Electricity Consumption Prediction Based On Machine Learning Algorithm

The energy problem restricts the social development.As one of the energy sources closely related to human development,electric power is allocated effectively through free trade between countries.Electricity consumption forecast is to start from the known demand for electricity and fully consider the impact of environment,holidays,months,weeks and other related factors to predict the demand for electricity in certain periods in the future.Power consumption forecast is an important part of power system planning and also the basis of power system operation.This thesis is based on machine learning algorithm,combined with the constructed power consumption data set to predict the future power consumption.To solve the common problem of low prediction accuracy in power consumption prediction,the GBDT(Gradient reconstructed Decision Tree)algorithm,SVM(Support vector machine)and LSTM(Long Short Term Memory)algorithm used in this thesis have effectively solved this problem.In addition,the feature weighting method is proposed for the phenomenon of rough treatment of feature engineering.The characteristic engineering method can effectively avoid the influence of abnormal temperature on the prediction result.Specifically,the research contents of this thesis are as follows:1.The building electricity consumption data with sufficient sample size was obtained,and the main characteristics affecting the electricity consumption were combined to construct the data set for model prediction.In combination with the data set,LSTM algorithm,GBDT algorithm,SVM algorithm and improved GBDT algorithm were respectively used for modeling and prediction.The parameters of the model and the scale of the model are adjusted to make the prediction accuracy reach the ideal state.Under the same conditions,the prediction results of the four models were compared by MAE(Monthly Absolute Error),MRE(Month Relative Error),MAPE(Mean Absolute Percentage Error)and RMSE(Root Mean Square Error).Experimental results show that the improved GBDT algorithm is better than other algorithms in generalization ability and the stability of the established prediction model.2.We put forward an improved scheme GBDT model predictions,which is based on local linear fitting forecast improvement program,and the partial linear function to fit discrete sample points in high dimensional space,we fitting the result obtained is the predictive value of a tree,the thesis gives the detailed mathematical deduction,together with data sets,and reasonable design of the experiment.Experimental results show that the local linear fitting method proposed in this thesis is feasible to improve the prediction of GBDT algorithm,which can effectively solve the problem of low prediction accuracy when the sample size is insufficient.In the comparative experiments conducted in this thesis,it is found that the improved GBDT algorithm has stronger robustness and stability than other algorithms under the same data set.