Non-Intrusive Load Identification Based on Retrainable Siamese Network.

Area of Science:

• Electrical Engineering
• Artificial Intelligence
• Smart Grid Technology

Background:

• Non-intrusive load monitoring (NILM) is crucial for smart grids and energy management, enabling load identification from single-point measurements.
• Current NILM methods excel at identifying pre-trained loads but struggle with scalability and identifying unknown appliances.
• The challenge of unknown load identification limits the widespread adoption and effectiveness of NILM systems.

Purpose of the Study:

• To develop a novel NILM method capable of identifying unknown electrical loads.
• To enhance the scalability and real-time adaptability of NILM systems.
• To improve the accuracy of load identification for previously unencountered devices.

Main Methods:

• A Siamese network architecture was proposed, combining a fixed Convolutional Neural Network (CNN) with two retrainable Back Propagation (BP) networks.
• The CNN extracts low-dimensional features from voltage-current (V-I) trajectories of detected unknown loads.
• Online retraining of the BP networks adapts the model in real-time, enhancing its representation ability for accurate identification.

Main Results:

• The proposed Siamese network achieved high accuracy in identifying unknown loads.
• Validation on WHITED and PLAID datasets demonstrated the method's effectiveness.
• Real-house environment tests confirmed the practicality and scalability on an embedded Linux system (STM32MP1).

Conclusions:

• The Siamese network-based NILM method offers a scalable solution for identifying unknown loads.
• Online retraining capability allows for continuous improvement and adaptation of the NILM system.
• The approach proves effective for real-world smart grid applications and energy management.