An Interpretable Voltage Sag Identification Method Based On Knowledge Graph

Accurate identification of voltage sag types is of great significance to maintaining stable operation of the power grid.The data-driven deep learning method has made remarkable achievements in the identification of sag types.However,most current models for power quality disturbance identification are black-box models.Although they have high recognition accuracy,the uninterpretable nature of the deep learning model is There are potential threats in the application of power quality disturbance identification,which has become an obstacle to limit its further application in the field of power quality.Therefore,it is necessary to study the interpretability model of voltage sags to realize accurate identification of sag event types,so as to take timely countermeasures and reduce economic losses.To overcome this weakness,this paper investigates the interpretability of deep models.The research on the interpretability of deep learning models aims to increase the transparency of deep neural network models and improve the interpretability of model decision results to users.The knowledge graph has endogenous interpretability,which contains human prior knowledge to form a structured knowledge system,including basic concepts,general rules,expert experience and other related structured information,which can be used for intelligent tasks,such as reasoning and decision-making,endowing network models with deeper understanding capabilities.Specifically,the innovation of this paper are as follow:(1)The knowledge atlas in the field of voltage sag is constructed,which effectively organizes and stores the massive data stored in the power system,promotes the integration of the knowledge atlas and the deep learning network model,helps to improve the understanding of the data characteristics of the model,and makes preparations for improving the interpretability of the voltage sag classification model.(2)The VGG16-Attention-LSTM sag classification model is built to make up for the shortcoming that the traditional single network model cannot align the image features with the semantic feature space.Through this model,the sag knowledge map is introduced,and the knowledge vector is embedded in the model and combined with the sag recording image for learning,which promotes the feature alignment between the knowledge space and the image space,realizes the guidance and constraint of expert experience knowledge and general rules on the network model,and further improves the interpretability of the deep learning model.(3)The semantic interpretation of sag recognition results using natural language is proposed,which overcomes the problems of experience error and violation of user’s cognition in using machine language interpretation,generates higher quality interpretation,and improves the user’s understanding of network model decision results.The experimental results show that the research results in this paper can help users understand the decision-making basis of the model,enhance human-machine mutual trust and human-machine coordination,and have practical reference and application value for improving the interpretability of the application of the deep learning model in the power system.