Research On Energy Consumption Prediction For Short And Medium Term Based On Improved ARIMA And XGBoost

With the increasing of energy consumption,the concentration of greenhouse gas increases gradually,which leads to global warming.Energy forecasting is very important for energy planning,management and saving.Energy consumption prediction can help users estimate potential energy savings,implement effective energy management,and reduce energy consumption.Using artificial intelligent methods make energy consumption predictions more accurate and energy consumption prediction models more universal,which is of engineering significance and academic value.In the problem of energy consumption prediction,time series model and machine learning algorithm are the common intelligent methods.This thesis studies and improves some shortcomings of ARIMA and XGBoost.The main work is as follows:(1)In the short-term energy consumption prediction problem,the parameter estimation method of ARIMA model has low accuracy.In order to improve the accuracy of the model,the p ARIMA model is proposed.The parameter estimation problem in the ARIMA model was transformed into an unconstrained optimization problem,the problem was solved by the improved powell method,and the p ARIMA model was proposed.Because the complex nonlinear characteristics of energy consumption data,EMD-p ARIMA model was proposed,which combines EMD and p ARIMA model.The results of comparative experiments using real data show that the EMD-p ARIMA model prediction accuracy is higher than that of EMD-ARIMA,p ARIMA,ARIMA,XGBoost,and SVM in short-term energy consumption prediction.(2)In the mid-term energy consumption prediction problem,the XGBoost algorithm parameters depend on empirical values,which affects the accuracy of the algorithm.In order to solve the parameter optimization problem of the XGBoost,the artificial fish swarm algorithm(AFSA)was selected to find the best parameter set of XGBoost.A New Adaptive Artificial Fish Swarm Algorithm(NAAFSA)was proposed to solve the problems of the poor performance of the local search and the falling into a local optimum value in the later stage.During the search of the NAAFSA,according to the behavior of this selection of artificial fish and the distance from the optimal value in the field of view,the step and visual of the next iteration process are adjusted.At the later stage of the algorithm,an elite strategy is introduced to retain the optimal individuals,and then the positions of other artificial fish are moved with a certain probability.The NAAFSA was applied to the parameter optimization of XGBoost,and the NA-XGBoost model was proposed,which is used for the medium-term energy consumption prediction with monthly cycle.The real data sets were used to compare with performance of the NA-XGBoost,A-XGBoost,GA-XGBoost,XGBoost,EMD-p ARIMA,and the results show that the validity and superiority of the NA-XGBoost model are best.(3)In order to improve the regression accuracy and generalizability of the energy consumption prediction model,the EMD-p ARIMA and NA-XGBoost models were combined in parallel and in tandem.The two hybrid models were used to make experiments to energy consumption predictions in hour,day and month periodicity.Comparing with other models,and the results show that the parallel model is more generalizable and higher prediction accuracy than the individual model in the three periodicity,and the tandem model performs better when there are many effective features.