Load Characteristic Analysis Methods For Electric Power Big Data

With the wide application of sensors in modern smart grids and the development of computer storage technology,the level of power grid informatization has been improving,and a large amount of data has been accumulated in various information systems of the power sector.The load data of power users is an important part of the big data of electricity and a true reflection of users’ electricity consumption behavior,which is the decision support for making policies in the field of electricity distribution.How to carry out effective and accurate load characteristics analysis of power users in a data-driven way,so as to obtain effective information that can help improve operational reliability and economic and social benefits,is one of the hot issues of concern for power-related departments.In this paper,a series of data-driven modeling methods for user load feature analysis based on the nonlinear and non-smooth global-local dynamic characteristics of user load information in electric power big data are proposed.The main research results are obtained as follows:（1）To address the problems of nonlinear and local dynamic features in electric power big data,and the single daily load curve cannot completely reflect the dynamic behavior of customers’ electricity consumption.By combining the K-means clustering algorithm,the temporal extension of neighborhood preserving embedding（TNPE）algorithm and Bayesian（Bayes）classification algorithm,a hierarchical classification model（H-TNPE-Bayes）for the electricity consumption industry is proposed.The results show that the proposed classification model can effectively identify the local dynamic features in the load data and obtain high classification accuracy with fewer training samples.（2）To address the problem that there are non-smooth and local dynamic features in electric power big data and the existing density-peaks clustering algorithm cannot solve the global-local multi-grain features in electricity consumption behavior,a multi-grain clustering model（MG-DPC）based on Variational mode decomposition（VMD）and densitypeaks fast search（DPC）is proposed.The validity of the model is verified using real load datasets and compared with different clustering algorithms for direct clustering of the original load data,and the experimental results show that the model can achieve effective clustering at different time granularities.（3）To address the problems of nonsmooth and autocorrelated features in electric power big data,and the prediction accuracy is not only related to the model structure but also related to the temporal characteristics of the input data.By combining the phase space reconstruction algorithm,the variational modal decomposition（VMD）algorithm,and the extreme learning machine（ELM）algorithm,an extreme learning machine-based load multigranularity forecasting model（MG-ELM）is proposed.The case results show that the prediction accuracy of the load is affected by the size of the time window of the input data,and the size of the optimal time window differs for different modal components.The optimal time window size can be effectively estimated by the chaotic phase space reconstruction method,which can improve the prediction accuracy of the load model.