Confidence-Based, Collaborative, Distributed Continual Learning Framework for Non-Intrusive Load Monitoring in Smart...

Area of Science:

• Electrical Engineering
• Machine Learning
• Smart Grid Technology

Background:

• Non-Intrusive Load Monitoring (NILM) is crucial for smart grids and energy management, enabling appliance-level energy consumption analysis from aggregated data.
• Real-time NILM systems need continuous learning from new client data due to diverse appliances, but face inter-client conflicts and catastrophic forgetting in distributed settings.

Purpose of the Study:

• To address challenges in distributed multi-client continual learning for NILM, specifically inter-client knowledge conflicts and catastrophic forgetting.
• To propose a novel confidence-based collaborative distributed continual learning framework for enhanced NILM effectiveness and accuracy.

Main Methods:

• Developed a lightweight layer-wise dual-supervised autoencoder (LWDSAE) model for smart meter deployment, supporting load identification and confidence-based collaboration.
• Implemented a confidence judgment method using signal reconstruction deviations to enhance individual client load identification performance through model collaboration.
• Introduced an anomaly sample detection-driven method for updating model portfolios to maintain optimal local performance under quantity constraints.

Main Results:

• The proposed framework demonstrates sustained performance improvements in distributed continual learning scenarios for NILM.
• Evaluations on public datasets and real-world applications show consistent outperformance compared to state-of-the-art methods.
• The confidence-based collaboration and portfolio update methods effectively mitigate inter-client conflicts and catastrophic forgetting.

Conclusions:

• The confidence-based collaborative distributed continual learning framework offers a robust solution for real-time NILM systems facing evolving appliance data.
• The LWDSAE model and associated update strategies provide a scalable and effective approach for enhancing NILM accuracy in distributed environments.
• This research contributes to advancing smart grid capabilities by improving the reliability and adaptability of energy disaggregation techniques.