Daily Gas Load Forecasting Based On Deep Learning

The accurate forecasting result of daily gas load is related to the reliability of gas supply and directly affects the economic benefit of gas enterprises.Since the historic data of daily gas load is a random non-stationary time series,the forecasting with error-free is virtually impossible at present.Thereby,the increasing scholars are dedicated to the improvement of forecasting accuracy.Due to its ability to effectively deal with various complex prediction problems,deep learning has gradually become one of the research focus in the field of artificial intelligence.Applying deep learning to gas daily load forecasting is of great significance to improve forecasting accuracy and generalization performance.Therefore,this paper focuses on the daily gas load characteristics and forecasting models of three cities with different gas-utilizing structures: city A with mainly civil users city B with mainly business customer and city C with mainly industrial user.The main research work and conclusions are as follows:(1)Pre-processing the historical data of daily gas load.The experimental results show that the modified multiple regression method is better than the linear interpolation method and the multiple regression method in filling missing data for consecutive days.It is obvious that the isolation Forest is more efficient and accurate than the Local Outlier Factor algorithm for outlier data detection.(2)Comparing and analyzing the variation law of daily gas load in the year,month,and holidays in different cities,and using relevant analysis methods to reseach the effect of various factors on the daily load.The analysis results show that the periodicity of cities A and B is more obvious than that of city C.The daily load of city A in cooling period of the mild season and city B during the period of holiday have stronger fluctuation.The gas-utilizing structure affects the abrupt direction and magnitude of the holiday load for seven consecutive days,but has a weaker effect on the holiday for three consecutive days.(3)Considering the time sequence characteristics of daily gas load comprehensively,a forecasting model of long short-term memory network was established for three cities from the aspects of input variables,time information mining,gradient problem,optimization algorithm and different seasonal characteristics.The forecasting results show that the forecasting performance of the long short-term memory network forecasting model is better than the auto regressive integrated moving aaverage model,the back propagation neural network model,and the recurrent neural network model.The mean absolute percentage error range of each season is between 1.82% and 4.74%.(4)Considering that the long term memory network has insufficient ability to extract local feature information,a hybrid model combining convolutional neural network is proposed.The hybrid model can effectively increase the forecasting accuracy and generalization performance.In summary,the deep learning algorithm can effectively increase the accuracy of daily gas load forecast.The research train of thought in this paper can provide reference for the establishment of other city forecasting models,which has certain practical significance.