Research On Power Load Forecasting Based On Load Curve Clustering And Adaboost-CART Combination Model

With the rapid development of smart grid in China,under the premise of ensuring the safety and reliability of high-voltage side transmission and distribution,the low-voltage side management system has attracted much attention because of its real-time and effective control of users'electric energy.Accurate power load forecasting is a prerequisite for load planning in management system and effectively provides data basis for energy-saving optimization.Compared with large power grid,the load of low-voltage side is greatly affected by external factors with strong volatility.The traditional single model forecasting method generally cannot get satisfactory results because it cannot fully reflect the changing rules and characteristics of low-voltage side load.However,the combined prediction model can synthesize more information and has stronger adaptive ability,so as to obtain better prediction accuracy,which has become a research hotspot in recent years.This paper aims to improve the accuracy of power load forecasting based on the electricity consumption of low-voltage users on the campus.The main works in this thesis are summarized as follows:(1)Features construction in the load prediction model.In the feature engineering,two factors are considered,one is the time and space influence factors,such as week attribute,time attribute,whether or not the working day,and the other is the meteorological factors,such as temperature,humidity,wind speed,meteorological index and so on.Because of the possible problems of related features and redundant features,14 features are selected as the input features of the prediction model by combining the feature recursive clearance algorithm and random forest algorithm.(2)For the selection of the initial clustering center of K-means algorithm in load curve clustering,this paper proposes that the average value of similarity matrix between data objects is used as the density radius to calculate the density value of each clustering object,and then the maximum density value,the maximum difference degree and the minimum compactness are used as the selection criteria of the initial clustering center.Then,for the problem of determining the best clustering number k value,an elbow criterion based on SSE is designed to automatically and accurately select the best clustering number.(3)Based on the above clustering analysis results,a prediction model based on the integration tree algorithm is established.After clustering analysis,R~2 value of random forest,GBDT and XGboost increased by 0.2,0.64 and 0.64 respectively,while RMSE value of them decreased by 6.6,16.26 and 18.74 respectively.For any day in any prediction model,the forecasting accuracy is greatly improved after the clustering analysis results are brought in.(4)Furthermore,a combined forecasting model based on Adaboost-CART is proposed.By using CART regression tree algorithm,the three integrated tree models are combined in nonlinear optimization,which avoids the complex calculation of weighting coefficient in the traditional combination model.Adaboost algorithm is used to train cart so as to adaptively adjust the combination form of each prediction model in CART,and ultimately furthermore improve the prediction accuracy.Experimental results indicate that this method has higher prediction accuracy than single prediction models and traditional combined prediction models.Compared with the optimal single model and the entropy weight method,theR~2 value of combined method is 0.935,which is 0.015 and 0.013 higher than the first two.while RMSE value is 47.61,which decreased by 4.61 and 4.49 respectively.effect when modeling and forecasting the data collected by the energy-saving regulatory platform,which can provide effective guarantee for system management and a strong data basis for energy-saving optimization.