Non-intrusive appliance classification in smart grids using swin transformer mixtures

As smart grids evolve, non-intrusive appliance classification has become an important enabler of demand-side energy management and grid intelligence. However, accurately identifying appliances from aggregate low-resolution power measurements remains challenging due to concurrent appliance operation, temporal variability, and similar consumption signatures across appliances. To address these challenges, we transform consumption windows into complementary time-series images to provide richer spatiotemporal representations and enable robust image-based learning and classification. We propose a Mixture of Swin Transformers (MoST) in which a pretrained Swin backbone extracts features from stacked multi-modal images while a mixture-of-experts transformer head performs adaptive appliance classification. To improve sample efficiency and accelerate convergence, we fine-tune ImageNet-pretrained Swin weights via transfer learning. For resource-constrained deployment, we further optimize MoST using Neural Architecture Search (NAS). Experiments on ENERTALK and REFIT show that MoST with multi-modal fusion outperforms existing methods, achieving up to 100% multi-class accuracy and over 94% multi-label F1-score.