Accurate situational assessment and severity rating are of great importance to the voltage stability of power grid. Traditional approaches depend heavily on the network parameters and component models, which restrict their applications. In this paper, an unsupervised situational assessment scheme is proposed to achieve a voltage stability margin-based, three-class situation categorization via the knowledge-aided siamese autoencoder and k-Means clustering. The distribution characteristic of voltage stability margin is utilized to provide support for searching optimal feature subspace that enables k-Means to minimize intra-class and maximize inter-class differences through the siamese architecture. Experiments on IEEE-39 system prove that the proposed scheme outperforms classical approaches in multiple indicators, which proves it a useful situational assessment tool for power grid voltage stability monitoring.