XFed-PQD: Explainable Federated Learning for Power Quality Disturbance Classification in Edge-Enabled Smart Distribution Networks

Abstract

Power-quality disturbance (PQD) recognition is increasingly being pushed toward the sensing point rather than handled only at a central server. That shift creates a useful opportunity for distributed learning, but it also exposes two practical questions: how much performance is lost when raw waveforms remain local, and whether the resulting decisions can still be interpreted. This paper examines both questions through XFed-PQD, an explainable federated-learning framework for 17 single and compound PQD categories in edge-enabled smart distribution networks. The framework combines a compact multi-branch one-dimensional convolutional network with FedAvg and FedProx training, temporal attribution, robustness testing, and deployment profiling. The controlled experiment used three seeds, 17,000 reference waveforms, ten fully participating clients, 15 centralized epochs, and 25 federated rounds. Centralized training produced 91.35% ± 0.70% accuracy and 91.33% ± 0.70% macro-F1. With FedAvg and an IID partition, the same model reached 83.89% ± 0.78% accuracy and 83.85% ± 0.89% macro-F1 while exchanging 47.57 MB of model traffic, 83.4% less than the CNN-LSTM benchmark. Client heterogeneity had a much larger effect: Dirichlet label skew reduced macro-F1 to 69.41% ± 1.70%, and FedProx reached 69.67% ± 2.05%. The resulting 0.26-percentage-point average gain was small and did not point in the same direction for every seed. In the extended Seed-42 analysis, local-to-global Integrated-Gradients cosine similarity was 0.924 ± 0.053. The same checkpoint was highly sensitive to added noise, with macro-F1 falling from 83.82% on clean data to 54.97%, 17.01%, and 3.71% at 40, 30, and 20 dB. Dynamic INT8 quantization changed macro-F1 by -0.13 percentage points, reduced serialized size by 3.44%, and increased mean CPU latency. The dataset audit also uncovered a repeated healthy waveform that crossed data splits and appeared once under the Swell label; for that reason, three-seed sensitivity results excluding the true healthy class are reported. Taken together, the results show that explainable federated PQD monitoring is feasible under the evaluated conditions, while also making clear that client heterogeneity, duplicated waveforms, noise sensitivity, and hardware-dependent runtime behavior remain central concerns.