Situation Awareness Of User Side Of Power Distribution System Based On Data-driven And Machine Learning Methods

Considering situation awareness technology(SAT)on user side which is composed of extraction,identification and predictive evaluation,a situation awareness technology based on data-driven and machine learning methods is proposed by making full use of AMI(advancing metering infrastructure)data.Four kinds of research works are performed including abnormity assessment for low-voltage users,troubleshooting priority ranking of none-consumption users(NCUs),operation health status monitoring of special transformers,and short-term load forecasting of low-voltage users in this dissertation.Combined with the perception extraction and identification technology in SAT,an abnormity assessment algorithm based on CRITIC(CRiteria Importance Though Intercrieria Correlation)and radar chart methods is proposed for low-voltage users.First,indexes that characterize the consumers abnormal features of power consumption and supplies are extracted from the original AMI data.Then,an abnormity assessment algorithm is used to determine power consumers' abnormal features of power consumption and supplies by using the extracted indexes,in which the weights of indexes are determined by the CRITIC method and the assessment value of abnormal features is determined by the improved radar chart method.Next,the abnormity assessment algorithm is used again to assess power consumers' power consumption and supply abnormities.Finally,the effectiveness of the proposed algorithm is demonstrated in case studies by troubleshooting results of low-voltage users in Zhejiang Province,China.Combined with the perception extraction and identification technology in SAT,a troubleshooting priority ranking algorithm of NCUs based on decision tree and data-driven methods is proposed.First,decision tree is utilized to determine types of NCUs.Then,the key factors suitable for NCUs filtering are determined to filter NCUs that cannot be screened by the decision tree.On this basis,CRITIC and radar chart methods are adopted to determine weights of the key factors and to determine the filtering results of NCUs,respectively.Finally,the NCUs power-supplied by an actual power supply station in Zhejiang province are served for demonstrating the proposed algorithm of NCUs filtering,and the simulation and on-site inspection results show that the proposed data-driven method is effective for screening out the abnormal NCUs.Combined with the perception extraction and predictive evaluation technology in SAT,an operation health status monitoring algorithm of special transformers is proposed based on BIRCH(Balanced Iterative Reducing and Clustering Using Hierarchies)clustering and Gaussian Cloud methods.The algorithm is composed of two parts,i.e.,the offline and online parts.For offline part: The operating indexes of special transformers are extracted based on historical operating data,and Gaussian clouds of normal operation conditions of the special transformers are determined by BIRCH clustering and Gaussian cloud methods.For online part: The operating indexes of special transformers are extracted based on real-time operation data,and Gaussian clouds of real-time operating conditions of special transformers are determined by BIRCH clustering and Gaussian cloud methods.Finally,the monitoring results of operating health status are determined by the distance between the normal Gaussian clouds and the real-time Gaussian clouds of special transformers.The algorithm in this dissertation earns a rather good accuracy verified by the abnormal sample database.Combined with the perception extraction and predictive evaluation technology in SAT,a short-term load forecasting method for low-voltage users based on deep belief neural network is proposed.First,the load data of low-voltage users are normalized,and the DBN(Deep Belief Network)training sample sets and test sample sets are formed based on the normalized load data.Then,the DBN model is trained and adjusted by the training sample sets and test sample sets respectively.Finally,the history short-term load data of the users are input to predict the users load data in next time.The short-term load forecasting algorithm in this paper has high accuracy verified by the actual load data.In summary,data-driven and machine learning methods are adopted in this paper,combined with perception extraction,identification and predictive evaluation technology in SAT,to provide power security and energy-saving services for low-voltage users in power distribution systems.