Non-intrusive Appliance Load Disaggregation Using Graph-based Signal Processing

With the further progress of smart grid,smart power utilization is one of the important parts.Load monitoring can be used as a key technology of the smart power utilization and is also the foundation to carry out energy conservation.This thesis conducted a load monitoring method based on the theory of graph signal processing——non-intrusive appliance load disaggregation.This method operate on total active power data that is already collected by ordinary smart meters,using software analysis and tools to decompose it into major contributing household appliance loads,then we acquire power consumption details of residential users.The randomness in total power changes of household is inevitable,there often exist abnormal values in measurement data,which will lead to load events wrongly detected.In order to ensure the reliability of monitoring data which is the basis of load disaggregation,in this thesis we will eliminate noise and outliers from the total power measurements by data preprocessing such as down-sampling and median filtering.We then use difference in power amplitude as the algorithm input and also the load characteristic,adopting event-based method.In contrast to method based on probabilities,this algorithm doesn't need any training to model appliance by applying graph signal processing to load disaggregation.We construct graph signal using power data and then seek the optimal solution for graph signal smoothness as reference to cluster and analyze load events.This algorithm consists of two main blocks,one is event detection and clustering,the other is feature matching utilizing amplitude and time interval meanwhile.Graph signal processing is applied in updating adaptive threshold,clustering and feature matching.Final outputs are positive and negative edges for load identification,then it's possible to calculate disaggregated power consumption for each appliance.This thesis finally evaluate the system performance by several metrics both from local and overall.We implement this algorithm by MATLAB and proposed dichotomy to determine the scaling factor of the adjacency matrix,gives partial load disaggregation results,energy proportion and metric evaluation results.In general,the algorithm can work at low sampling rate and tolerate a degree of electric noise with small complexity.It's suitable to apply to household load disaggregation,proving the feasibility of graph signal processing in the field of load disaggregation.