The Research Based On Improved Deep Learning Algorithm In Household Power Load Disaggregation

Previous studies have shown that by detecting the power consumption of users and feeding back their information to users,users can be motivated to actively adjust the mode of power consumption and plan the time of power consumption reasonably,so as to achieve the goal of saving power.At the same time,the monitoring information plays an important role in improving the planning and management of power grid,enhancing residents’ awareness of energy conservation,and calling for national energy conservation and emission reduction.The non-invasive load disaggregation method has more advantages than the traditional invasive load disaggregation method.It can effectively decompose electric energy without disturbing the normal activities of the family.At present,deep learning has been successfully applied to speech recognition and machine translation,and has shown excellent learning ability and feature extraction ability.However,there are few studies on load disaggregation using in-depth learning,and the disaggregation effect of the method used is not ideal.Therefore,based on the improved deep learning method,this paper carries out the non-invasive load disaggregation research for household electrical load.Through the analysis and comparison of the deep learning related models,this paper selects the recurrent neural network.In the recurrent neural network,there are two different structures: long short-term memory neural network(LSTM)and gated recurrent unit neural network(GRU).In this paper,LSTM and GRU are studied,and the corresponding non-intrusive load disaggregation model is successfully established.The model is improved and extended to improve the accuracy of load disaggregation.The improved measures in this paper include two aspects.One is the attention mechanism,which capable of processing input information with a long sequence,and a bidirectional architecture that can capture more useful information from a lot of data.The second is the sequence-to-point structure to effectively reduce the calculation time.In this paper,four load disaggregation models are established based on LSTM and GRU respectively,and the REDD data set(Reference Energy Disaggregation Dataset,REDD)is used for data training and testing.The load disaggregation target is shared among six families in the REDD data set.Dishwasher,refrigerator,electric light and microwave are four types of loads.The test results were evaluated by the six evaluation indicator,include recall rate,accuracy,F1-Score,accuracy,total energy relative error and mean square error.The simulation results show that the improved deep learning algorithm can effectively decompose the household appliances,and it also proves that the introduced bidirectional architecture and attention mechanism can optimize the disaggregation process.At the same time,the paper compares the disaggregation results of LSTM and GRU.The experimental results show that GRU has better disaggregation accuracy than LSTM;and the sequence-to-point system used can effectively save network training time.Through the research in this paper,it provides an effective way to solve the load disaggregation problem of household electricity.