| In today’s society,under the background of increasing power demand,increasing power quality requirements,and continuous access to distributed power sources,the reactive power characteristics of the power grid have changed dramatically,the operating status has become more complicated and incomplete observation areas may appear in some district.These situations pose a serious threat to the economic and safe operation of the distribution network.In this paper,with the goal of realizing datamodel-drive,a method for voltage optimization and regulation of distribution network combining deep learning and physical model is designed to eliminate the voltage risk caused by high-permeability distributed power access,to optimize distribution scheduling strategies under different operating conditions of the power grid,to limit the switching frequency of discrete devices and to improve the accuracy of real-time voltage control of distribution networks with incomplete observation areas.The article first studies the principle of photovoltaic equipment,SC,OLTC and other controllable equipment,preprocess the data and then establishes a physical model of the distribution network system.Meanwhile,data is collected to build historical strategy library and situation awareness neural network.Finally,two data-model-drive distribution network voltage control method are proposed: based on scene discrimination,use discrete equipment to do voltage scheduling in day-ahead period;use continuous equipment to do dynamic optimization;based on historical strategy library assistance,do real-time voltage control.The main research contents are as follows:1)Day-ahead voltage scheduling and short time optimization based on scene identification.Aiming at the problem of complex and changeable operation status of the distribution network,an ELM neural network model for sensing the index value of the scene division is designed.By entering the predicted 24-hour index influencing factors,the voltage deviation and voltage fluctuation indexes are sensed to realize four types of operating scenarios the day before the division.Considering the economics of the distribution network and the number of discrete equipment switching,according to the independent optimization goals and constraints in each type of scene,finally formulated a proactive and predictable discrete equipment scheduling plan.In shorttime period,under the premise that the scene and discrete equipment are unchanged,in order to cope with the load fluctuation and the previous planning error,a multi-objective optimization model that considers scene objectives,economy and safe operation is established.In the end,the short-term control scheme of dynamic equipment is solved.2)Real-time voltage control based on historical strategy library assistance.In the real-time period,due to the incomplete observation area and other errors,the accuracy of the method of obtaining the photovoltaic control quantity through the traditional substitution voltage sensitivity calculation is low.In response to this problem,this paper introduces a deep learning method and designs a real-time voltage control strategy that uses the ELM neural network algorithm to establish a historical strategy library model and uses the historical strategy library to assist in amending the approximate voltage sensitivity calculation method.The neural network algorithm takes the normal operating state of the device as an input and solves the problem of insufficient data for deep learning.The strategy library model is used to verify the results of the approximate sensitivity method and achieved real-time voltage over-limit recovery that meets accuracy requirements after multiple iterations.In this paper,an IEEE 33 node network is revised to simulate three typical scenes in day-ahead and short time period,along with two typical scenes in real-time period.Finally the rationality and effectiveness of the proposed scheme are verified. |
