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

Hussein Al-Bazzaz, Muhammad Azam, Manar Amayri, Nizar BouguilaSensors 2023 · DOI

TL;DR

This paper introduces a novel unsupervised machine learning framework for segmenting smart meter energy consumption data. By integrating an expectation-maximization algorithm with the Minimum Message Length criterion and a bounded asymmetric generalized Gaussian mixture model, the approach simultaneously selects relevant features and identifies optimal consumption clusters, offering a more robust and interpretable solution for utility companies.

What Problem Does This Paper Address?

The proliferation of smart meters has led to an explosion of high-resolution energy consumption data. Analyzing this data to understand consumer behavior and identify patterns for energy efficiency and demand response programs is crucial. However, this data presents several challenges: non-Gaussian distributions, an unknown number of consumption clusters, and high dimensionality with varying feature importance. Traditional clustering methods often struggle with these issues, requiring separate steps for feature selection and cluster number determination, which can be computationally expensive and suboptimal. This paper aims to address these challenges by developing a unified, unsupervised approach that can effectively handle complex energy consumption data, infer the true number of clusters, and identify the most informative features simultaneously.

Methods

The core of this work is a novel unsupervised learning framework that combines several key components:

  • Bounded Asymmetric Generalized Gaussian Mixture Model (BAGGDMM): This is the proposed probabilistic model for representing energy consumption data. Unlike standard Gaussian models, BAGGDMM can capture asymmetric distributions and bounded support regions, which are more representative of real-world energy data.
  • Expectation-Maximization (EM) Algorithm: This iterative algorithm is used to estimate the parameters of the mixture model and the feature weights.
  • Minimum Message Length (MML) Criterion: This information-theoretic model selection criterion is employed to determine the optimal number of clusters and simultaneously perform feature selection. MML aims to find the model that provides the most compressed representation of the data, effectively balancing model complexity and data fit.
  • Feature Saliency: The model incorporates feature weights that quantify the importance of each feature, allowing the algorithm to focus on the most informative attributes and effectively perform feature selection.

The novelty lies in the integration of these components into a single framework that performs model selection, parameter estimation, and feature selection concurrently, avoiding the need for separate, iterative steps.

Datasets

The authors evaluated their proposed model using two types of datasets:

  • Synthetic Datasets: These datasets were generated to mimic real-world smart meter data characteristics, including asymmetric distributions and varied shapes. They were created by clustering real-life consumption profiles and then adding Gaussian white noise with different parameters. Two synthetic datasets were used, one with five clusters and another with eight clusters.
  • Real-Life Smart Meter Datasets:
    • Irish Smart Meter Data: This dataset, obtained from the Commission for Energy Regulation (CER), contains hourly energy consumption readings from 3639 Irish residential consumers over a period of approximately 1.5 years. PCA was used for feature extraction, resulting in a 250-feature dataset.
  • UK Power Networks Smart Meter Data: This dataset comprises half-hourly energy usage readings from 3891 London households collected between November 2011 and February 2014. Nine features were extracted to summarize consumer behavior, including relative average power for different time periods, mean standard deviation, seasonal score, weekend-vs-weekday difference score, tariff type, and geo-demographic group.

Key Insights and Ideas

  • Unified Approach for Model and Feature Selection: The most significant insight is the successful integration of model selection (determining the number of clusters) and feature selection within a single unsupervised framework using MML. This contrasts with traditional methods that require separate, often iterative, steps.
  • Importance of Bounded Distributions: The paper highlights that real-world energy consumption data often exhibits bounded support, making bounded mixture models more appropriate than unbounded ones for accurate modeling.
  • Feature Saliency for Interpretability: By assigning weights to features, the model not only improves clustering performance but also provides insights into which aspects of energy consumption are most discriminative for segmentation.
  • Privacy-Preserving Analysis: The authors emphasize that their approach can analyze energy consumption patterns without needing to access potentially sensitive personal information, aligning with privacy concerns.
  • HPC for Scalability: The implementation of the model on a High-Performance Computing (HPC) facility demonstrates its scalability and efficiency for processing large-scale smart meter data, crucial for real-world utility applications.

Key Results

The proposed FSBAGGMM model demonstrated superior performance across various metrics when compared to other mixture models (FSAGGMM, BAGGMM, AGGMM) on both synthetic and real-life datasets.

  • Synthetic Data: The model achieved high accuracy, with ACC scores of 95.569% and 91.856% on the first and second synthetic datasets, respectively. It also correctly inferred the optimal number of clusters in both cases.
  • Real-Life Data:
    • Irish Dataset: The MML criterion correctly identified 3 clusters, and the proposed model outperformed others across several performance metrics, including S (0.250), CH (7.377), and DI (0.253).
  • UK Dataset: The MML criterion identified 4 clusters, and the proposed model again showed superior performance, with S (0.319), CH (1984.843), and DI (0.027) being the best among the compared models.
  • Performance Improvement: The paper states that the proposed mixture model outperforms its non-bounded variant by an average of 7.828%.

Limitations and Future Work

  • Limitations:
    • The current framework is unsupervised and cannot leverage ground truth labels.
  • While the model is robust, the paper acknowledges that many applications might benefit from supervised or semi-supervised approaches.
  • Future Work:
    • Propose a learning method for the mixture model that incorporates ground truth labels to optimize model parameters.
  • Further explore the synergy between HPC and edge cloud computing for real-time smart meter data processing.

Summary updated

BibTeX
@article{albazzaz2023unsupervised,
  author = {Hussein Al-Bazzaz and Muhammad Azam and Manar Amayri and Nizar Bouguila},
  title = {{Unsupervised Mixture Models on the Edge for Smart Energy Consumption Segmentation with Feature Saliency}},
  journal = {Sensors},
  year = {2023},
  doi = {10.3390/s23198296},
}
Figure