Research On Deep Learning Method Based User Classification And Energy Disaggregation

The reform of the power market has prompted the power sector to begin to be user-oriented and provide users with high-quality and personalized energy services.With the wide application of smart meters,the massive user consumption data recorded by smart meters not only makes it possible to use energy data analysis,but also creates conditions for intelligent and personalized management of the power grid.At the same time,with the rapid development of deep learning methods,deep learning methods provide a feasible solution for extracting user energy consumption characteristics in complex environments.In order to further play the role of deep learning method in in-depth analysis of energy consumption data,this thesis conducts research on deep learning based user classification and energy consumption disaggregation,and proposes an auto-learning edge weights graph convolution based user classification network and an improved temporal convolutional network based parallel non-intrusive energy decomposition method.The main work of this thesis is as follows:（1）The related knowledge of user classification and non-intrusive energy use decomposition based on user energy consumption data is introduced,and the solutions to related problems are given.In addition,several typical deep learning methods used in the following research are elaborated,and the advantages and disadvantages of different methods are dissected.（2）In order to realize the user classification based on smart meter data and user information,an auto-learning edge weight based graph convolution classification network（AEW-GCN）is proposed.In this network,firstly,a attention mechanism based graph transformation layer is constructed to realize the conversion and learning from discrete ammeter data to graph data;secondly,features are extracted from graph data structure and graph node data respectively.The residual network with jump connection is introduced to filter the features,and the feature reconstruction is generated based on the screening features.Finally,the graph integral classer is constructed to output the classification results.To verify the classification performance of this method,it is applied to the actual user classification experiment of CER dataset,and many classification methods such as SVM,CNN and GRU are compared.The experimental results show that the proposed AEW-GCN network has good performance on the actual user classification problem.（3）In order to explore the specific equipment energy consumption of users,an improved temporal convolution network based（DBB-TCN）parallel non-invasive energy decomposition method is proposed.In this method,firstly,a data division method based on peak-valley interval is proposed to realize the effective classification of the high and low energy consumption interval of the original data.Secondly,in the learning of equipment energy consumption characteristics,the long-term trend characteristics and short-term dynamic characteristics of the equipment are extracted respectively to achieve the purpose of parallel disaggregation.Specifically,the parallel network uses an improved temporal convolution network（DBB-TCN）to extract long-term trend features and convolution auto-encoders（CAE）to extract short-term dynamic characteristics respectively.Finally,the output of long-term and short-term decomposed subnets is organically unified in the full connection layer,and the results of network decomposition are output.To prove the effectiveness of the proposed method,experiments are carried out on two energy disaggregation datasets,UK-DALE and REDD,respectively,and compared with a variety of existing classification methods.The experimental results show that the proposed parallel decomposition method based on improved temporal convolution network can effectively realize user energy decomposition and has better performance than other methods.（4）The user energy analysis methods introduced in this thesis are summarized,the shortcomings of related methods in user classification and energy disaggregation are reflected,and the next research is prospected.