Abstract: Aiming to address the challenges of the current power system fault risk warning methods in accurately mapping the meteorological elements and faults with complex coupling and its incapability in realizing multi-dimensional meteorological feature extraction under multiple types of extreme weather as well as the problem of low fault warning accuracy, in this paper we propose a combination of supervised and unsupervised learning as an accurate warning method for extreme weather-induced faults. First, the meteorological elements and fault correlations at the moment of failure in extreme weather scenarios are analyzed, and the hierarchical analysis method is utilized to assign weights to the meteorological elements, so as to realize the risk-adaptive assignments between the meteorological elements and the faults. Subsequently, a supervised network based on the random forest algorithm is embedded after the decoder of the unsupervised learning network, aiming to enhance the correlation between the risk weight features and power faults. Finally, the support vector mechanism is employed to construct an early warning model. The risk weight features, equipment data and environmental data are utilized as inputs to realize the fault early warning of the power system. Experimental analysis demonstrates that the fault warning accuracy reaches 87.13%, with the warning system capable of providing alerts 15.19 minutes prior to the occurrence of incidents