Building a Graph Signal Processing Model Using Dynamic Time Warping for Load Disaggregation.

Area of Science:

• Electrical Engineering
• Data Science
• Machine Learning

Background:

• Unsupervised Non-intrusive load monitoring (NILM) algorithms, particularly those using graph Laplacian regularization (GLR), have achieved state-of-the-art performance.
• Existing methods often require well-defined graph structures to effectively model correlations in smart meter data.

Purpose of the Study:

• To propose a novel unsupervised approach for designing an underlying graph to model correlations within time-series smart meter measurements.
• To enhance the performance of NILM by improving the graph construction and data interpretation process.

Main Methods:

• Developed a variable-length data segmentation approach to extract potential events from smart meter data.
• Assigned measurements associated with extracted events to graph nodes.
• Employed dynamic time warping to define the graph's adjacency matrix and introduced a robust cluster labeling technique.

Main Results:

• The proposed method demonstrated improved classification performance on four different datasets.
• Achieved up to a 10% improvement in classification performance compared to competing unsupervised NILM approaches.
• The novel graph design and labeling strategy effectively models correlations in time-series smart meter data.

Conclusions:

• The novel unsupervised approach for graph design significantly enhances NILM performance.
• The method offers a robust and effective way to model correlations in smart meter data for improved energy disaggregation.
• This work advances unsupervised NILM techniques by providing a more sophisticated graph-based modeling framework.