Abstract: In response to challenges posed by existing transient stability feature selection methods, which often encounter limitations in searching for the optimum combination of critical features and lack an objective criterion for determining the optimal number of key features, this paper introduces a novel approach. A transient power angle stability key feature selection method that seamlessly integrates multi-head self-attention (MHSA) and the Boruta algorithm. A deep neural network (DNN) with an MHSA model is initially constructed to execute transient stability assessments directly on the input grid features. The model dynamically adjusts attention weights during training, focusing on key features. Subsequently, the Boruta algorithm is employed to determine the number of key features. It generates a combination of real and virtual features, which the MHSA model trains to select the actual features that are higher than the maximum virtual feature weight, and the model autonomously determines the optimal number of key features. An analysis is conducted on the IEEE 39-node and 118-node systems to validate the proposed method. The results demonstrate that this approach ensures evaluation accuracy while significantly reducing the number of input features. Moreover, the key features identified exhibit higher evaluation accuracy than traditional methods.