The work is devoted to countering cyber attacks on electrical power systems. For this purpose, a method is proposed for predicting system states based on the previous dynamics of its functioning. The method is two-stage and consists of the following steps: 1) the stage of building a functioning model - collecting information about the states of the system, forming states, clustering states (to combine close states into one), building a recurrent artificial neural network model, training the model; 2) the stage of modeling the behavior of the system – obtaining the current state, predicting future states, checking states for attacks, assessing current security. The method is implemented in the form of a software prototype (using the architecture of an artificial neural network – LSTM), with the help of which a number of experiments were carried out on the HAI (HIL-based Augmented ICS) dataset. As a result of the experiments, the accuracy of the trained model was assessed and the probability of correctly predicting several future states of the electric power system was determined depending on various parameters (history length, number of artificial neural network nodes, etc.). The main controversial research issues are highlighted and their explanations are given.