Research On Non-intrusive Load Monitoring Algorithm Based On Data Fusion And Neural Network

In the aspect of intelligent power consumption,load monitoring is one of its important technologies,which can play an important role in the acquisition of power consumption information for users and maintain a good interactive relationship between power grid enterprises and users.Intrusive load monitoring is the main method used in load monitoring.It needs to install data acquisition device in all electrical equipment.Non-intrusive load monitoring only needs to install data acquisition device in smart meter.Therefore,in view of the high cost of traditional intrusive power monitoring methods and the difficulty of popularization and application,non-intrusive power monitoring has more prospects.The research on non-invasive load monitoring algorithm mainly focuses on feature extraction and load decomposition algorithm.High overlap and poor performance of partial load feature are always difficult problems in feature extraction.The research difficulty of load decomposition algorithm is how to improve the classification accuracy.The innovation of the algorithm is as follows:1)Based on the BOP(bag of pattern)model,this paper improves the feature extraction method,namely DF-BOP algorithm,which makes the feature set smaller and improves its generalization ability;2)The non-traditional feature time probability table is fused with the traditional power feature to make the feature more distinguishable;3)On the basis of DF-BOP feature extraction method,BP neural network is used as load decomposition algorithm,and it is improved twice.One is to optimize the weight and threshold of neural network by particle swarm optimization algorithm,the other is to improve the iterative process,which effectively improves the classification accuracy of load identification algorithm.Finally,through a number of comparative experiments,it is proved that:1)DF-BOP algorithm has better classification performance,higher classification accuracy and less time-consuming than BOP and BOSS algorithm;2)The new composite feature is combined by non-traditional feature time probability table and traditional power feature,and the performance of the composite feature is better than other feature combinations;3)The improved BP neural network effectively improves the classification accuracy.Finally,the nilm algorithm proposed in this paper achieves 92.24%accuracy on Plaid dataset,which is higher than the traditional integer programming algorithm.