A Non-Intrusive Load Monitoring Method For Edge Computing

Non-Intrusive Load Monitoring(NILM)has emerged to reduce equipment costs and user adoption barriers,with non-intrusive load identification as a promising branch.However,in current traditional neural network load recognition algorithms,most researchers focus only on the network’s recognition rate,disregarding parameters,computational complexity,and hardware requirements.This results in the need for high-performance devices for load identification.Currently,NILM data is uploaded to cloud servers for computation,which incurs high costs for server construction and maintenance,contradicting the requirements of NILM.Therefore,this study aims to introduce edge computing into NILM and conducts relevant research and experiments.Firstly,the dataset,data preprocessing methods,and the commonly used event detection algorithm,CUSUM sliding window algorithm,are introduced.Then,the steady-state features and transient features commonly used in NILM are described.Three electrical features,V-If trajectory,reactive power,and current,are selected for load identification.To improve the recognition rate,an algorithm is proposed to blend these three features into a fusion RGB feature image and is compared with similar algorithms through experiments.Finally,a lightweight load recognition model based on attention mechanism,suitable for edge computing,is proposed.Leveraging the advantages of CNN in image classification,the existing lightweight neural network,ShuffleNet V2,is improved.By combining the fused RGB image containing multiple information of electrical appliances,load identification is performed.Experimental results show that this method significantly reduces the hardware requirements for edge computing,with a 97%reduction in computational power demand and a 95.6%reduction in storage demand.Moreover,it achieves a recognition accuracy of 98.5%on the PLAID dataset.