A Wasserstein-based distributionally robust neural network for non-intrusive load monitoring

Abstract

Non-intrusive load monitoring (NILM) is a technique that uses electrical data analysis to disaggregate the total energy consumption of a building or home into the energy consumption of individual appliances. To address the data uncertainty problem in non-intrusive load monitoring, this paper constructs an ambiguity set to improve the robustness of the model based on the distributionally robust optimization (DRO) framework using the Wasserstein metric. Also, for the hard-to-solve semi-infinite programming problem, a novel and computationally efficient upper-layer approximation is used to transform it into an easily solvable regularization problem. Two different data feature extraction methods are used on two open-source datasets, and the experimental results show that the proposed model has good robustness and performs better in identifying devices with large fluctuations. The improvement is about 6% compared to that of the convolutional neural network model without the addition of distributionally robust optimization. The proposed method supports transfer learning and can be added to the neural network in the form of a single-layer net, avoiding unnecessary training times, while ensuring accuracy.