Research On User Load Feature Identification Method Based On Data Mining Technology

The fine development of power market operation mechanism and auxiliary service has put forward higher requirements for the accuracy,real-time and intelligence of load information processing.However,with the increase of renewable energy and load types in power system,its complexity and coupling gradually strengthen.In addition,the wide application of intelligent data acquisition device has produced large amount,complex structure and various kinds of power data information.According to the characteristics of user load data,it is important to study the effective and efficient load data mining and analysis technology,load clustering classification algorithm and load model extraction method for reliable operation of power grid,intelligent control of power consumption,personalized care of user power consumption,etc.Although the research on load classification is rich in recent years,it needs further improvement in the efficiency,quality and interpretability of classification results.In view of the above three aspects,this paper uses load classification method based on the combination of time sequence trajectory feature learning and supervised or unsupervised algorithm to identify the user power consumption behavior pattern.(1)Aiming at the problem that the fuzzy boundary samples in the process of clustering daily load curve lead to the large clustering quality and data dimension,which leads to the low efficiency of calculation,a clustering algorithm based on SVD kicic is proposed.Firstly,the original data is reduced by singular value decomposition,and then the daily load curve clustering analysis is carried out by maximizing the distance between classes by kicic algorithm.The data collected from load curve sampling at each time is regarded as a dimension.The weight value of dimension is extracted by SVD,and the cluster samples with weight of each dimension are formed.Then,the objective function is constructed by combining the distance between classes and the distance between classes,which can ensure the minimum distance and the maximum distance between classes,and realize the efficient and accurate clustering of load.Accurate load tag is the basis of the follow-up time sequence track feature learning.Therefore,the k-media algorithm is used to calculate the cluster center,and more accurate load label data is selected from the clustering results.(2)The existing load classification research often ignores the interpretability of classification results.A feature extraction method of time series trajectory is proposed.Firstly,a shapelet subsequence set with strong interpretability is obtained by FLAG algorithm.Then,based on shapelet transformation technology,temporal trajectory features are extracted from labeled data and unlabeled data,and training set and test set are obtained respectively for the construction of subsequent classification model.(3)Aiming at the over fitting phenomenon and "black box" problem of traditional strong classifiers,a random forest weak classifier set algorithm based on time series trajectory feature learning is used to classify load curves efficiently and accurately,and provide interpretable shapelet classification basis.Based on a large number of load curve data of a city,the timeliness,accuracy and stability of strong classifier algorithms such as neural network,support vector machine,decision tree and random forest weak classifier cluster algorithm are compared.Through the evaluation index,the random forest algorithm based on time series trajectory feature is selected as the classifier of load user fine classification.The simulation results show that the proposed model has high timeliness,accuracy and interpretability.According to the power load user fine classification model proposed in this paper,power enterprises can classify users more accurately and efficiently,mine and study the classification basis,predict user categories,and provide different power supply needs for different users.