Deep learning algorithms have excellent performance in power system transient stability assessment, but the incomprehensibility of the assessment results and the uncontrollability of the decision-making process hinder their practical adoption by industry. A transient stability assessment model based on the Transformer encoder is proposed. The rules that the model focuses on and learns can be interpreted and analyzed by the attention weights of the features. Thus a model updating method is proposed which employs physical information in combination with interpretable results to guide model optimization. From the perspective of the loss function, the attention weight distribution of the model to the features is adjusted in a fine-tuned way to enhance the mining for the instability patterns. In the process of fine-tuning, an attention-guiding function is introduced to increase the attention weights to the key generators of specific instability patterns, so as to reduce the misclassification of specific instability patterns. In this way the overall prediction accuracy can be improved. The performance of the proposed method is verified on the IEEE39-bus system and the East China power grid system.