Research On Knowledge Graph Technology For Digital Twin Model Of Station Equipment

Although the current level of station operation and maintenance in China is very high,there are still many problems in the application of equipment operation and inspection,personnel safety monitoring,robot autonomous operation and so on.It is urgent to achieve precise location mapping through the physical characterization of power grid and its operating environment,and to achieve three-dimensional spatial analysis of multi-source and multi-type data.Digital twin model is a virtual representation model of physical equipment or physical model,that is,to complete the precise location mapping of three-dimensional space of power grid equipment in virtual space.With the continuous development and maturity of Internet of Things,artificial intelligence,in-depth learning and other technologies,the data generated by intelligent devices such as power grid operation,control and monitoring are soaring geometrically,and the data generated by virtual devices in the construction of digital twin model of power grid are also growing synchronously.Therefore,it is necessary to use intelligent means to convert a large number of heterogeneous data in the digital twin model into knowledge,so as to assist relevant personnel to make decisions quickly to clear the fault when the power grid fails.As a semantic network,knowledge mapping can model entities,concepts,attributes and their relationships in the real world.It has strong expressive power and modeling flexibility,and can be widely used in simple query services such as equipment query,line query and fault query,as well as intelligent question answering and knowledge reasoning services.In this thesis,the power grid knowledge map is taken as the research object,and two different knowledge map representation learning methods are proposed according to the scale of the station and the complexity of the association between the equipment.Firstly,the embedding principle and equation design ideas of the model are understood on the basis of the existing knowledge mapping embedding model,and then the corresponding knowledge representation and learning methods of knowledge mapping are designed according to the characteristics of different stations.Aiming at the characteristics of large scale station,large number of equipment and complex relationship between equipment,a knowledge representation and learning method based on path information and grouping feature interaction is designed.In the embedded part of the model,the way of grouping feature interaction is used,and the relationship is regarded as a three-dimensional space rotation operation between the head and tail entities,which reduces the model parameters to a certain extent.In the aspect of path reasoning,the graph attention network is introduced,and the corresponding influence weights are imposed on different paths,so that the interference is avoided to a certain extent,and the effect of knowledge reasoning is optimized.Aiming at the characteristics of small number of overall equipment,simple structure and low degree of association between equipment in small stations,a knowledge representation learning method based on neighborhood graph attention network is designed,a neighborhood graph attention layer is added on the basis of the existing model,a relative attention coefficient is designed,the weight coefficients of different neighboring nodes in the neighborhood of a central node are optimized,and the central node is corrected.Considering the influence of the relationship on the head and tail entities,the circular convolution is introduced to further improve the degree of information blending between entities and relationships,and to seek deeper features in the limited information.Compared with the experimental results of several benchmark models,the results show that the performance of the designed model is better than that of the benchmark model.