TransSense: A multilevel attention and rotary positional embedding‐based transformer for load disaggregation

Abstract Non‐intrusive load monitoring (NILM) aims to decompose total electricity usage into appliance‐specific signals, facilitating detailed energy management without the need for further hardware installation. However, precisely isolating the consumption of specific appliances remains difficult because of overlapping signatures, transient fluctuations, and noise interference. In this study, we propose TransSense as a transformer‐based NILM framework designed to overcome these challenges by using multiscale feature extraction and improved temporal modelling. TransSense integrates global and local attention processes to capture both long‐term consumption trends and short‐term switching events to balance comprehensive contextual comprehension with resilience to localized noise. The model incorporates rotary positional embeddings to capture relative temporal connections in energy consumption patterns. This approach dramatically enhances sequential sensitivity compared to conventional absolute positional encoding. We also implement a hierarchical feature extraction module that includes pooling operations and transformer blocks to reduce redundancy and enhance significant temporal features across various time scales. The results of an extensive experimental evaluation on the Reference Energy Disaggregation Data Set (REDD) and UK Domestic Appliance‐Level Electricity (UK‐Dale) datasets show that the TransSense model performed better than the conventional benchmark methods used for comparison. These results reflect its enhanced predictive accuracy, robust identification of different load states, and strong resilience to noise. Our findings underscore TransSense's potential as a scalable and generalizable solution for intelligent energy disaggregation on future smart grid infrastructure.