To address the susceptibility of existing traveling-wave head identification methods to noise and wavefront distortion

as well as the lack of interpretability in improved approaches based on deep learning

a fault traveling-wave head identification method based on Granger causality is proposed. First

the propagation characteristics of fault traveling waves are analyzed

and a mathematical model of the traveling-wave signal at the detection terminal is established to reveal the causal relationships among sliced signal segments

which are determined by the reflection

refraction

and transmission processes. Subsequently

adjacent sub-samples are constructed as potential causal pairs

and the Granger causality method is employed to calculate the causal strength between each pair of adjacent sub-samples. Feature and adjacency matrices are then constructed accordingly. Finally

a graph neural network and long short-term memory (GNN-LSTM) network model is developed to accurately identify the arrival time of the fault traveling-wave head. Experimental results demonstrate that the proposed method exhibits good adaptability to both hybrid and multi-branch lines and is less affected by variations in transition resistance

noise

fault inception angle

and fault type. It also maintains high detection accuracy under changes in distribution network topology

demonstrating good engineering application prospects.