Abstract: To address the issues of insufficient timeliness and weak generalization capability in fault warning for 110 kV transmission lines, this paper takes a feeder under the jurisdiction of a power supply company as an example. By analyzing the fault evolution mechanism and screening sensitive characteristic parameters using the Pearson correlation coefficient, a fault prediction framework based on the long short-term memory network-attention mechanism is proposed. This method constructs an end-to-end model to achieve deep extraction of temporal features and adaptive allocation of weights for critical moments, while formulating a hierarchical warning strategy. The research results demonstrate that the proposed method effectively captures the coupling patterns of multiple parameters and significantly reduces the fault misreporting rate.