A Leakage Fault Diagnosis Method for Low-Voltage Distribution Areas Based on Multidimensional Feature Extraction and LightGBM

Abstract

After photovoltaic (PV) generation is connected to a low-voltage distribution area, the residual current becomes more complex, noisier, and more variable in amplitude, which may cause conventional residual-current protection devices to mis-operate or fail to operate. To address this issue, this paper proposes a leakage fault diagnosis method that integrates multidimensional feature extraction, principal component analysis (PCA), and a LightGBM classifier. Time-domain, frequency-domain, entropy, and wavelet-packet time-frequency energy features are extracted from the residual currents of eight feeders, forming 176 original features. PCA is then used to map highly correlated features into an orthogonal space, and 170 principal components are retained as model inputs. Finally, Bayesian optimization is adopted to tune the key hyperparameters of LightGBM. Experimental results show that the proposed PCA-LightGBM method achieves an accuracy of 96.83% and an F1-score of 96.84% on the test set, with a training time of 4.98 s and a single-sample diagnosis time of 0.024 ms, demonstrating high diagnostic accuracy, redundancy resistance, and deployment efficiency.