Multivariate Constrained Elastic Matching With Application in Real-Time Energy Disaggregation

Abstract

Non-Intrusive Load Monitoring (NILM) aims to estimate the power consumption of electrical appliances from the aggregated power consumption. While recent machine learning approaches have demonstrated very high disaggregation accuracies, ensuring real-time capability is crucial in NILM’s hardware implementations. We propose a constrained elastic matching approach for NILM to reduce execution time significantly. Our approach was tested on two datasets (REDD and AMPds2). The reported performance is on average 93.2% in terms of estimation accuracy for deferrable loads using the AMPds2 dataset. The proposed approach reduces execution time by a factor of ten compared to unconstrained elastic matching techniques, achieving per-frame inference times of 3.5–12.1 ms depending on the hardware platform and model size. Memory usage for the largest model is approximately 7.5 MB, and reducing the model to 10% of reference signatures lowers active power consumption from 12.1 W to 5.2 W, representing a 57% energy saving with minimal accuracy loss. Furthermore, the proposed approach has been evaluated on five different microprocessors, demonstrating consistent runtime reduction and enabling real-time implementation of elastic matching based NILM with large reference databases.