【目的】随着现代电网的快速发展，电能质量问题不可忽视,其不仅影响电力系统的安全运行，还可能导致设备损坏和电能浪费。针对传统的电能质量分类方法存在识别效果差、计算复杂度高和数据隐私问题，本文提出了一种结合联邦学习和边缘计算的联邦加权重采样与混合优化(FedWRHO)算法。【方法】首先，设计了一种结合卷积神经网络(CNN)和长短期记忆(LSTM)网络的混合模型，以高效提取电能质量信号的时空特征，并在边缘设备上进行本地训练。然后，将边缘计算与联邦学习相结合，设计了一个分布式的联邦学习框架，使各边缘节点利用本地数据进行训练，并在云端聚合模型，提高整体分类性能。最后，为了验证所提的FedWRHO算法及其结合CNN-LSTM的混合模型的有效性，进行了详细的试验验证。【结果】结果表明，模型对14类电能信号的分类准确率较高，大多数类别的准确率接近或超过95%；边缘训练的数据传输量和存储需求显著低于集中训练。【结论】本文所提方法，不仅在分类性能和数据隐私保护方面表现优越，还解决了计算资源分配和适应性方面的问题，具有广泛的应用前景。

[Objective] With the rapid development of modern power grids, power quality issues have become significant concerns in power systems. These issues, including voltage fluctuations, harmonics, transient pulses, and voltage interruptions, not only affect the safe operation of power systems but can also lead to equipment damage and energy waste. To address the limitations of traditional power quality classification methods, which suffer from poor recognition performance, high computational complexity, and data privacy concerns, this paper proposed a federated weighted resampling and hybrid optimization (FedWRHO) algorithm, which leverages federated learning and edge computing. [Method] First, a hybrid model integrating convolutional neural network (CNN) and long short-term memory (LSTM) network was designed to efficiently extract spatiotemporal features of power quality signals, with local training performed on edge devices. Subsequently, a distributed federated learning framework was established by integrating edge computing with federated learning, allowing edge nodes to utilize local data for training while aggregating models in the cloud to improve overall model performance. Finally, detailed experimental validation was conducted to assess the effectiveness of the proposed FedWRHO algorithm and its hybrid CNN-LSTM model. [Results] The findings indicated that the model achieved high classification accuracy for 14 types of power signals, with most categories reaching or exceeding 95% accuracy. Additionally, the data transmission volume and storage requirements for edge training were significantly lower than those for centralized training. [Conclusion] The method proposed in this paper not only demonstrates superior classification performance and data privacy protection but also effectively addresses issues related to computational resource allocation and adaptability, indicating broad application potential.

国家青年科学基金项目(52102477)