基于知识蒸馏的边缘侧电能质量扰动识别研究

Research on Edge Side Power Quality Disturbance Recognition Based on Knowledge Distillation

传统的云端电能质量扰动识别方式下,海量分布式电能质量数据会给网络负载带来巨大压力。为降低云端识别延迟,采用边缘侧扰动识别的方式,但是边缘侧智能终端计算资源有限,无法部署大规模深度神经网络。文章针对边缘侧智能终端计算资源有限和扰动识别准确率降低的问题,提出了一种基于知识蒸馏的边缘侧电能质量扰动识别方法。首先将云端训练好的性能稳定但复杂度高的深度神经网络模型进行知识蒸馏,生成一个结构简洁且运算量小的模型;然后再将蒸馏后的优化模型下发并部署在配电物联网边缘侧,直接执行电能质量扰动分类识别计算。实验结果显示,相比于现有的知识蒸馏算法,经过本方法优化过的小模型准确率提高了1.4%~2.46%。同时,与传统的云端识别方式比较,边缘侧扰动识别的数据传输速率需求降低了99.993%。表明在边端计算资源有限的前提下,基于知识蒸馏的边缘侧电能质量扰动识别方法能够满足准确率和实时性的需求。

Under the traditional cloud power quality disturbance identification method,massive distributed power quality data will bring great pressure on the network load.In order to reduce the cloud identification delay,the edge-side disturbance identification method is adopted,but the edge-side intelligent terminal has limited computing resources and cannot deploy large-scale deep neural network.In order to solve the problems of limited computing resources and low identification accuracy of disturbance in edge intelligent terminals,a method of edge side power quality disturbance recognition based on knowledge distillation was proposed.Firstly,the deep neural network model trained in cloud with stable performance but high complexity is distilled to generate a model with simple structure and small computation.Then the distilled optimization model is distributed on the edge of distribution Internet of things to directly perform power quality disturbance classification and identification calculation.Experimental results show that compared with the existing knowledge distillation algorithm,the accuracy of the optimized small model improved by 1.4%~2.46%.At the same time,compared with the traditional cloud identification method,the data transmission rate requirement of edge side disturbance identification is reduced by 99.993%.This indicates that the edge side power quality disturbance recognition based on knowledge distillation can meet the requirements of accuracy and real-time performance under the premise of limited edge-side computing resources.