Research On Event Based Non-intrusive Load Monitoring Algorithm

Energy is the material basis for the development of human society from beginning to end.Load Monitoring(LM)can realize the monitoring of customers’ electricity consumption behavior,enable the power supply side to cut peaks and fill valleys,improve utilization of resources,and help the customer side to improve electricity consumption habits and reduce power waste,it is of great significance to save energy.Non-intrusive load monitoring(NILM)has become the mainstream of technology with its advantages of low cost and privacy protection.This paper presents a study on the event detection and load identification components of non-intrusive load monitoring,two event detection algorithms are proposed: the event detection algorithm based on double sliding window and the event detection algorithm based on voting variance.On this basis,a load identification algorithm of random forest based on genetic optimization is conceived and practiced,so as to improve the effect of non-invasive load monitoring as a whole.To begin with,an event detection algorithm based on double sliding windows is suggested to address the problems of imprecise positioning of load throwing time,wrong steady-state judgment before and after throwing event,and low accuracy of load identification in current non-intrusive load monitoring.By moving the sliding window twice,after comparing the variance with the threshold to preliminarily determine the fluctuation position,increase the judgment of the data changes on both sides of the fluctuation,and accurately locate the steady-state time before and after load switching.The experimental outcomes suggest that the designed event detection algorithm corrects more than 40% of the throwing event windows,can significantly as well improve the data quality of the steady-state current feature samples,and improves 8.74% in the load identification accuracy.The next,for the existing event detection method in the high-power appliances after the switch on missed detection of low-power appliance switching events and the event occurrence time positioning inaccurate problem,proposed a voting variance-based fluctuation detection algorithm.It selects the sample point with the largest variance that satisfies the conditions within a time series as the switch event occurrence point through a voting mechanism,and later uses the extreme difference to determine the event end point.Compared with the event records in the blue dataset,the effect on the start time and end time of positioning switching events is accurate,F1 score reaches about 95%,and the error of detected start time accounts for 95.30%in the two sampling points.It is still robust when the aggregate power level changes,and the load fluctuation range can be detected to be about 40 W ～ 2000 W.After the above two algorithms are used for event detection,the load characteristics are extracted for load identification.For the problem of low accuracy of load recognition by a single machine learning algorithm at low sampling rate,a load recognition algorithm based on random forest with genetic optimization is devised and executed.It couples the idea of genetic optimization with the classification method of random forest to achieve load identification.After obtaining load features using a double sliding window based event detection algorithm,the algorithm is evaluated using the BLUED dataset,which effectively enhanced the load recognition accuracy.