Abstract
Traditional non-intrusive load monitoring (NILM) algorithms often struggle with high data requirements, the inability to identify unknown devices, and poor versatility across different environments. This paper presents an adaptive load identification model based on a Siamese network that utilizes Voltage-Current (V-I) trajectories and active power as load signatures. By shifting from classification to similarity-based matching, the model enables few-shot learning and dynamic identification of new loads. Evaluated on the PLAID and COOLL datasets, the proposed system achieves high accuracy and demonstrates superior generalization capabilities compared to standard classification models.
Introduction
Non-Intrusive Load Monitoring (NILM) is a critical technology for smart grids, enabling the disaggregation of total household energy consumption into individual appliance data using a single point of measurement. Despite its potential, current machine learning approaches face significant barriers. Most models are "data-hungry," requiring extensive labeled datasets for training. Furthermore, they typically operate as closed-world classifiers, failing when encountering previously unseen appliances. Finally, models often lack versatility, performing poorly when transferred from one household to another due to variations in electrical characteristics.
Methodology
The proposed solution leverages a Siamese network architecture to address these challenges. Unlike traditional classifiers, a Siamese network consists of two identical sub-networks that share weights. It processes two inputs simultaneously to calculate a similarity score, determining if they belong to the same category.
Load Signatures: V-I Trajectories
The primary feature used for identification is the V-I trajectory, which captures the relationship between voltage and current over a single cycle. This 2-D representation serves as a unique "fingerprint" for different appliance types (e.g., resistive, inductive, or capacitive).
The raw voltage and current signals are preprocessed into $32 \times 32$ binarized images. A line-filling algorithm is applied to ensure the trajectory is continuous, compensating for high-speed current transitions that might otherwise appear as gaps in the digital image.
Adaptive Workflow
- Feature Library: A repository of known device signatures is maintained.
- Similarity Matching: New load detections are compared against the library using the Siamese network.
- Active Power Filter: To resolve ambiguities between devices with similar V-I shapes (like a laptop vs. a low-power LED), active power ($P$) is used as a secondary filtering criterion.
- Dynamic Learning: If a detected load's similarity score falls below a predefined threshold ($TH_0$), it is flagged as a "new" device and automatically added to the feature library.
Experimental Evaluation
The model was tested using the PLAID (Plug-Level Appliance Identification Dataset) and COOLL (Controlled On/Off Loads Library) datasets. To test generalization, the model was trained on data from only one household (House 6) in the PLAID dataset and evaluated across the remaining 55 houses.
Results
The Siamese network approach demonstrated exceptional performance, achieving an $F_1$ score of 97.88% on PLAID and 97.00% on COOLL. Notably, it outperformed standard Convolutional Neural Networks (CNNs) by 3.88% in accuracy and surpassed Weighted Recursive Graph (WRG) methods by nearly 10%. The system's "plug-and-play" capability was verified as it successfully built feature libraries in new environments without retraining.
Conclusion
By reframing NILM as a similarity-matching problem, this research provides a robust, few-shot learning framework that is both adaptive and versatile. While the computational cost of matching against large feature libraries remains a challenge for real-time scaling, the Siamese network significantly advances the practicality of non-intrusive load identification in diverse real-world settings.
Summary updated