Marketing Service Center (Metrology Center) of State Grid Zhejiang Electric Power Co., Ltd., Hangzhou 310000, China

The electric energy meter, as an important data source in the low-voltage distribution network, plays a crucial role in ensuring the fairness of electricity transactions through the management of measurement anomalies. Therefore, a model of weakly-correlated measurement anomaly recognition of line loss based on neural network with variable weights is designed, whose inputs are freezing values of power and voltage on the electric energy meter, as well as partial archival information. The main part of the model is the residual neural network to calculate the measurement anomaly index. First, aimed at the accuracy degradation caused by insufficient input information, a neural network algorithm with variable weights is proposed, in which the network weights are no longer considered to remain constant after training, instead, some of them are regarded as variable weight layers formed by the superposition of several eigenstate weights. Then, through the node correlation state that is not able to be directly used as network input, the superimposing coefficients are calculated via composite density matrices with dimensions, which significantly improves the model accuracy. Finally, by using the augmented sample set for training and testing, the validity of the residual neural network recognition model with variable weights is comparatively verified. Furthermore, the model is tested and applied to the control of a high-loss distribution station area in a city in Zhejiang Province of China to confirm its detection effect in actual working conditions.