Research On Knowledge Distillation Algorithm Of Power Quality Disturbance Identification Model Based On Cloud-Edge Collaboration

In the new power distribution Internet of Things(IoT),power quality disturbance identification is a complex and diverse problem.In the large-scale power distribution IoT business,the difficulty of this task lies in the limited computing and storage resources,low recognition efficiency and heavy load pressure of massive data communication transmission.This paper proposes a power quality disturbance identification model distillation algorithm and framework based on cloud-edge collaboration,aiming to reduce cloud-edge resource utilization,improve the speed of accurate identification of power quality disturbances,and reduce the data transmission load of the power distribution IoT.The main work and contributions of this paper are as follows:(1)Aiming at the problem of limited resources available for the power quality disturbance identification business,this paper introduces a knowledge distillation algorithm.The traditional large-scale power quality disturbance identification model(the teacher model)is distilled to form a new model(the student model)with accurate identification,simple structure and low computational complexity.In the training stage,the student model fully learns the knowledge of the teacher model through the knowledge distillation algorithm.The experimental results show that the student model(the proposed model)trained by the knowledge distillation algorithm has a higher recognition accuracy than the student model trained alone;compared with the teacher model,the model designed in this paper has faster recognition speed and less memory usage.(2)In order to further improve the performance of the power quality disturbance identification model after knowledge distillation,this paper designs a hybrid knowledge distillation algorithm based on data relations,called Data Relational Knowledge Distillation(DR-KD).When the teacher model guides the student model in training,the data relational and response knowledge of the teacher are passed on to the students,so that the students can better learn the teacher’s classification ability.Experimental results show that this method effectively improves the accuracy of the model after distillation,making the student model closer to the recognition effect of the teacher model,and is more suitable for power quality disturbance data and recognition models than other existing knowledge distillation algorithms.(3)To address the massive power quality disturbance data transmission pressure and limited edge resources,this paper proposes a model distillation framework for power quality disturbance identification based on cloud-edge collaboration.This framework makes full use of the respective advantages of the cloud and edge,and focuses on the three parts of data transmission,model update,and disturbance recognition and response,and conducts related experiments on the cloud and edge.Experimental results show that the proposed framework has the advantages of small data transmission volume,strong model update performance and short response time for disturbance recognition.The proposed framework can effectively reduce the bandwidth pressure in the transmission process of the power distribution IoT,and ensure the efficiency and speed of real-time disturbance recognition on the basis of improving the recognition accuracy.In summary,the proposed method can reduce the occupation of cloud-side resources in the large-scale power distribution IoT,improve the speed of accurate identification of power quality disturbances,and relieve the bandwidth pressure of massive data transmission in the power distribution IoT.