Resarch On Non-intrusive Load Monitoring Algorithm Based On Deep Neural Network

Non-intrusive load monitoring(NILM)is a technology that monitors and identifies the load inside a house through a low-cost,low-risk,high-feasibility,and easy-to-maintain advanced measurement approach.By measuring and analyzing the electrical features with the main meter at the incoming line,the NILM algorithms can indirectly assess the load situation inside the house.NILM is divided into two categories,one is closely related to energy-saving consumption and intelligent power distribution,which is called consumption monitoringoriented NILM,this type of problem focuses mainly on the energy consumption of appliances within the house for fine-grained analysis;the other is fault detection-oriented NILM that can enable timely interruption in case an arcing event happens,this type of problem is to determine whether an arcing event occurs inside the house,to rapidly identify an abnormal type of load.In this paper,two categories of NILM problems are modeled and deep neural network algorithms in NILM fields are studied in depth,with the following main results:First,a novel multi-scale residual network based on dilated convolution is proposed out of the consideration of the nature of NILM problems.The network is able to preempt the gradient disappearance and gradient explosion as the network deepening and the performance enhanced.In addition,the network optimizes the traditional processing of the time-series problem from the perspective of convolution by using its parallelism.Moreover,the novel multi-scale structure proposed in this paper is able to perform feature extraction across the input feature maps from multiple receptive fields.In this paper,the characteristics and the novelty of the network are described in detail,and two network structures are proposed under two NILM categories with their own specificity.Then,an experiment case for the consumption monitoring-oriented NILM is designed.By using the popular public dataset UK-Dale,the generation of the neural network training set,validation set and test set is detailed and the rules of data filtering and preprocessing are described.In addition,the loss function is described,several mainstream networks in the same field are introduced,then comparison experiments between these networks and the novel network proposed in this paper on the same dataset and evaluation metrics are performed.The network proposed in this paper outperforms other networks in terms of F1 values and MAE for all five appliances and has good performance in terms of spatial complexity and temporal complexity,proving the superiority of the network.Finally,an experiment case for the fault detection-oriented NILM is designed.Similarly,by using the public dataset IAED,the generation of neural network training set,validation set and test set is detailed and the preprocessing rules for the data are illustrated.Also,the loss function,the mainstream networks in the field of image classification are described.Then these networks are compared with the novel multi-scale network on the same dataset for experiments.Finally,the proposed network in this paper achieves the highest F1 scores 94.16% and 87.59%on HR-mix mixed cohort and LR-mix cohort,respectively.It also performs well on some individual load types,proving the superiority of the network.