Research On Machine Learning-Based Two-Stage Task Offloading Optimization For Power Distribution Internet Of Things

Under the development trend of new power system,the power grid has gradually entered the digital and information technology,thus developing into a smart grid.The traditional smart grid has many disadvantages,and it is difficult to support the transformation to a new power system.Edge computing has brought many advantages to the smart grid.Edge devices can collect and process data closer to users,reduce data transmission delay and network congestion,improve efficiency and response speed,and edge computing enables the smart grid to better adapt to different load changes and needs,and achieve more flexible and reliable energy distribution and management.The increase in the number of services in the distribution network results in a large increase in task data.Power Distribution Internet of Things(PDIo T)is the specific application of the Internet of Things(Io T)in the distribution network.By deploying a large number of PDIo T equipment,collect parameters such as voltage,active power,reactive power,harmonics,etc.,to provide support for distribution network fault identification,status detection and other services.PDIo T utilizes a large number of devices and relies on 5G networks to collect and transmit tasks to edge servers for realtime data processing.However,how to dynamically select edge servers and channels to meet the energy saving and low-latency task offloading requirements of PDIo T devices still faces some technical challenges,such as the task offloading decision coupling between devices,the unavailable global state information,and the relationship between various quality of service(Qo S)indicators(such as energy efficiency and latency).Therefore,the research objective of thesis is the two-stage task unloading optimization algorithm in the distribution Internet of Things,and the main work is as follows:1.The system model is established for transmission rate,transmission delay and energy efficiency of transmission process,and the specific conditions of the three parameters are observed through three models.2.A joint optimization problem is constructed to maximize the weighted difference between energy efficiency and delay of components in the distribution network of things through three models;Then,the joint optimization problem is decomposed into a large scale server selection problem and a small scale channel selection problem to achieve better results.3.A two-stage task uninstallation algorithm based on machine learning is presented.In the first stage,the offloading strategy of matching game theory is adopted.Through three models,each device and server can obtain the priority value of each other,and the terminal and edge server can choose from each other in two directions,so as to solve the problem of server selection in large time scale through bilateral matching algorithm;In the second stage,based on the reinforcement learning method that can achieve the maximum reward in machine learning,an adaptive learning method is proposed ε-greedy algorithm uses the average cumulative reward to dynamically adjust the adaptive factor,so as to solve the problem of small time scale channel selection.At last,the simulation results indicate that the proposed algorithm has better performance in terms of energy efficiency and delay compared with the task unloading delay-first matching algorithm and the task unloading strategy based on the matching theory.