Fixed-point Temperature Prediction Based On Heterogeneous Fusion Model And Temporal Convolutional Network

The world-famous 2022 Winter Olympic Games will be held in Beijing and Zhangjiakou,and 70% of the snow projects are held in Chongli competition area of Zhangjiakou.Meteorological conditions are uncontrollable factors,it is one of the key factors that affect the success of the Winter Olympic Games.There are many peaks and ravines in Chongli Competition area.At the same time,the accumulation of meteorological data is less.Therefore,it is a great challenge to improve the accuracy of weather forecast under complex terrain.In addition to the small probability of blizzard,low visibility,strong wind,dust and other high impact weather events,it is also difficult to forecast the temperature of conventional elements.At the same time,the fixed-point forecast of temperature is also one of the meteorological problems that the Beijing Organising Committee for the 2022 Olympic Winter Games pays close attention to.Although the numerical prediction model has made great progress in temperature prediction,the prediction accuracy is still low in the mountainous complex terrain.The accuracy needs to be corrected by many methods.In the thesis,the research on temperature prediction of complex terrain in small and medium-scale 5 km precision area is carried out.The main work and innovations include the following parts.(1)The core observation station data set of Chongli competition area is constructed.According to the location information of meteorological observation stations,accurate retrieval and positioning is proposed to solve the demand of temperature prediction in the core competition area of Winter Olympic Games.The construction of point-to-point core competition area data set is realized.The bilinear interpolation method is used to extract the initial field data.Historical observation data are extracted in a unified time scale.To ensure the time sequence of the data,a three-dimensional tensor data set is constructed.(2)Data sets and feature elements are optimized.In order to find out the law of temperature change in the core competition area better,the best forecast period of the winter observation data in recent two years is analyzed.The best time series were selected for temperature prediction.The embedded method based on regression tree is used to complete the feature importance analysis.According to the prior knowledge of meteorology,some initial physical quantities are optimized and the initial characteristics are recalculated.(3)A heterogeneous fusion model based on Stacking strategy is proposed to improve the accuracy of fixed-point prediction.A heterogeneous fusion model based on Stacking strategy is constructed by using multiple integration strategies and single regression algorithm in machine learning.The experimental results show that,the accuracy of the heterogeneous fusion model is improved by about 15%,and the temperature prediction error was strictly controlled within 2?,compared with the traditional numerical prediction model.(4)A fixed-point temperature prediction model based on Temporal Convolutional Network(TCN)is constructed.In order to further optimize the predicted value and enhance the correlation between sequences,a fixed-point temperature prediction method based on Temporal Convolutional Network is proposed.The experimental results show that the error dispersion of TCN model is less than that of heterogeneous fusion model,which effectively improves the problem that the prediction accuracy of heterogeneous fusion model declines with time.