Underwater acoustic sensor networks (UASNs) have attracted significant attention due to their potential applications in military surveillance and disaster early warning. However, due to the low data rate, high latency, and instability of underwater acoustic communication, UASNs are highly vulnerable to cascading failures triggered by the malfunction of key nodes, which can lead to network collapse. To address this challenge, we establish a cascading failure model tailored for dynamic traffic UASNs and propose a multi-criteria key node prediction (MC-KNP) algorithm based on the novel finding that nodes exhibit varying importance levels under different network traffic conditions. Although experimental results demonstrate that MC-KNP algorithm outperforms others approaches (e.g., degree, betweenness, and load) in accurately predicting key nodes for dynamic traffic UASNs, it suffers from high computational complexity. To address this limitation, we propose a key node prediction framework named KNP-GL, which integrates a graph convolutional network (GCN) to extract features that reflect both the structural roles of nodes and their potential impact on cascading failures, and a long short-term memory (LSTM) module to capture the temporal dynamics of cascading failures. Furthermore, based on the prediction results of KNP-GL framework, we design a mitigation strategy leveraging capacity expansion to improve network resilience against cascading failures. Experimental results show that KNP-GL framework achieves approximately 90% accuracy while reducing execution time from tens of seconds to tens of milliseconds. The proposed mitigation strategy further enhances network robustness, providing both theoretical insights and practical guidance for the development of high-reliability UASNs.