In order to solve the shortcoming of the low accuracy of ferroresonance fault identification of voltage transformers (VT) by the traditional microcomputer harmonic suppression device, the T-test and variance contribution rate were introduced to form an improved complete ensemble empirical mode decomposition (ICEEMD) method, and an identification method of VT ferroresonance fault based on ICEEMD feature extraction, kernel principal components analysis (KPCA) feature reduction, and random forest (RF) construction optimized by the whale optimization algorithm (WOA) was proposed. Firstly, taking the 10 kV neutral point non-grounded system as an example, the zero-sequence voltage signals under different working conditions of the distribution network are taken as input samples. Secondly, the ICEEMD was used to extract the feature of the signal, and the KPCA was used to reduce the feature dimension to form the sample set with the best feature dimension. Finally, the WOA was used to optimize the RF parameters to determine the optimal parameters, built a VT ferroresonance fault identification model, and classify the sample set. Experimental results show that the ICEEMD can effectively extract zero-sequence voltage signal features under different working states, and the accuracy of the proposed fault identification method reaches 98.33%.