Traditional fault detection methods are difficult to meet the requirements of complex fault modes, multidimensional data processing and real-time detection. Aiming at the problems existing in the fault detection of the existing electric power metering system, such as the lack of closed-loop system, insufficient fusion of spatio-temporal features, and the difficulty in balancing model lightweight and accuracy, this paper puts forward a series of improvement schemes. A CNN-LSTM hybrid model based on attention mechanism is constructed to strengthen the fusion of spatio-temporal features, a dynamic threshold generation algorithm combining kernel density estimation and sliding window statistics is designed to improve the adaptability of anomaly detection, and a three-level hierarchical classification framework is established to enhance the generalization ability of the model through ensemble learning. The overall technical framework adopts dynamic anomaly detection, spatio-temporal feature fusion, hierarchical classification and closed-loop process of situation quantitative evaluation to realize intelligent state perception of power metering system. The experiment is deployed in a provincial power grid company, and a large number of smart meter operation data are collected to construct a test set. The results show that the proposed dynamic kernel density estimation (KDE) algorithm is excellent in anomaly detection accuracy and false alarm rate, and the CNN-LSTM-Attention model has a higher F1-score in fault classification, and the compressed model shows significant advantages in edge computing performance. The situational awareness assessment model has high accuracy in quantitative assessment and spatial location of fault risk, which can effectively improve the situational awareness ability of power grid operation and provide new ideas and methods for fault detection and diagnosis of power metering system.
Zhao et al. (Sun,) studied this question.