Research On Combination Model Based On Feature Selection For Natural Gas Load Forecasting

Accurate natural gas load forecasting can provide gas purchasing standards for natural gas companies and stable gas supply for customers.In order to further solve the problem of natural gas supply and demand,so as to ensure the smooth use of natural gas,accurate natural gas load forecast is particularly important.The data set selected in this paper is A total of 1097 sets of daily natural gas load data for users A and B along the gas transmission pipeline of a natural gas company from October 6,2018 to October 6,2021.The ratio of training set to test set is 7:3.Firstly,in order to improve the quality of the data set,the mean value correction and wavelet denoising methods are used to deal with the outliers and noise in the load data.In the process of research,in order to avoid the larger order of magnitude data covering the smaller order of magnitude data,the data set is normalized to obtain reliable data.Redundancy of influencing factors will reduce the efficiency of load prediction.In order to eliminate redundancy factors,graphic method and feature selection Relief method are used to select and weight analysis six influencing factors including temperature,wind power,weather and date type,and it is concluded that the three influencing factors with maximum weight are minimum temperature,maximum temperature and average temperature.Secondly,in the four single models of ARIMA,BPNN,SVR and XGBoost,SVR and XGBoost are selected as the benchmark models according to the advantages and disadvantages and the analysis of the prediction results.In view of the difficult and time-consuming selection of parameters of a single model,PSO and SOA optimization algorithms were used to optimize key parameters of SVR and XGBoost respectively,and the longitudinal combination models PSO-SVR and SOA-XGBoost were obtained.In order to reduce the prediction error of the longitudinal combination model,A new horizontal and vertical combination model PSO-SVRSOA-XGBoost was obtained by assigning appropriate weights to the two vertical combination models using variance-covariance method.Finally,the experimental platform of load forecasting program was designed and written,and the daily natural gas load data set was substituted into seven load forecasting models for training and forecasting.The experimental results show that compared with the single model ARIMA,BPNN,SVR and XGBoost,the average absolute percentage error predicted by the horizontal and vertical combination model PSO-SVR-SOA-XGBoost increases by 15.492% and 13.315%,10.141%and 10.575%,7.514% and 8,respectively 101%,5.686% and 5.578%;Compared with the longitudinal combination model,PSO-SVR and SOA-XGBoost were increased by 2.531%,2.082%,1.08% and 0.544%,respectively.The PSO-SVR-SOA-XGBoost model proposed in this paper has a higher forecasting accuracy than other models in daily load forecasting,considering the influence of temperature.