Non-Intrusive Load Monitoring Based On Deep Learning And Edge Computing

Electric power load monitoring and its observation data could be extremely helpful to the resolution of numerous electric power issues,including power system planning & prediction,power grid stability simulative modeling,demand-response policy establishment & verification,power efficiency analyzing,etc.Traditional load monitoring system deploys sensors in electric outlets for a type or even a single of appliance to collect power usage information,which is of low cost-efficiency and maintainability,also challenging in transferring real-time sensing data.For this,NILM(a.k.a Non-Intrusive Load Monitoring)technology is designed to be deployed per room,factory and even building as basic units,and only requires collecting the electric power physics from electricity inlets,with which the power usage information within the observing area could be easily calculated by NILM algorithms.Hence,NILM system is able to effectively decrease the investments in sensor devices in an economical and environment friendly manner;Data generation is more controllable and it’s equipped with better functional extendibility and interoperability.We carried on NILM researches and discussions from the perspective of feature extraction,switch event detection,load disaggregation algorithm and system communication in this paper,where two load disaggregation models are proposed and evaluated for its generalization capability on real-life data;An edge computing architecture for NILM under Io T scenarios are proposed.Details are listed as below:Firstly,an ’Event Driven’ mode NILM neural network model is proposed,which comprises an activation detector and an appliance recognizer.Initially,the activation detector detects and extracts valid appliance activations from the sensing data,which will be passed to the appliance recognizer.In the aspect of the activation detecting algorithm,a judging strategy based on dispersion and power usage behavior regularity analysis is proposed;In the aspect of the appliance behavior recognizing algorithm,aimed at low sample frequency input feature extraction,a CNN architecture that can automatically extract activation signatures and recognize the load type is proposed.Plus,by the analysis of activation background power,power value fluctuation,etc.,three generic features are defined as supplements to the CNN extracted features.The ’Event Driven’ model only takes a single neural network to accomplish the task of load recognition,causing less computational costs.Secondly,an ’End to End’ mode NILM neural network model is proposed,comprising the feature extraction module and the feature mapping LSTM(Long Short Term Memory)module.For end to end models,the target categorical load’s time series of power value could be deduced after feeding the model with easily preprocessed sensing data.Comparatively,it requires more computational resources,due to the fact that every single load category requires at least one neural net to realize power disaggregation.However,end to end models are better at real-time capabilities and having a more detailed grasp of load’s power data.Last,from the angle of edge computing,we carried on the research of NILM systems for electric power Io T.Aimed at the communication bottleneck introduced by the particularity of smart meter’s deploy location(usually a remote switchbox which is covered with low network signal),we proposed an edge computing architecture for NILM,and discussed about the NILM task designation throughout the components.Coupled with containerization and virtualization,edge gates are provided with functional extendibility besides its basic computing capability,which reaches the demands of electric power utilities and enterprise clients;Based on the Co AP protocol,smart meters are able to achieve real-time power data transferring with low hardware resource occupation in a narrowband environment.