Aiming at the problems of low prediction accuracy and slow convergence speed of the existing power system network security situation prediction models. In this paper, a prediction method of artificial fish swarms algorithm-Extreme Learning Machine Algorithm (AFSA-ELM)is proposed. Firstly, the parameters of hidden layer nodes and corresponding nodes are randomly selected in the algorithm to achieve the purpose of unsupervised learning. Secondly, aiming at the problem that Extreme Learning Machine (ELM) is easy to fall into local optimum, a confidence interval strategy based on artificial fish swarms algorithm (AFSA) is proposed to limit the weight of input data and hidden layer, which can guide the result of falling into local optimum to global optimum. Through the experimental analysis, it is verified that the model has a good evaluation and prediction effect, and can predict the real-time network situation completely and accurately.