Non-Intrusive Load Disaggregation Using Sequence-to-Point Integrating External Attention Mechanism

Abstract

Non-intrusive load disaggregation (NILD) can deeply explore the value of customer power consumption data, providing an important reference for decision analysis such as power equipment fault monitoring and demand response. Aimed at the conflict between the training time and the accuracy of non-intrusive load disaggregation, a non-intrusive load disaggregation algorithm using sequence-to-point integrating external attention (EA) mechanism is proposed. First, the original data is pre-processed by data purification, normalization, and some other operations, and the train data is built with a same length window. The equipment feature is extracted through the encoder layer. Then, the feature weights of important parts are enhanced by introducing an external attention mechanism. Finally, the results are yielded through the decoder layer. Simulation calculation of the proposed model and the current mainstream model is performed using the publicly available datasets, REDD and UK-DALE, while the indicators of signal aggregate error, mean absolute error, normalized disaggregation error, model disaggregation curves, feature map, and user energe consumption are compared and analyzed. The proposed model overcomes the shortcomings of attention scattering in the convolutional layer, enhances the ability to extract and utilize effective information, and has a more accurate decomposition accuracy without increasing the training time cost.