<p>Timely detection and mitigation of cascading failures are critical to preventing large-scale power system blackouts and ensuring grid stability. The proposed framework suggests a new direction in finding the power system failure initiations using a deep learning framework based on an event-triggered hybrid system model. It combines the continuous dynamic behavior of the power grid and the discrete event-triggered transitions to conduct timely cascading failure detection. The proposed approach captures past knowledge about blackouts and system parameters by constructing a deep neural network with responses from relay protection, frequency regulation, and dispatching. The detection performance is enhanced by incorporating the physical nature of the power grid dynamic responses into the hybrid model. We show experimental results on a large-scale synthetic grid that proves this model outperforms traditional time-driven models both in terms of accuracy and efficiency. This work demonstrates the potential of deep learning to enhance the detection mechanisms and constitute a promising tool for grid operators in their efforts to avoid cascading failures and large-scale blackouts. Experimental validation on a large-scale 39-bus synthetic grid demonstrates that the proposed model achieves a 93.2% F1-score, improves detection accuracy by 12.6%, and reduces false alarms by over 30% compared to baseline methods. It also reduces simulation runtime by up to 70% compared to fixed-step models.</p>