Short-term Load Forecasting Based On Competitive Transfer Learning Of EMD And LSSVM

In recent years,with the distributed energy and new energy gradually into the smart grid,resulting in changes in load trends,how to accurately load forecast in a short time has become an important part of the safe and healthy operation of smart grid protection.At the same time,a large amount of data collected and collected by the smart grid system has a significant impact on the traditional data analysis model and presents a new challenge to the system how to exploit the load fluctuation factors.Due to highlight the randomness,periodicity and related trend of load fluctuation,it is necessary to explore a signal decomposition method which can improve the prediction accuracy by using the load change mechanism because of the high nonlinearity and nonstationarity of power load fluctuation.This paper is analyzed by empirical mode(EMD)decomposition,but EMD decomposition will have a boundary effect,this paper uses a signal decomposition method based on competitive migration learning to weaken the effect.In this paper,we propose a method based on EMD analysis of load fluctuation and the method of feature extraction in order to solve the problem of load change mechanism and feature extraction.First,the EMD decomposition of the load time series is obtained,and the random component,the periodic component and the trend component are obtained.Then,the relationship between the variation of each component and the candidate factors that affect the load fluctuation is analyzed,the load change mechanism is deduced,and the load forecasting eigenvalue is extracted.Finally,the de-redundancy of feature extraction is performed.This method not only analyzes the internal factors that affect the external factors of the load itself and the inherent laws of the load itself,but also highlights the time characteristics of the load and can remove the interference of the redundant candidate feature data.Experiments show that the method proposed in this paper can deduce the mechanism of load change,but also can effectively extract the characteristics.In addition,aiming at the boundary problem,this paper proposes an EMD method based on competitive migration learning,and realizes the boundary processing mechanism of "edge screening and edge extension".This method firstly extends the load data by competing with the multi-model predictions.Then,the competitive multi-model forecasting value is used as the source task of the migration learning,and the boundary effect of the EMD is weakened by moving it into the target task.At the same time,Finally,according to the characteristics of each component to train different LSSVM prediction model,and use it to each component were predicted,the reconstruction of the components of the forecast results to arrive at the final forecast value.In this paper,Guangdong Province load data set as a case study,the use of least squares support vector machine and other different algorithms and competitive migration learning model for comparative study.The results show that the proposed method can improve the prediction accuracy and achieve the expected effect.