Abstract

Partial discharge (PD) localization is essential for the reliability and safety of high-voltage equipment. Conventional fingerprinting-based neural networks often converge slowly and risk local minima, while compressed sensing (CS) achieves high accuracy but at significant computational cost. This paper proposes a hybrid UHF PD localization framework that integrates affinity propagation clustering, particle swarm optimization-initialized backpropagation (PSO-BP) neural networks, and singular value decomposition (SVD)-enhanced CS refinement. In the offline stage, RSSI fingerprints are clustered to reduce noise and used to train PSO-BP models for coarse localization. In the online stage, the PSO-BP output guides a reduced CS reconstruction for fine refinement. Simulations on a 25×25 m² area with four UHF sensors demonstrate that the hybrid framework achieves ~35% lower training error and ~40% faster convergence than NN, with a mean localization error of 1.23 m and 82.4% of cases within 2 m. Runtime is reduced to 0.067 ms/event, confirming the method’s suitability for real-time PD monitoring in substations.

Key words: Partial discharge; Ultra High Frequency sensors; Received signal strength indicator fingerprinting; Particle swarm optimization-Backpropagation neural network; compressed sensing; Orthogonal Matching Pursuit; localization.