Transparent early warning technology for power system icing terminal status based on attention mechanism-enhanced CNN

Abstract Reliable icing early warning requires simultaneous assessment of line icing, terminal health, and communication availability. This study develops a transparent-connection early warning method based on a multi-source state matrix and an attention-enhanced convolutional neural network. Environmental icing, conductor mechanics, terminal operation, and link-quality variables are aligned within event-grouped sliding windows and jointly classified into normal, icing-risk, communication, power-supply, sensor, and comprehensive high-risk states. Multi-scale convolution extracts short-, medium-, and long-term patterns, while channel and temporal attention emphasize discriminative variables and evolution periods. The Attention-CNN achieved 96.84% accuracy, a 96.24% F1-score, and an AUC of 0.986; transparent-connection anomalies were identified with 95.86% overall accuracy. Under seven complex operating scenarios, average accuracy and recall were 93.42% and 92.87%, respectively. These results indicate that joint state modeling improves terminal observability, anomaly-source localization, and graded warning support for transmission-line icing monitoring under coupled physical and communication disturbances.