Lightweight embedded AI architecture for real-time transient fault identification in distributed photovoltaic systems

Abstract

This paper proposes a lightweight embedded inference architecture based on the fusion of temporal convolutional networks and channel attention mechanisms to address the challenge of real-time transient fault identification in distributed photovoltaic systems. Multi-layer dilated causal convolutions are used in parallel to capture multi-scale high-frequency harmonic attenuation and low-frequency phase jump characteristics in transient currents. An attention module adaptively recalibrates feature weights to suppress environmental noise interference. Structured pruning and fixed-point quantization techniques are used to deploy the model to resource-constrained embedded terminals to achieve real-time inference at the edge. Experimental results show that even in extreme noise environments where the signal-to-noise ratio drops to −10 dB, the fault identification accuracy of the proposed method remains at 84.5%, and the absolute reconstruction error during the fault transient process is consistently below 3 A. This research achieves localized, high-precision, real-time analysis of the transient characteristics of distributed photovoltaic faults, improving the response speed and robustness of the power grid protection system.