Research On Power Consumption Prediction And Anomaly Detection Based On Deep Learning

Power energy consumption is one of the most widely existing energy consumption at present.With the continuous increase of power energy acquisition equipment,more and more power energy data are collected and stored,which makes the current power energy data increasingly rich.How to analyze and utilize the collected power consumption data has become a problem of great concern at present.Power consumption data is a kind of time series data.At present,the main direction of time series data analysis is deep learning,which has become the first choice for processing time series data due to its strong data fitting ability and stability.Based on the deep learning architecture,this paper proposes the prediction method of power consumption and the abnormal detection method of power consumption.The main research contents are as follows:(1)The method of decomposition first and then prediction is adopted to solve the problem of difficulty in prediction due to the nonlinear and unstable characteristics of power consumption data.Empirical mode decomposition was adopted to decompose the power consumption data,and components with more linear and regular characteristics were obtained.Long short-term memory networks are then used to predict.The experimental results show that,compared with the single LSTM prediction,the EMD-LSTM method is more accurate in predicting power consumption.(2)The particle swarm optimization algorithm is adopted to optimize the internal parameters of the long short-term memory network in view of the problem that internal parameters of the long short-term memory network need to be selected according to human experience.This method improves the rationality of parameter selection in the network and can better play the performance of the deep learning network.The experimental results show that the PSO-LSTM method is more accurate in predicting power consumption than that of selecting the internal parameters of LSTM network according to human experience.(3)For the problem that PSO algorithm is easy to fall into the local optimal solution and the mode aliasing problem of EMD algorithm,A seagull optimization method with stronger global search ability and a complete ensemble empirical mode decomposition with adaptive noise method with better data decomposition effect are introduced.Furthermore,a combined prediction method of power consumption based on CEEMDAN method,SOA algorithm and LSTM network is proposed.(3)In view of the poor reconstruction effect of the traditional autoencoder on electrical energy consumption data,the gated recurrent unit network of deep learning is combined with the autoencoder structure.The encoder and decoder parts of the traditional autoencoder are realized by gated recurrent unit network,which gives full play to the data feature extraction capability of gated recurrent unit and data reconstruction function of autoencoder structure.Then,the GRU deep autoencoder model is used to detect abnormal electrical energy consumption data,and the abnormal points are detected according to the reconstruction error between the original data and the reconstructed data.The experimental results show that the power consumption anomaly detection method based on GRU depth autoencoder can reconstruct the power consumption data well,and the power consumption anomaly detection performance is good.