[Paper Review] Lightweight and Scalable Transfer Learning Framework for Load Disaggregation
The paper introduces RefQuery, an appliance-conditioned NILM framework that keeps a pretrained disaggregation network frozen and learns lightweight per-appliance embeddings to enable scalable, edge-friendly transfer learning for multiple appliances. It demonstrates competitive accuracy with reduced adaptation cost across cross-domain datasets.
Non-Intrusive Load Monitoring (NILM) aims to estimate appliance-level consumption from aggregate electrical signals recorded at a single measurement point. In recent years, the field has increasingly adopted deep learning approaches; however, cross-domain generalization remains a persistent challenge due to variations in appliance characteristics, usage patterns, and background loads across homes. Transfer learning provides a practical paradigm to adapt models with limited target data. However, existing methods often assume a fixed appliance set, lack flexibility for evolving real-world deployments, remain unsuitable for edge devices, or scale poorly for real-time operation. This paper proposes RefQuery, a scalable multi-appliance, multi-task NILM framework that conditions disaggregation on compact appliance fingerprints, allowing one shared model to serve many appliances without a fixed output set. RefQuery keeps a pretrained disaggregation network fully frozen and adapts to a target home by learning only a per-appliance embedding during a lightweight backpropagation stage. Experiments on three public datasets demonstrate that RefQuery delivers a strong accuracy-efficiency trade-off against single-appliance and multi-appliance baselines, including modern Transformer-based methods. These results support RefQuery as a practical path toward scalable, real-time NILM on resource-constrained edge devices.
Motivation & Objective
- Motivate NILM for edge devices with constraints like low sampling rates and privacy requirements.
- Develop a scalable, multi-appliance NILM framework that avoids fixed output sets and large per-appliance models.
- Enable lightweight target-domain adaptation by learning compact appliance embeddings.
- Demonstrate cross-domain transfer efficiency and real-time feasibility on resource-constrained hardware.
Proposed method
- Use a shared 1-D CNN feature extractor to encode both reference and query windows into embeddings of dimension E.
- Condition disaggregation on concatenated reference and query embeddings with interaction terms (e_r, e_q, (e_q - e_r)^2, e_q ⊙ e_r) via a multitask head.
- Jointly predict appliance power (with z-normalization) and ON/OFF state (with sigmoid) per input window.
- Stage I training on a source domain with fully labeled data to learn reusable, appliance-conditioned disaggregation.
- Stage II target-domain adaptation freezes the backbone and learns a per-appliance reference embedding e_r^(k) of dimension E using lightweight backpropagation.
- Stage III inference uses the fixed embeddings to disaggregate target mains in real time.
Experimental results
Research questions
- RQ1Can a single shared model disaggregate multiple appliances without a fixed output set by conditioning on compact appliance embeddings?
- RQ2Does a lightweight per-appliance embedding adaptation enable effective cross-domain transfer with limited target data?
- RQ3Is the proposed RefQuery approach scalable and efficient enough for edge-deployed, real-time NILM under resource constraints?
Key findings
- RefQuery achieves competitive accuracy compared with state-of-the-art baselines across cross-domain NILM datasets (UK-DALE, REDD) and scenarios.
- In full-data adaptation, RefQuery delivers strong MAE and F1 performance, with consistency particularly high for multi-state appliances like dishwasher and washing machine.
- In limited-data adaptation, RefQuery shows the strongest overall performance, especially in F1 scores, and maintains low adaptation cost by updating only E parameters per appliance.
- RefQuery demonstrates favorable accuracy-efficiency trade-offs and reduced storage/adaptation requirements, supporting practical on-device NILM and incremental appliance addition.
Read this paper
Open the full text in the Nubint reader and ask about any passage as you read.
Read this paper on NubintThis review was generated by AI from the original paper. Please check the paper before citing it.