Anomaly Detection for Power Quality Analysis Using Smart Metering Systems.

Area of Science:

• Electrical Engineering
• Computer Science
• Data Science

Background:

• Power quality analysis is critical for complex systems and large plants.
• Rapid anomaly detection in voltage and current signals is essential for system availability and cost-effective maintenance.
• Existing methods may lack the speed and efficiency required for on-line detection.

Purpose of the Study:

• To implement and evaluate machine learning-based anomaly detection algorithms for on-line power quality monitoring.
• To assess the effectiveness of One Class Support Vector Machine (OCSVM), Isolation Forest (IF), and Angle-Based Outlier Detection (ABOD) for rapid anomaly identification.
• To demonstrate the feasibility of deploying these algorithms directly on sensor nodes for low-complexity, efficient signal processing.

Main Methods:

• Utilized machine learning algorithms: One Class Support Vector Machine (OCSVM), Isolation Forest (IF), and Angle-Based Outlier Detection (ABOD).
• Applied algorithms for on-line clustering and anomaly detection of voltage and current signals.
• Established an experimental platform for method evaluation, using consistent hyperparameters and Principal Component Analysis (PCA) for data processing.

Main Results:

• The proposed anomaly detection algorithms demonstrated rapid and efficient identification of signal anomalies.
• Models achieved high performance metrics, including 100% recall and up to a 92% F1 score.
• The computational simplicity allows for direct implementation on sensor nodes.

Conclusions:

• Machine learning algorithms, specifically OCSVM, IF, and ABOD, are effective tools for on-line power quality anomaly detection.
• The proposed solution offers a low-complexity, efficient method for enhancing system availability and reducing maintenance.
• Further classification algorithms can be employed for in-depth investigation once anomalies are detected by the initial algorithms.