With the construction of power grid information system, the power data presents a complex data multiple structure. In order to meet the demand of efficient response dispatching and to enable timely and accurate fault location in complex fault situations in the fault diagnosis process after the occurrence of faults, this research proposes a knowledge graph-based power fault inference diagnosis method (KG-PFD). The method proposes a graph-based inference model for equipment entities based on the established knowledge graph for power faults. The fault entities are firstly mined for feature labels in temporal order. After that, entity slicing feature coding and temporal dynamic feature coding are introduced to the attention mechanism to learn the local structure information of devices within different time levels step by step. Finally, the importance weights of the inference targets are evaluated, and the relationship inference among the faulty entities of the equipment is realized to assist real-time assisted decision making. The practical application shows that the inference accuracy of the method is through 86%, and the knowledge experience obtained by entity relationship inference through the graph database can provide effective fault disposal reference solutions for dispatchers and technical support for dispatch center staff as well as maintenance and operation personnel.