Reactive Power Optimization Of Distribution System Based On Data Driven And Deep Learning

Reactive Power Optimization is an important means to maintain the safe,reliable and economical operation of the distribution network,as well as an effective measure to improve the voltage quality of the network and reduce line loss.Through reactive power optimization and rational allocation of reactive power to make it balanced on the spot,not only can reduce the line loss in the power grid,but also can improve the system voltage level,can bring economic and social benefits.As a lot of distributed power generation(DG)and electric vehicle(EV)are connected to the distribution network,the distribution network structure becomes more complex,and higher requirements are put forward for reactive power optimization.This paper mainly proposes a reactive power optimization(RPO)method for distribution network based on data driven.It gets rid of the limitation of the traditional reactive power optimization model and parameters,and uses the neural network learning to propose the optimization strategy more quickly.Specifically,the starting point is to reduce the distribution network line loss and voltage deviation.Firstly,the reactive power optimization of the distribution network is carried out through the traditional genetic algorithm.The reactive power control is mainly carried out by transformer tap and capacitor switching.At the same time,using the System Line Loss and Voltage Deviation two indicators to verify its effectiveness.Then,according to all kinds of electrical data and environmental data at each moment in the historical big data,as well as the corresponding reactive power optimization control strategy at that moment,the neural network is used to learn the mapping relationship between them,and the data is used as input,and the corresponding strategy is used as output to train the network.After training the network,input the data of the current time to be optimized,and the corresponding strategy can be directly obtained for reactive power optimization.Two kinds of networks are used for testing.One is to extract the characteristic indexes of the data by using the method of free entropy,and then to learn the characteristic indexes and strategies through BP neural network.The other is through deep learning,using convolutional neural network(CNN).This method has more powerful autonomous feature learning ability,and it does not need to carry out preliminary feature processing on the data.Instead,it directly uses the original feature data of the distribution system and the corresponding strategy to learn.Finally,using the historical data of one and a half years,the case study of each method are tested and compared in the improved IEEE37 node distribution system.The results show that the method based on CNN deep learning has better performance.It basically learns the mapping relation between all kinds of data and reactive power control strategy,which can significantly reduce network loss and node voltage deviation,and is even better than the traditional method at some moments.At the same time,it no longer relies on the model and parameters of the distribution system,and can provide online decisions more quickly than traditional methods.By discussing the optimization effect of each method under different DG permeability and historical data volume,it is shown that the method based on deep learning has good optimization effect even under new unknown scenarios,verifies the adaptability,robustness and generalization ability of the method based on CNN,which provides a new way for the application of big data and deep learning technology in reactive power optimization of distribution network.