Lite-NILMFormer: A Lightweight Nonintrusive Load Monitoring Method for Smart Metering Applications

Abstract

Nonintrusive load monitoring (NILM) extracts appliance-level energy consumption from aggregate power measurements and is essential for smart metering systems. However, Transformer-based methods such as NILMFormer suffer from high computational complexity and distribution shifts across households, limiting edge deployment and measurement consistency. To address these issues, this article proposes Lite-NILMFormer, a lightweight NILM method for smart metering applications that formulates load disaggregation as a physics-informed measurement inversion process. It combines recurrent weight sharing and low-rank parameterization (LRP) with an edge enhancement module (EEM) and convolutional gating to reduce model complexity while preserving switching transients and local waveform features. Experiments on five appliances from UK-DALE and REFIT show 67.9 % fewer parameters and 28.9 % fewer floating-point operations (FLOPs) than NILMFormer while preserving competitive accuracy and energy consistency; bidirectional cross-dataset tests further support robustness across households and datasets. These results indicate that Lite-NILMFormer improves deployability for resource-constrained smart metering systems while retaining competitive measurement-oriented performance.