Application of Self-supervised Learning in Non-intrusive Load Monitoring

Abstract

With the proposal of smart grid, the demand of both source and load for fine monitoring and control of power load is becoming increasingly prominent. Non-intrusive load monitoring is a technical means to better meet this demand. However, the research at home and abroad focuses on the existing data sets and labeled data to improve the accuracy of load identification, while the research on the training method of the model under the massive unlabeled monitoring data in the actual scene is still in a relatively blank stage. Aiming at the problem of how to make full use of unlabeled monitoring data for model training, a non-intrusive-load monitoring method based on self-supervised learning is proposed in this paper. This method designs a self-supervised learning task, so that the model can make full use of the massive unlabeled monitoring data for training, eliminating the step of manually labelling the data; Based on the encoder-decoder structure, a deep learning model is established, and the load is identified through the load characteristic vector output by the encoder, so that the method has generalization performance. In this paper, AMPds2 data set is used to verify the method, and test examples verify the effectiveness of the method.

An intelligent non-intrusive load monitoring model based on power encoding and convolutional state modules

“…First, the accuracy of disaggregation for similar simultaneously switched devices can be further improved.Secondly, industrial equipment consumes more electricity and encompasses a wider variety of device types except for NILM for household devices; implementing NILM for industrial equipment effectively have a significant impact on energy conservation. Finally, developing unsupervised or semi supervised methods is crucial for practical applications for NILM tasks [49]; addressing at achieving low classification error in situations with limited or no labeled device data remains an important research topic [50].…”