Load Clustering Analysis And Self Learning Control In Distribution Network

The development of big data technology has brought many new ideas and challenges for the power system analysis and control.Accompanied by the development of distributed power,energy storage,monitoring and protection devices,the traditional distribution network has gradually evolved into an active distribution network with observability and controllability.Combined with big data in power system,this thesis has conducted related researches from the aspect of planning and operation in distribution network.The second chapter puts forward the information collection and load estimation method of distribution network based on open source information.The following third chapter completes the urban spatial load distribution situation perception based on the improved clustering algorithm.Then the fourth chapter discusses the self-learning optimization control method of the distribution network based on the depth enhanced learning algorithm and the effectiveness of the method is demonstrated by the simulation analysis in the fifth chapters.Because the distribution network is at the end of the power system,it has the characteristics of wide distribution,large scale and difficult data collection.The existing system can not meet the requirement of measurement configuration,so is necessary to use other means to collect relevant power information.This thesis takes full advantage of the open source information provided by the network,uses data mining technology to obtain valuable power user information,and constructs a sample set of urban load spatial distribution through data screening and load estimation,which provides a new idea about accurate urban spatial load estimation.This thesis then uses the improved clustering algorithm based on the peak density which takes the local density and spacing of the sample points of each power user as the index to perceive the spatial distribution of urban load.Taking each cluster as the basic unit,by analyzing the head's spatial coordinates,the total load,the load density,the total number of users,the maximum radius,average radius,the urban power load situation perception model with cluster attribute is established.By comparing with the actual distribution network area,the development trend of the new regional load center,the migration of regional load center and the change of regional load density can be accurately grasped and it provides a reference for the distribution network planning,such as the distributed power supply and the active load access position.The distribution network with distributed generation and active load access has the characteristics of complex power flow analysis and diverse control methods,it is difficult to establish a universal model for optimal control.Based on the data provided by the measurement device,this thesis proposes a self-learning optimal control method based on data-driven of active distribution network: The principle of self-learning optimization control based on Markov decision process is given by analyzing the observation status,control means and operation target of distribution network.Then the state-action evaluation function is defined to provide reference for the selection of control action,and in combination with the ?-greedy strategy,the distribution network adopts exploratory control in the initial stage of self-learning and empirical control in the middle and later stages.Finally,combining the actual operating samples,the distribution network can autonomously select the control means to achieve the operation target based on the observation status by using neural networks to fit similar conditions.At the end of this thesis,the extended IEEE33 active distribution network is taken as an example to illustrate the problem of the voltage limit brought by the distributed power supply.By controlling the charge and discharge status of the electric vehicle charging pile,the safe and stable operation of the distribution network is realized with less switching times,thus verifying the feasibility and effectiveness of data driven self-learning and optimization control method in distribution network.