Research On Power Load Pattern Recognition Method Based On Improved K-means Clustering Algorithm

With the intensification of smart grid and power system reform,more and more smart measurement devices have accumulated a large amount of load data.With the diversification of power grid services,data mining methods are more widely used in power grid systems to support the operation of power grid.Electric load clustering is the early basis of grid system planning,load modeling,time-of-use electricity price,demand side management,load forecasting,etc.Traditional load classification methods based on user types can no longer meet the needs of the grid business.According to the daily load curve clustering,the users can be finely divided,an accurate load pattern recognition model can also help the grid staff to realize the user's load pattern discrimination,can help the power company to find out its electricity load pattern,It provides theoretical basis for load analysis,prediction and decision-making of power systems.According to the summary of research methods of electric load clustering,this paper carries out the following work on electric load clustering and electric load pattern recognition:(1)Analyze the characteristics and structure of power load,comb the commonly used methods of load clustering,and introduce the pre-processing flow of load data.(2)For the early-stage dimensionality reduction algorithm used in load data,PCA and SAMMON mapping and dimensionality reduction methods based on characteristic indexes are compared and analyzed.The efficiency and clustering effects are analyzed,and the reference opinion on selecting the method of load clustering dimension reduction is given.(3)For the most commonly used k-means algorithm in load clustering,a comprehensive analysis and discussion are made on the selection of the number of clusters and the sensitivity to the initial clustering center.A method based on the GSA elbow criterion was introduced to determine the number of clusters.For the initial clustering center problem,a density feature and dissimilarity property based on load data was proposed to construct the Huffman tree,and the nodes of the Huffman tree were deleted by backwardst to get the initial cluster center.The validity of the above improved algorithm is verified in the test data set.The improved algorithm can reduce the number of iterations and obtain a stable clustering result.(4)On the basis of the above load clustering,an accurate load pattern recognition model is implemented through a gradient boosting decision tree model.Finally,a set of platforms for load clustering analysis and visualization is implemented based on B/S architecture.