Non-intrusive Load Monitoring System Based On Deep Learning And Ensemble Calibration Technology

With the advancement of technology,many smart meters now can provide real-time and detailed information of total residential current or total power.If we can make full use of this information,it will greatly help us to save energy.The non-intrusive load monitoring method can decompose the power consumption and the working state of a specific load according to the aggregated information of intelligent electricity meter,so as to help people save energy.This paper proposes a non-intrusive load monitoring system based on deep learning and ensemble calibration technology.The load monitoring system proposed in this paper can be divided into four parts: event detection,transient signal recognition,feature extraction,and system calibration.In the event detection part,it is completely different from the previous anomaly detection methods.This paper performs event detection as a classification task,uses power envelope and neural network to classify the collected circuit signals,and uses some training tricks to improve the model.The final event detection module is implemented using the trained event detection model and sliding window method.In the load identification part,this paper use convolution adaptive filtering to perform simple processing on the current signal,and use a bidirectional recurrent neural network to identify the transient current signal in the sliding window where the event is detected.Finally,this paper proposes a load identification model based on the steady-state differential current and fusion features,and add the no-load category to calibrate the misdetection that occurred in the previous power envelope-based event detection module.In addition,the bagging recognition integration method is used to ensemble the load identification result of this module and the previous instantaneous current identification module as the final identification result.All the experiments in this article are analyzed on the event detection-based data set BLUED.From the experimental results,it is completely feasible to use the power envelope and neural network by performing event detection as a classification task.And the detection capability is also completely better than other anomaly detection algorithms.In addition,the load recognition algorithm using the differential current method and the fusion features proposed in this paper is completely feasible from the separate experimental results,and the stability of the entire load monitor system has also been improved after using the ensemble idea to calibrate the previous event detection and transient current recognition model.