Research On Application Of Improved Deep Recurrent Neural Network In Non-invasive Load Monitoring

The proportion of energy consumption is increasing with the increasing energy demand of residential and commercial buildings.Using non-intrusive load monitoring technology to obtain detailed load information inside the building to improve energy use is of great significance to alleviate the energy problem.Deep recurrent neural networks has achieved great success in the fields of speech recognition,machine translation and so on in recent years.Considering that the load data used for nonintrusive load monitoring is the same as the data type used for language recognition,it belongs to time series data.As a consequence,this study will use the long-term memory network,gated recurrent unit and antisymmetric recurrent neural network to build deep learning models for non-intrusive load monitoring.In addition,the attention mechanism is introduced into the models to improve the performance of the models.Finally,Bayesian optimization algorithm is used to improve the hyperparameters of the models to further improve the performance of the models.First of all,this study discusses the principle and structure of long-term memory network,gated recurrent unit and antisymmetric recurrent neural network.Besides,it discusses a common method to reduce training error in deep learning and a method to reduce generalization error.Secondly,a deep learning environment for non-intrusive load monitoring is built,and the REDD data set used in the experiment is introduced.Moreover,some of the data are selected as the experimental data.Three kinds of recurrent neural networks are used to build non-intrusive load monitoring models for this data,and the load of fridge,microwave,dishwasher and electric furnace in the data set are used as the target appliances for load monitoring to test the performance of the models.The results of the three models are compared.Thirdly,the working principle of attention mechanism in deep learning is analyzed,and the attention mechanism is introduced into the load monitoring models to improve the performance of the models.In addition,the results of the models after the introduction of attention mechanism are compared with the previous results.The influence of introducing attention mechanism on the performance of load monitoring models is analyzed.Finally,the hyperparameters are not necessarily the best hyperparameters of the models because the hyperparameters of the above models are all set according to experience.Bayesian optimization algorithm is used to improve the hyperparameters of the models in order to find the best hyperparameters of the models to further improve the performance of the models.The load monitoring results of the improved models are compared with the previous results,and the influence of Bayesian optimization on the performance of the models is analyzed.Besides,the results of the improved three models are compared.Moreover,the performance of each model when different appliances are used as target appliances is analyzed.