A neural network classification of isolated substorms was performed, taking into account the features characterizing the peculiarities of generation of different substorm phases. For this purpose, the following classification features were chosen: the duration of the nucleation phase, the development phase, the recovery phase, and the duration of the substorm as a whole, as well as the behavior of the Bz component of the interplanetary magnetic field (IMF). The latter feature is understood as the southward rotation of the Bz component of the IMF, which determines the beginning of the nucleation phase of the substorm. These features are adopted as input series for the self-learning neural network models being created. The result of the classification neural networks is the formation of graphical images of the set of the above classification features, each of which contains information on the duration of the phases of the considered substorms. Classification neural network experiments allow us to divide substorms into five classes. The physical features of the selected classes consist in the cause-and-effect relationships between the duration of substorm phases and solar wind parameters and MMP features.