Multi-State Energy Classifier to Evaluate the Performance of the NILM Algorithm.

Area of Science:

• * Energy Systems Engineering
• * Data Science
• * Electrical Engineering

Background:

• * Smart meter deployment enables Non-Intrusive Load Monitoring (NILM) for appliance energy analysis.
• * NILM estimates individual appliance energy use from a single measurement point.
• * Current NILM evaluation metrics are insufficient for multi-state devices like refrigerators and heat pumps.

Purpose of the Study:

• * To demonstrate the inadequacy of existing metrics for evaluating NILM techniques.
• * To propose a novel metric for more realistic and accurate NILM performance assessment.
• * To improve the evaluation of NILM for complex, multi-state appliances.

Main Methods:

• * Developed a new metric integrating event classification and operational state energy estimation.
• * Utilized unsupervised clustering to identify device operational states from labeled datasets.
• * Computed a penalty threshold for predictions deviating from ground truth.
• * Experimentally evaluated state-of-the-art NILM techniques on real-world power consumption data.

Main Results:

• * Showcased the limitations of current metrics in evaluating NILM performance.
• * Proposed metric provides a more comprehensive assessment of NILM techniques.
• * The new metric accurately accounts for the operational states of multi-state devices.
• * Experimental validation confirmed the effectiveness of the proposed evaluation approach.

Conclusions:

• * Existing NILM evaluation metrics require significant improvement, especially for multi-state devices.
• * The proposed metric offers a more robust and accurate method for NILM technique assessment.
• * This work contributes to advancing the field of NILM by providing a better evaluation framework.