To mitigate the limitations associated with the arduous and time-consuming identification of conventional self-organized criticality, this paper presents a novel t-SNE-BP-based framework for discerning self-organized criticality within power systems. Firstly, the OPA model is employed to simulate cascading failures and acquire the resultant loss in system load, which subsequently serves as the observed parameter for M-K validation, facilitating the construction of the state dataset. Secondly, harnessing the dimensionality reduction advantages offered by t-SNE and the learning capabilities inherent to a neural network endowed with optimized hyperparameters, an innovative t-SNE-BP-based neural network model is introduced. Lastly, through comprehensive case studies conducted on the IEEE-39 node system, the proposed model is demonstrated to surpass alternative methodologies in terms of heightened accuracy and reduced identification time. These findings effectively corroborate the efficacy and superiority of the model, furnishing a solid theoretical foundation and compelling evidence for averting major power outage incidents.