Research On Non-invasive Load Identification Algorithm Based On Artificial Neural Network

Non-intrusive Load Monitoring(NILM)is a method to obtain the detailed information of specific electrical appliances in a household by analyzing the total load information of household electricity consumption.It can be divided into two parts: load identification and load decomposition.Giving the feedback of obtained specific electricity consumption information to users and power supply companies can promote the efficient utilization of energy and the construction of smart power grid.Compared with traditional methods such as HMM and clustering algorithm,deep neural network method has outstanding performance in non-intrusive load monitoring and has gained wide attention.This paper will be detailed in two parts,load identification and load decomposition.The task of load identification can be divided into three parts: event detection,feature extraction and load classification.For the event detection task,CUSUM algorithm is improved in this paper.By adding the dynamic threshold item and dynamic noise item that reflect the load fluctuation,the number of missed and false detection of the event detection of this method is reduced;In addition,BP neural network is applied to the event detection task,and good experimental results are obtained.For the feature extraction task,this paper uses event-based steady-state current feature,named difference current,as the feature of load classification,and uses three data processing methods,including phase alignment,switch separation and data enhancement,to effectively improve the accuracy of load classification.For load classification task,this paper uses CNN model for load classification,and proposes that BP neural network with single hidden layer can also complete load classification task.Based on the above methods,the task of load identification in this paper has achieved high accuracy and low computational complexity.The whole load identification algorithm process is simple and effective,with good performance,providing a new research idea of "simple network + appropriate features" for load identification research.For load decomposition task,because CNN network has strong feature extraction ability,this paper uses CNN-based method to process load decomposition as a fitting task.Since the power data is a sequence with specific significance,GRU network that can learn the characteristics of the sequence is considered;Compared with unidirectional network,bidirectional network has the advantage of accepting both forward and backward memories,so in this chapter,Bi GRU method is applied to load decomposition;Finally,attention mechanism is used for getting a larger receptive field in order to capture more global information.Experiments show that the Bi GRU +attention method in this chapter has good load decomposition performance.