Abstract:

To address the difficulty in fault feature extraction for multi-branched distribution network fault location under weak fault conditions, this paper proposes a fault location method based on multi-dimensional feature extraction, which integrates the convolutional neural network (CNN), bidirectional gated recurrent unit (BiGRU), and attention mechanism (ATT). Firstly, the traveling wave characteristics of different fault locations and fault branches are analyzed. A wavefront calibration method based on the line segment detector (LSD) is employed to extract information features such as the coordinates, amplitudes, and slope of fault wavefronts, and principal component analysis (PCA) is used to construct a multi-dimensional fault-feature space that maps to fault locations. Then, a CNN-BiGRU-ATT fault location model is established to deeply explore the correlations between temporal features, amplitude features, and fault locations. Finally, classification and regression tasks are integrated to achieve both fault section identification and precise fault location. Under the condition of limited samples, the fault section location accuracy reaches 99.6429%, the accurate location error is 55.77 m, and the cross-condition error is as low as 2.95 m. The results show that the proposed model can effectively correlate multi-dimensional fault features with fault information, and exhibits superior stability of fault location accuracy and scenario generalization ability compared with the comparative models.

Key words: fault location, multi-branched distribution network, LSD, multi-dimensional fault features, CNN-BiGRU-ATT