Power Flow For Grids With Unfixed Topologies By Using Deep Neural Network

With the continuous development of new energy technologies and information technologies,a new energy utilization system,which is called Energy Internet,links power systems,renewable energy systems and Internet systems together.Energy Internet is committed to improving the energy efficiency of production and consumption side through big data analysis,machine learning and other novel technologies,and ultimately make the energy supply and demand can be adjusted dynamically in real time.As a critical problem in power systems,power flow provides fundamental and vital information in system planning,operational management,and control strategy.Therefore,it is of great practical significance to carry out research on applying deep learning which is a new artificial intelligence technology to power flow problem.Based on the power flow problem of power system,the research content of this subject is to explore the application of the relevant models and technology of deep leaning in power flow.The trained stacked auto-encoder can solve the power flow of the grid with any node number not greater than n and apply it to occasions where power flow is required,such as N-1 security verification in power system.The research done in this thesis is a new exploration of power flow,aiming to expand the application of deep learning in power system and make a preliminary work for the follow-up work.Compared to other power flow methodologies,this method is a direct method where we only need a series of basic operation of the matrix phase to obtain the result.Thus,the calculation speed is fast enough for online power flow and there is no case where convergence can not be achieved.Finally,the proposed model will be implemented by Tensorflow and trained by deep learning server.The main content contains the following four aspects:(1)First of all,a brief introduction of the development of power flow and deep learning has been adopted by reading a large number of literatures.On this basis,the essence of power flow and the correlation between power flow and deep learning will be analyzed.(2)A deep neural network called stacked auto-encoder has been used to solve the power flow problem of any grid with unfixed topologies within n nodes on the basis of expounding the connection between the essence of power flow and deep learning model.The proposed method will be performed on grids with nodes less than or equal to 50 and compared with the traditional feedforward neural network and Newton method.Results show that the stacked auto-encoder can be used for fast power flow and it is superior to traditional neural network in training and test performance,and there is no convergence problem.(3)The method above has two defects.For one thing,it is necessary to pre-train each layer to obtain the corresponding initialization parameters before training and it will certainly increase the overall training time.For another,when the scale of grid is expanded,a problem called degeneration often occurs along with the exacerbation of stacked auto-encoder.To solve these problems,we introduce a deep neural network called residual network with batch normalization.The proposed method will be performed on grids with nodes less than or equal to 14.The trained model will be applied to distribution network with 14 nodes and compared with the shallow feedforward neural network,deep feedforward neural network,deep neural network with ReLU activation function and deep neural network with ReLU activation function and batch normalization.Results show that deep residual network have better performance over other networks with the same structure(4)As the scale of grid expands,the input sparsity of the deep neural network will be greatly improved and the dimension will be geometrically increased.Therefore,inspired by the word embedding technology,we further propose a embedding technology based on the triple loss function for the processing of network input.This technique can narrow the dimension of the input data by mapping the input data to a new vector space,so as to fully exploits the underlying characteristics of the input data.The proposed method will be performed on grids with nodes less than or equal to 14.The trained model will be applied to the IEEE standard 5-node benchmark and compared with the corresponding network with no embedding layer.Results show that the embedding technology can not only reduce the training time of deep neural network,but also improve the effect to a certain extent.