Research On The Method Of Hotel Online Sales Forecast Based On The Blending Of LightGBM And WaveNet

In recent years,China's tourism market has maintained a continuous development trend.The Internet has penetrated into all aspects of national life.People tend to book hotels through the Internet.OTAs with a large number of hotels have accumulated massive hotel and consumption data.Hotel sales is the basis and core of revenue management.If the existing technologies and methods predict the sales changes of specific hotels in a certain period of time in the future,on the one hand,it will help guide the hotel inventory management and improve the benefit of revenue management,on the other hand,it will also improve the user experience and bring benefits to the whole platform.Due to the influence of many factors,such as the properties of hotels,price,holiday and so on,the traditional time series forecasting methods can not fully mine the data characteristics and deal with the high-dimensional and non-linear big data efficiently,additionally,the prediction error is large.Aiming at this problem,this paper puts forward an online sales forecasting method based on the linear blending of lightgbm and WaveNet model.The main work of the study is as follows:Analyze and process historical data,including missing value and dirty data,analyze sales data of different dimensions,describe the characteristics of hotel online sales data,and select a reasonable prediction time range.Organizational feature engineering,one is the basic feature,including the basic data in the dimensions of region,business district,hotel attribute,holiday time,etc.Encode the unique characteristics of Chinese holiday time,which is coded according to the unique characteristics of holiday time in China,and one-hot encoding is used in the non digital feature;The other is the mathematical statistical feature based on time sequence,which is combined in different time windows and combinations according to the target prediction time range.Different statistical data are extracted from cross feature combination.On the one hand,this paper constructs a hotel online sales forecasting model based on LightGBM.LightGBM has the characteristics of small memory demand,fast computing speed and low communication cost.On the other hand,it constructs a hotel online sales forecasting model based on WaveNet,which is an autoregressive deep generation model.It uses multi-layer causality convolution to express the temporal characteristics by expanding the output receptivity field,and the residual network structure and parameterized skip connection are used to improve the training depth.It achieves better results for audio analysis than all previous models in deep learning.Combined with the sequence to sequence prediction framework,it innovatively applies the improved WaveNet to the hotel online sales prediction.Because the two models have irreplaceable advantages,finally,the two models are linear blended to improve the accuracy of model prediction.A series of experiments using the real sales data of Ota platform prove the efficiency of the algorithm model.