Semi-Supervised Graph-Based Non-Intrusive Load Monitoring Method Combining Transformer and Graph Laplacian Smoothing

With the widespread adoption of smart meters and the advancement of smart grid technologies, demand-side management—exemplified by real-time monitoring and load management—has been playing an increasingly vital role in maintaining grid stability. Non-intrusive load monitoring (NILM), which can estimate the power usage of individual appliances based solely on whole-house power readings without additional equipment, serves as a low-cost and highly sensitive monitoring method. It provides data support for demand-side management in load management and has attracted significant attention in recent years. Machine learning methods based on graph signal processing (GSP) concepts have been widely applied to low-sampling-rate NILM tasks. However, the original aggregate power sequence features are difficult to separate, and enhancing feature selection can improve GSP performance in NILM tasks. This paper proposes a new semi-supervised GSP-based NILM approach by integrating neural network feature extraction with GSP. A transformer-based encoder-decoder is trained in a supervised manner to extract high-dimensional features from the encoder outputs, which are then used to generate graphs. After calculating graph weights using a Gaussian kernel function, graph Laplacian learning is applied to smooth the signals and infer unknown labels in a semi-supervised way. This yields the operational states of each appliance for load disaggregation. This method leverages the strong feature extraction capabilities of neural networks to enhance graph signal processing performance. The approach is validated on public datasets, outperforming traditional machine learning methods and two other GSP algorithms across multiple evaluation metrics.