Abstract: Aiming at the problem that it is difficult to obtain small probability samples in the transmission line extreme freezing disaster warning model and it is difficult to learn online when extracting dynamic relations of features, an adaptive kernel dynamic latent variable algorithm and a transmission line extreme freezing disaster warning model are proposed. Firstly, the model uses normal data to construct an offline model based on the kernel dynamic latent variable (KDLV) and obtains the statistical limit \textT_\lim ^2. Then, the model adaptive update criterion is introduced to improve the approximate linear dependence algorithm (ALD), and the improved ALD algorithm is used to update the statistical limit \textT_\lim ^2, so as to adaptively extract dynamic latent variable features. Finally, the KDLV model is used to calculate the statistic of the test set data \textT_\textnew^2, and whether the statistic of the test set data \textT_\textnew^2 exceeds the statistical limit \textT_\lim ^2 is used as the judgment criterion. This paper uses the ice coverage data of a transmission line in northeastern Yunnan for experimental verification. Compared with the dynamic latent variable, KDLV, dynamic internal principal component analysis and time neighbor preserving embedding methods, the proposed method has the highest disaster warning accuracy, the lowest missed alarm rate and the lowest false alarm rate.