基于Lissajous轨迹的电能质量扰动边-云协同高效辨识框架

Lissajous Locus-based Efficient Identification Framework for Power Quality Disturbances Based on Edge-Cloud Collaboration

边缘计算的就地实时诊断模式能够有效解决传统电能质量扰动云集中识别模式存在的强信息时延问题。针对当前电能质量扰动边-云协同辨识方法中识别算法普适性差、训练模型规模庞大且下发边缘端会造成精度缺陷的难题,提出了一种基于Lissajous轨迹的电能质量扰动边-云协同辨识框架。首先,根据图像处理领域最新进展,提出了基于双相Lissajous轨迹的视觉转换方法,将扰动信号转换成具有特殊形状的轨迹图像。然后,为了在增强特征捕捉能力的同时降低计算复杂度,开发了一种轻量级循环挤压卷积神经网络以执行主辨识任务。通过边-云共享权重参数,所提框架能够实现扰动的实时辨识。为持续优化模型性能,设计了一个更深层的云端网络以辅助模型更新。最后,基于IEEE标准仿真数据集和变电站实测扰动数据集验证了所提框架的有效性。结果表明,该框架在取得优异扰动辨识泛化性能的同时,实现了云端与边缘端识别模型的同步轻量化,并通过边-云权重交互避免了训练模型下发所造成的性能损失。

The on-site real-time diagnosis mode of edge computing can effectively solve the strong information delay problem existing in the traditional cloud-based identification mode of power quality disturbances.Faced with the problems of poor universality of identification algorithms,the large size of the training model,and accuracy defects caused by sending it to the edge in the current identification method for power quality disturbance based on edge-cloud collaboration,this paper proposes a Lissajous locus-based identification framework for power quality disturbances based on edge-cloud collaboration.Firstly,leveraging the latest advancements in the image processing field,a visual conversion method based on the double-phase Lissajous locus is proposed to convert disturbance signals into locus images with special shapes.Secondly,to enhance the feature capture ability while reducing the computational complexity,a lightweight cyclic squeeze convolutional neural network is developed to perform primary identification tasks.By sharing weight parameters of edge-cloud,the proposed framework can achieve the realtime disturbance identification.To continuously optimize the model performance,a deeper network is designed at the cloud to assist in model updating.Finally,the effectiveness of the proposed framework is verified based on the IEEE standard simulation dataset and the real-time measured disturbance dataset from substations.The results show that this framework achieves excellent disturbance identification generalization performance while realizing the simultaneous lightweight of cloud-edge identification models,and avoids performance losses caused by the distribution of the training model through edge-cloud weight interaction.