Abstract: It is one of the important approaches to improve the interpretability and performance by integrating knowledge into machine learning models. In this paper, a new physics-embedded machine learning framework for power system stability analysis is proposed, in which the model is trained with the prior knowledge described by the differential-algebraic equations of the dynamic process of the power system fault. Compared to the traditional methods purely relying on the massive data, the physics-embedded machine learning model directly simulates the physical process. The physical equations contained in the data are used to guide the training procedure of the neural network and constrain the decision space of the machine learning. The dynamic curves of the fault produced by the model show explicit physical meaning, which makes the results more explainable. Meanwhile, the physics-embedded framework significantly reduces the demand of the samples, which provides new ways for the few shot learning and parameter identification when the machine learning model is applied to a real system.