Research On State Estimation Method Based On Depth Belief Neural Network Theory

With the continuous expansion of the power system scale and the continuous interconnection of regional power grids,the traditional methods for identifying bad data and estimating the state are faced with such multi-dimensional and heterogeneous data.The difficulty is greatly increased,and which is impossible to tap latent information and patterns from massive data.In order to improve the reliability of the current system,in the context of today's data acquisition and storage technologies,deep learning has been widely applied because of its strong features,self extracting ability and processing high-dimensional,nonlinear data and other advantages.This paper starts with the Deep Belief Network-Deep Neural Network(DBN-DNN)model and deeply studies the detection and identification methods of bad data and state estimation.The main research is as follows:Firstly,this paper analyzes and studies the model and architecture of deep learning.Analyzes the widely used Convolutional Neural Network(CNN),Deep Neural Network(DNN)and Deep Belief Network(DBN),and restricts Boltzmann Machine(Restricted).Boltzmann Machine,RBM)studied DBN feature extraction capabilities and introduced its training process.Secondly,a classification model based on deep belief neural network for bad data identification is constructed based on the analysis of three deep learning models.Then several factors was introduced that affect the network training,mainly analyzed the classification ability of DBN-DNN under different learning rates,and deduced the data correction formula based on the problem of network frame changes under actual conditions,and then proposed to overcome the residual submergence.Under the premise of residual pollution,the bad data identification method adapting to the change of the grid was adopted,and the simulation of the data collected in a medium-sized city was verified.Finally,a state estimation model based on DBN-DNN is constructed by replacing softmax classification function by using PReLU(Parametric Rectified Linear Unit)function with very strong nonlinear fitting ability to deal with the low efficiency of large batch data processing in traditional state estimation.After training the model with training data,a simulation experiment was conducted on a test set of a medium-sized city,and thesituation when the structure of the grid changed was analyzed.Simulations show that the DBN-DNN estimation model can guarantee the estimation.On the premise of accuracy,reducing the calculation time and increasing the efficiency can overcome the change of the network structure.