in this study provides information support for management tasks of tourist attraction such as crowd control, service arrangement, etc..
XIE Qian , LU Ming , XIE Chunshan . High-temporal-frequency Forecast of Tourist Flow for Tourist Attraction based on LBS and Deep Learning[J]. Journal of Geo-information Science, 2023 , 25(2) : 298 -310 . DOI: 10.12082/dqxxkx.2023.220231
表1 深度学习模型的超参数与操作环境Tab. 1 The hyperparameter and operating environment of the deep learning model |
参数类型 | 超参数 | 超参数设置 | |
---|---|---|---|
网络参数 | 网络层数(Layers) | 5 | |
神经元数(Neurons) | 120,72,72,60 | ||
激活函数(Activation) | ReLU | ||
优化参数 | 优化器(Optimizer) | Adam | |
学习率(Learning rate) | 0.001 | ||
批量大小(Batch size) | 8 | ||
正规化与训练参数 | 随机失活(Dropout) | 0.1 | |
训练代数(Epochs) | 100 | ||
操作环境 | TensorFlow 2.3, Python 3.8, Anaconda 2.1 |
表2 本实验所使用的6种深度学习模型Tab. 2 Six deep learning models used by this experiment |
神经元算法类型 | |||
---|---|---|---|
GRU算法 | LSTM算法 | RNN算法 | |
双向循环神经网络 | DBi-GRU | DBi-LSTM | DBi-RNN |
常规循环神经网络 | D-GRU | D-LSTM | D-RNN |
表3 4种误差指标(R2、RMSE、MAE和MAPE)的评估结果Tab. 3 Evaluation results of the four error criteria (R2、RMSE、MAE and MAPE) |
测跨度 | 评估指标 | 深度学习模型 | ||||||
---|---|---|---|---|---|---|---|---|
双向循环网络模型 | 常规循环网络模型 | |||||||
DBi-GRU | DBi-LSTM | DBi-RNN | D-GRU | D-LSTM | D-RNN | |||
未来3日 客流量预测 | R2 | 0.948 | 0.928 | 0.860 | 0.927 | 0.869 | 0.799 | |
RMSE | 218.986 | 259.703 | 361.092 | 260.642 | 349.179 | 432.325 | ||
MAE | 131.914 | 148.097 | 199.625 | 155.404 | 211.489 | 259.284 | ||
MAPE | 0.340 | 0.362 | 0.652 | 0.368 | 0.469 | 0.549 | ||
高峰时段 客流量预测 | R2 | 0.918 | 0.898 | 0.679 | 0.824 | 0.723 | 0.466 | |
RMSE | 290.761 | 323.151 | 574.414 | 425.663 | 533.749 | 740.998 | ||
MAE | 208.730 | 233.964 | 391.663 | 282.890 | 396.007 | 593.208 | ||
MAPE | 0.112 | 0.125 | 0.241 | 0.158 | 0.250 | 0.452 |
表4 DM检验结果Tab. 4 Results of DM test |
评估的类型 | 评估的模型 | DM指标 | ||
---|---|---|---|---|
模型1 | 模型2 | DM值 | p值 | |
DBi-GRU模型的性能评估 | DBi-GRU | DBi-LSTM | -1.499 | 0.138 |
DBi-RNN | -1.682* | 0.097 | ||
D-GRU | -1.671* | 0.099 | ||
D-LSTM | -2.008** | 0.005 | ||
D-RNN | -1.719* | 0.090 | ||
双向网络模型与常规网络模型的 性能对比评估 | DBi-GRU | D-GRU | -1.671* | 0.099 |
DBi-LSTM | D-LSTM | -1.772* | 0.081 | |
DBi-RNN | D-RNN | -1.298 | 0.198 | |
常规网络模型的性能评估 | D-GRU | D-LSTM | -1.709* | 0.092 |
D-RNN | -1.738* | 0.086 |
注:“*”和“**”分别代表2个模型的性能差异在10%和5%的水平上显著。 |
[1] |
|
[2] |
|
[3] |
|
[4] |
|
[5] |
|
[6] |
|
[7] |
|
[8] |
余向洋, 胡善风, 朱国兴, 等. 基于LS-SVM方法的景区客流中期预测研究[J]. 旅游学刊, 2013, 28(4):75-82.
[
|
[9] |
|
[10] |
|
[11] |
|
[12] |
|
[13] |
段莉琼, 宫辉力, 刘少俊, 等. 基于客源地的聚类-ARIMA模型的短期旅游需求预测——以天津欢乐谷主题公园为例[J]. 地域研究与开发, 2017, 36(3):108-112.
[
|
[14] |
|
[15] |
|
[16] |
|
[17] |
黄先开, 张丽峰, 丁于思. 百度指数与旅游景区游客量的关系及预测研究——以北京故宫为例[J]. 旅游学刊, 2013, 28(11):93-100.
[
|
[18] |
梁昌勇, 马银超, 陈荣, 等. 基于SVR-ARMA组合模型的日旅游需求预测[J]. 管理工程学报, 2015, 29(1):122-127.
[
|
[19] |
|
[20] |
|
[21] |
|
[22] |
|
[23] |
|
[24] |
|
[25] |
|
[26] |
|
[27] |
|
[28] |
|
[29] |
|
[30] |
腾讯公司. 腾讯位置大数据[EB/OL].[2021-03-23]. https://heat.qq.com/bigdata/index.html.
[
|
[31] |
于丙辰, 陈刚. 基于腾讯区域热力图的庐山核心景区客流研究[J]. 国土与自然资源研究, 2017(2):83-89.
[
|
[32] |
|
[33] |
|
[34] |
|
[35] |
|
[36] |
Worldweatheronline. Shenzhen weather[EB/OL]. [ 2022- 01-20]. https://www.worldweatheronline.com
|
[37] |
|
[38] |
|
[39] |
|
[40] |
|
[41] |
|
[42] |
|
[43] |
|
/
〈 | 〉 |