Research On Deep Learning-Based Classification Of Smart Meter Users

As the development trend of power system,smart grids have attracted more and more attention from various countries in the world.The development of intelligent metering technology up with bidirectional flows between information in both ends of power grids have laid an important position for smart grids in power system.As one of the terminals in smart grids,smart meter can assemble a tremendous of data,which provides data support for improving customer service and energy efficiency.In this paper,a deep learning-based data processing strategy for smart meters is proposed to improve users’ service.According to answer of the trial residential customers’ survey provided by Commission for Energy Regulation(CER),which contains a large amount of customers’ social demographic information includes age,family status and employment status and so on,thus it is necessary to filtering and screening data.This paper proposed a deep convolutional neural network based smart meter data feature extraction method in allusion to highly nonlinear relationship between characteristic data above and electricity consumption.Method above can automatically obtain the complex relationship between electricity consumption in different time periods and social demographic status of customers,thus helps learning features from different datasets flexibly.A simple deep neural network model is proposed in this paper to optimize the structure of deep convolutional neural network above and further improve the performance of data recognition.Model above identifies and classifies customers’ social demographic information according to features automatically extracted by convolution neural network from dataset used in this article,which helps solving the complexity problem of calculation and overfitting problem in high dimensional samples.Therefore,the proposed model has obvious advantages in predicting accuracy.Module of Transformer is added in optimized algorithms based on recurrent neural network to improve the accuracy of customer classification,which makes more accurate feature recognition for the time series of customers’ smart meter data.Compared with other neural networks,it is also simpler in calculation and more advanced in structure.Further,proposed Transformer model performs better in learning nonlinear characteristics of sequential data,which can memorize sequential locations and share weights especially in dealing with problem of information overload in long time series.At the end of paper,calculation results of the dataset verified the validity and accuracy of model and algorithms proposed compared with the others.