Electricity Consumption Behavior Mining Methods Based On Big Data In Smart Grid

Large volume data resources have been accumulated increasingly with the development of smart grid.Data processing and mining requires higher im-provement due to the enhanced data sampling frequency,dramatically in-creased data volume and various data types.The electricity load data,which make up a large proportion of big data in smart grid,strictly indicate the elec-tricity consumption demands of consumers.As a result,it is worth to conduct in-depth mining on such load data for obtaining significant knowledge of elec-tricity consumption behaviors,which can assist decision support in the man-agement of electricity power distribution and consumption.However,electric-ity load data have typical characteristics including large volume,high dimen-sionality,various resources and types,high velocity and low value density,and also have close relationships with large volume data of society,economy,envi-ronment and geography.Therefore,in order to improve the application quality and scope of electricity load data,it is essential to research and develop effec-tive data analysis and mining methods according to diverse data characteristics and applications.This dissertation studies electricity consumption behavior mining meth-ods on three aspects,which are behavior characteristics,behavior correlation and behavior evolution,based on the characteristics of electricity load data and also taking into account of the applications required to be achieved.The con-tributions of this dissertation are summarized as follows:1.In terms of miming methods for electricity consumption behavior char-acteristics,first,a fused load curve clustering method is proposed for improving the validity of clustering on high dimensional load curves,so that typical load patterns of each electricity consumer can be extracted.This method employs wavelet transform to decrease the dimensions of load data,and then adopts an algorithm called cluster fusion to fuse and optimize the clustering results of two groups of features that are generated by wavelet transform.Second,after the load pattern extraction for a large number of electricity consumers,a hybrid model for consumer categorization based on load pattern similarity is proposed for obtaining explicit consumer categories and achieving new consumer classi-fication.This model adopts the algorithm of characteristic identification to identify distinct consumer categories with their characteristics from the results of load pattern clustering.Then the problem is extended from unsupervised clustering to supervised classification,which is conducting new consumer clas-sification by using the characteristics of consumer categories as labels.2.In terms of correlation analysis methods for electricity consumption be-havior,a canonical correlation analysis method with an optimal result selection mechanism is proposed for analyzing the correlations between two or more multivariable datasets,which aims to study the influences of various factors on electricity consumption behaviors.For the electricity,gas and climate datasets of each consumer,this method first groups all daily data in three datasets based on the result of multivariable daily load data clustering,and then conduct ca-nonical correlation analysis for every triplet that includes the electricity,gas and climate data in one day.The selection mechanism is used to select the op-timal result of each canonical correlation analysis based on an inverse process of the prediction method using canonical correlation analysis.Finally,all re-sults are gathered and drawn as curves with varying daily canonical correlations and weights,which can indicates fine-grained canonical correlations between factors and canonical weights of all variables in a certain period.3.In terms of analysis methods for electricity consumption behavior evo-lution,an incremental clustering method with probability strategy is proposed for updating load patterns,which can be used for load pattern evolution analysis and anomalous pattern detection,based on existed load patterns and new elec-tricity load data.This method first conducts new daily load data clustering to extract their load patterns,each of which is then processed in the phase of load pattern intergradation based on a judgment whether being kept as a new one or integrated with an existed one.Finally,an extra clustering is performed to mod-ify all integrated load patterns.In load pattern intergradation and modification,probability strategy is employed in distance measure and cluster centers calcu-lation for achieving the optimal incremental clustering results.Several import parameters are updated after one incremental clustering so that the method can be performed continuously with coming new load data.The comparison be-tween updated load patterns from sequential incremental clustering can show the evolution of electricity consumption behaviors,so that any anomalies can be detected.The above three methods and one model form a data mining system that centers on electricity load data with the consideration of data correlation and dynamism.They technically support the in-depth mining on electricity con-sumption behaviors and assist decision support system for improving the oper-ation efficiency of smart grid.