Embedded Artificial Neural Networks for High-Precision Fault Location in Distribution Networks with High DG Penetration

Abstract

The large-scale integration of Distributed Generation (DG) into radial distribution networks has weakened the performance of conventional impedance-based protection schemes, mainly because of bidirectional power flow and the associated infeed effect. To overcome these limitations, this paper introduces a lightweight, data-driven diagnostic framework designed for real-time implementation on resource-constrained embedded platforms. The proposed methodology relies on a dual-stage neural inference engine operating on a high-dimensional feature set that includes symmetrical components and wavelet-based transient descriptors extracted from signals sampled at 10 kHz. A key contribution of this work lies in the design of a feature space aimed at reducing DG-related bias effects, as illustrated through PCA and t-SNE analyses. Extensive validation on a dataset of more than 23,200 fault scenarios shows that the proposed method clearly improves upon conventional approaches, which produced localization errors greater than 14 km under active-grid conditions. The optimized framework combines a Multi-Layer Perceptron (MLP) for fault-type classification, reaching 99.44% accuracy, with a regression stage that lowers the Mean Absolute Error (MAE) to 78.9 m, corresponding to a 99.2% improvement over conventional estimation. In addition, a hardware-oriented benchmarking study against SVM, k-NN, and Random Forest models shows that the MLP offers the most favorable trade-off, with a deterministic execution time of 48.2 μs and a memory footprint of 284 KB. Overall, these results suggest that the proposed solution is a strong candidate for future integration into industrial protection relays, with potential benefits for both grid reliability and maintenance efficiency.