Natural disasters and climatic factors are the main reasons for the unplanned outages of the novel power system. With the high proportion of renewable energy access and the digital upgrade of power grids

the sensitivity of novel power system to meteorological disturbances has increased. Greater attention should be paid to meteorological disasters disaster of prevention and mitigation in power grid. In view of the correlation and patterns between meteorological factors and power grid faults

this study introduces a power grid fault early warning approach utilizing a meteorological sparse autoencoder (SAE) combined with support vector machine (SVM) technology. The method is developed by analyzing the relationship and patterns between meteorological conditions and power grid failures through the combination of sparse autoencoder and SVM

the sensitivity and anti-interference ability of power grid fault warning with a high proportion of power electronic devices (PED) are significantly improved. To address the dataset imbalance in historical meteorological and grid operation data under typical scenarios of modern power systems

this study employs the synthetic minority over-sampling technique (SMOTE) for data balancing. The initial weights of meteorological factors are optimized based on minimum information entropy principles. A stacked SAE

fine-tuned via backpropagation

is utilized to extract critical meteorological features and predict disaster types. By integrating SVM

an adaptive fault early warning model is developed to align with the flexible operational demands of contemporary power grids

enabling accurate correlation analysis between meteorological conditions and fault patterns in advanced grid architectures. Case studies demonstrate the effectiveness of the proposed model in disaster prediction and fault-type identification

offering actionable insights for enhancing the resilience of novel power systems.