Aiming at the complexity challenges of new power system fault diagnosis, this paper systematically studies the innovative application of artificial intelligence technology. By integrating deep learning (LSTM/GNN), knowledge-enhanced large language modeling and multimodal technology, an intelligent diagnosis system covering the whole chain of generation, transmission and distribution is constructed. Key technological breakthroughs include: dynamic threshold warning to achieve >95% diagnostic accuracy for hydropower stations; knowledge-enhanced model to solve the bottleneck of zero-sample defect identification (accuracy of 54.17%); and spatial–temporal mapping technology to reach 200-m fault localization in transmission grids. Practice has shown that: the early warning system on the power generation side avoids non-stopping accidents, Graph2Text technology for distribution networks supports 95% topology identification accuracy, and microgrid optimization reduces LCOE to 0.508 USD/kWh. The research further proposes quantum—classical computing, neural symbolic systems and other frontier directions, which will promote the fault diagnosis from after-analysis to prevention and provide core support for the construction of a new type of high-reliability power system. Provide core support.