Threshold-Free Event Detection for Non-Intrusive Load Monitoring Using Recurrence Plot and Convolutional Neural Networks
Men‐Shen Tsai, Yen‐Kuang Lin · 2024
Event detection is a crucial process in Non-intrusive Load Monitoring(NILM). Correct detection results not only can improve the accuracy of subsequent load identification but also quickly understand the real-time status of the system, so as to maintain the stability of power supply and the security of power dispatching. In the process of electricity consumption, the change of the switching state may cause an electrical appliance to be cut off or started, which is regarded as an event in the load identification process. However, after actual measurement, it is found that in the start process of the switch, some existing household appliances may produce a large surge when starting, while some may produce only a small one. Therefore, if the method of setting the threshold is adopted to determine whether an event has occurred, it will be difficult to obtain the appropriate threshold. For this reason, in this paper, a new feature is proposed to determine whether an event has occurred. This feature is that when the root-mean-square(RMS) current is converted to Recurrence PLOT(RP), the RP will present a geometry in different colors, so that there will be a clear distinction between events and non-events. Therefore, if such images are used to train Convolution Neural Network(CNN), events and non-events can be classified in a way without threshold configuration, and event detection can be completed. In this study, two public data sets were used to verify the suitability of the method. As shown by experimental data, the proposed method can achieve an identification rate of 98%”,99%, indicating that this method is better than other existing methods.