Power Data Mining Technology Based On Deep Learning And Transfer Learning

Power industry is an important basic energy industry for the country development,and the first basic industry for the economy,which controls the lifeblood of the country.As the scale of the power grids continues to expand and the operating conditions become increasingly complex,the scope and frequency of data collection for power grids continue to increase.How to make use of the huge power data,improve the utilization rate of the data,and provide a theoretical basis for the safety and reliability of power grid operation become new research topics.The power data is large in quantity,which has many types and low value density.It also needs immediate processing.How to efficiently mine and analyze it,extract valuable information,and serve practical problems is a challenging problem.According to the characteristics of the power data,this paper uses the artificial intelligence methods such as deep learning and transfer learning to establish a data mining model for power grid fault detection,fault diagnosis and load forecasting,extracting the correlation features of data and improving the accuracy and effectiveness of data mining.The main research contents and contributions of this paper are as follows:First,for the problems of little fault abnormal data,the local minimum,and the vanishing or explosion gradient in the traditional neural network,this paper proposes a line trip fault detection network based on stacked sparse autoencoder,using SSAE for unsupervised learning and mining high dimensional sparse features.PCA is introduced to compress features.The Gaussian kernel SVM classifier is used for the final fault identification.Second,considering that the underlying characteri,stics of fault type,the vanishing gradient of RNN,and the overfitting problem,this paper propeses a line trip fault diagnosis network based on MLSTM network.Three weighted LSTM subnetworks are utilized to extract and fuse the temporal features of the fault electrical quantity.Dropout and Batch normalization layers are used to solve the overfitting problem caused by few fault samples in power systems.Then,for the comprehensive influencing factors of load forecasting,electricity consump-tion characteristics of users,network parameters and convergence speed,this paper proposes a short-term load forecasting network for power grid users based on GRU networks.The cluster analysis algorithm is used to decrease the interference of different electrical consumption char-acteristics.The environmental information is quantified for the auxiliary input.The historical load data and the quantified environmental information are used as the input of the network to mine the deep relationship between load forecasting and multi-source information.Finally,in order to solve the problem of sample collection and further improve data utilization and network performance,this paper proposes corresponding data mining models based on transfer learning and MMD for the above three data mining methods.MMD is used to measure the distribution difference between the source domain and the target domain data.Then the transfer learning models are chosen and adjusted according to MMD.In this process,the valuable knowledge of the source domain is transferred to the target domain so that the negative transfer is prevented.In general,the research of this paper aims at the power data mining problem based on deep learning and transfer learning.Data mining models are designed for the main data mining tasks such as fault detection,fault diagnosis and load forecasting.According to the real data experi-ment of China Southern Power Grid,the proposed method effectively improves the performance and efficiency of data mining.