Research On Non-Intrusive Household Electrical Load Disaggregation Based On Deep Learning And Its Applications

In the framework of the smart grid,enhanced awareness of residential power consumption patterns and improved efficiency may enable the electricity supply department to adjust the generation plan in time to achieve “peak-load-shifting”.Nonintrusive load monitoring enables the contribution of a single appliance which can be disaggregated from household power consumption.This technique can capture detailed power consumption characteristics,which can be an essential promotion for improving home electricity management as well as cutting costs.In this paper,the non-intrusive load disaggregation method,home electricity load simulation data generation method,and home appliance usage optimization methods are studied for households.The details are as follows:(1)Non-intrusive load decomposition,compared to non-intrusive load monitoring solutions based on event detection,can disaggregate the power consumption of overall households straight into that of single appliances.In response to the drawback that the existing neural network-based load disaggregation model is poor for long-time activationtype appliances,this paper proposes a non-intrusive load disaggregation model based on the Self-Attention mechanism.The proposed model uses steady-state active power variation as the load feature,utilizes convolutional neural networks for feature extraction,and enhances the modeling of long-range correlations in the input sequence by SelfAttention,enabling the model to extract the most relevant features to each point.Afterward,the hyper-parameters in the model are adjusted using Bayesian Optimization to improve the accuracy.Finally,a model is trained and tested on the UK-dale dataset,and results show that the model has low errors and good generalization performance in the new environment.(2)The scarcity of data is a major constraint to further improving the performance of deep learning algorithms.Non-intrusive load disaggregation algorithms based on Deep Neural Networks rely on large amounts of appliance data,and the cost of measuring this data in the field is too high,which makes the generation of simulation data a feasible alternative.However,most studies generate simulation data by random combinations of multiple appliance data,which is difficult to reflect the actual home power consumption in this way.For this analysis,this paper analyzed the changes in the number of active occupants and the activity of electrical usage in 24 h by the number of household members,age,type of profession,and education level.Then,the appliance power consumption data is generated using a Variational Auto-Encoder-based data generation model.After that,based on electrical usage activity analysis,the household electrical load is generated by overlapping the power consumption of all electrical appliances.In the end,the generated simulation data is compared with the real data from the public dataset to check the effectiveness.(3)It is necessary to obtain the usage of appliances when optimizing household power consumption behavior.Several studies extracted appliance usage patterns from previous data or simulated appliance usage during the day with random probabilities.Nevertheless,appliance usage patterns vary greatly from household to household,and it is not feasible to extract it from the rest of the historical data when appliance power consumption is not available.In this regard,this paper obtains information on the earliest turn-on period,latest turn-on period,highest operating frequency period,and power consumption of each appliance through a non-intrusive load disaggregation technique,which provides the basis for home appliance scheduling optimization.According to users’ demand and operation characteristics of various types of household appliances,these are classified into a fixed load,interruptible load,and shiftable load,and their power consumption models are established.Then,a home appliance scheduling optimization model was established to optimize the operation time of appliances with the objective of cost-cutting,while also taking into account the influence of scheduling on the discomfort of customers.Finally,the real power consumption data is extracted from the UK-dale,Refit dataset and solved by PSO algorithm to validate the effectiveness of the method.