Abstract:With the gradual complication of power grid systems, traditional power grid fault handling decisions have been unable to cope with the complex and changeable power grid environment. Facing the defects such as large errors, low accuracy and high cost existing in the current decision-making methods for power grid fault handling, the deep Q-network and random forest were introduced to optimize the relational graph convolutional network, and a machine learning model for power grid fault identification and location was constructed. On this basis, the knowledge graph was studied and introduced, and an intelligent decision-making architecture for power grid fault handling based on knowledge graph and machine learning was proposed. The results show that this architecture has an accuracy rate of 95.8%, an error of 0.94%, a recall rate of 96.4%, the F1 score and AUC value are 96.3% and 0.961 respectively, the processing delay is 240.1ms, and the memory usage is 171.4MB. The cost-benefit ratio was 363.7%, the power supply restoration rate and decision update efficiency were 99.3% and 93.3% respectively, and the fault handling risk index was 0.186. Research shows that this architecture can combine the advantages of efficient calculation, reasoning and learning of physical mechanisms and knowledge-based analysis methods to conduct real-time analysis of power grid faults. And it has effectively improved the diagnostic accuracy and processing efficiency of power grid faults, providing comprehensive and intelligent decision support for power grid faults.