［Purposes]This paper aims to study a stator single-phase ground fault protection method based on a convolutional neural network (CNN) algorithm, so as to improve the reliability of stator single-phase ground fault protection for Powerformers. [Methods] First, an improved variational mode decomposition (VMD) method was used to process fault time-series data. Next, combined kurtosis, comprehensive energy entropy, and comprehensive concavity coefficients were extracted from the decomposed intrinsic mode function (IMF) components to form a fused feature vector. Then, a radiative padding strategy was applied to enhance the feature vector dimensionality, which was input into the CNN algorithm to determine the fault identification results of the Powerformer. Finally, to verify the method’s applicability under different operating conditions, a system simulation model consisting of three Powerformers was built by using the PSCAD/EMTDC power system simulation software. [Findings]The proposed protection method enhances identification accuracy, significantly reduces the impact of different neutral point grounding methods, fault initial angles, fault locations, and transition resistances, and demonstrates stronger noise resistance. [Conclusions] The proposed protection method achieves high identification accuracy and strong reliability, making it suitable for stator single-phase ground fault protection of Powerformers under various operating conditions.