Cloud-edge collaborative state monitoring and fault diagnosis for smart distribution system operation and maintenance

Abstract

Smart distribution systems integrate distributed photovoltaics, electric-vehicle chargers, power-electronic loads, and automated field devices, making feeder operating states highly dynamic and increasing the risk that local disturbances propagate into service interruptions. This paper proposes a cloud-edge collaborative state-monitoring and fault-diagnosis method, EIQ-CaT, for real-time operation and maintenance (O&M) of low-voltage distribution areas. The edge node converts synchronized voltage-current streams into continuous wavelet transform (CWT) time-frequency maps and uses depthwise separable convolution, channel attention, a compact temporal Transformer, and 8-bit integer (INT8) quantization to identify normal, sag, swell, interruption, harmonic, flicker, transient, and compound states. A confidence and safety gate retains high-confidence diagnosis and bounded control decisions locally, whereas uncertain events, compressed waveforms, and multimodal asset evidence are uploaded for cloud digital-twin review and model update. Camera, infrared, and light detection and ranging (LiDAR) flags are used only as diagnostic context, thereby preserving electrical interpretability. Tests on 120,000 simulated and field-noise-emulated windows achieved 99.23% accuracy, 99.18% macro-F1, 18.6 ms inference latency, and a 1.42 MB model size. In a control-in-the-loop feeder emulation, the method reduced alarm-to-command latency from 186.7 to 27.9 ms, achieved a 98.91% correct response-command rate, and completed 92.6% of events locally. The results show that cloud-edge intelligence can support continuous online monitoring, rapid abnormal-state diagnosis, and reliability-oriented automation of critical distribution infrastructure.