ELECTRIcity: An Efficient Transformer for Non-Intrusive Load Monitoring.

Area of Science:

• Artificial Intelligence
• Energy Systems
• Machine Learning

Background:

• Non-Intrusive Load Monitoring (NILM) uses aggregated household signals to infer individual appliance energy consumption.
• Sequence-to-sequence deep learning models are current state-of-the-art for NILM, but recurrent models have limitations.
• Accurate appliance disaggregation is crucial for energy management and grid optimization.

Purpose of the Study:

• To introduce ELECTRIcity, a novel transformer-based architecture for NILM.
• To leverage attention mechanisms for enhanced extraction of dependencies in power signals.
• To develop an efficient training routine for improved performance and reduced training time.

Main Methods:

• Proposed a transformer-based deep learning architecture named ELECTRIcity for NILM.
• Utilized attention mechanisms to capture global dependencies between aggregate and appliance signals.
• Implemented an unsupervised pre-training and supervised fine-tuning training strategy.

Main Results:

• ELECTRIcity demonstrates superior performance compared to existing state-of-the-art NILM methods.
• The model achieves high accuracy in estimating appliance power consumption signals.
• The proposed training routine reduces training time while improving predictive accuracy.

Conclusions:

• Transformer architectures, specifically ELECTRIcity, offer a powerful alternative to recurrent models for NILM.
• ELECTRIcity provides accurate appliance disaggregation with reduced data preprocessing and efficient training.
• The approach advances the field of energy monitoring through improved deep learning techniques.