Unsupervised Mixture Models on the Edge for Smart Energy Consumption Segmentation with Feature Saliency.

Area of Science:

• Data Mining and Machine Learning
• Energy Systems Analysis

Background:

• Smart meter data granularity has increased significantly, presenting challenges for traditional clustering methods.
• High-resolution data exhibits non-Gaussian distributions, unknown cluster counts, and high dimensionality, complicating pattern analysis.

Purpose of the Study:

• To develop an innovative learning framework for effective clustering of high-resolution smart meter data.
• To enable concurrent feature and model selection for improved energy consumption pattern discernment.

Main Methods:

• Integration of the expectation-maximization algorithm with the minimum message length criterion.
• Proposal of a bounded asymmetric generalized Gaussian mixture model with feature saliency.
• Validation using three feature extraction methods across synthetic and real-world smart meter datasets.

Main Results:

• The proposed algorithm demonstrates superior clustering efficacy compared to state-of-the-art methods.
• Identified clusters effectively highlight variations in residential energy consumption patterns.
• Achieved an average performance improvement of 7.828% over the non-bounded variant of the mixture model.

Conclusions:

• The developed framework provides actionable insights for utility companies in demand reduction efforts.
• The method is robust and applicable in real-world smart meter environments, including edge cloud computing.
• The proposed bounded asymmetric generalized Gaussian mixture model offers significant advantages over other tested models.