Real-Time Detection and Classification of Power Quality Disturbances

Mahsa Mozaffari, Keval Doshi, Yasin YilmazSensors 2022 · DOI

TL;DR

This paper introduces a novel, data-driven sequential method for real-time detection and classification of power quality disturbances (PQDs) in smart grids. By extending a supervised anomaly detection approach, the method achieves faster and more accurate identification of various PQDs compared to existing techniques, crucial for grid stability and equipment protection.

What Problem Does This Paper Address?

Power grids are increasingly susceptible to power quality disturbances (PQDs) due to the integration of renewable energy sources and the proliferation of modern electronic devices. These disturbances, such as voltage sags, swells, and transients, can cause significant damage to sensitive equipment, disrupt grid operations, and lead to economic losses. While numerous methods exist for detecting and classifying PQDs, there's a persistent need for approaches that are not only accurate but also capable of real-time operation, providing timely alerts and enabling swift countermeasures. Existing methods often struggle with high computational complexity, limiting their real-time applicability, or they focus solely on detection without providing classification, requiring additional steps. This paper aims to bridge this gap by proposing a unified, fast, and accurate sequential method for both detecting and classifying PQDs.

Methods

The core of this work lies in a novel supervised sequential detection method, termed "Vector-ODIT," which builds upon the principles of the Online Discrepancy Test (ODIT). Unlike ODIT, which is semi-supervised and only trains on nominal data, Vector-ODIT leverages both nominal and disturbance data for training. This supervised approach allows for more tailored and effective anomaly detection.

The method extracts a set of statistical features (e.g., RMS, standard deviation, autocorrelation, entropy, waveform length, zero crossing, average fluctuation) from voltage waveforms within small, sliding windows. These features are then used to compute an "anomaly evidence" score for each observation. This score is recursively updated, and an alarm is triggered when it exceeds a predefined threshold.

To achieve classification, the Vector-ODIT extends the detection framework into a multi-hypothesis testing problem. For each potential disturbance type (including the "no disturbance" case), a separate decision statistic is maintained. When any of these statistics exceed their respective thresholds, a disturbance is detected, and its type is identified based on which statistic triggered the alarm. This approach allows for simultaneous detection and classification.

Datasets

The paper does not explicitly mention or introduce new benchmark datasets. Instead, it relies on simulated data generated using MATLAB/Simulink SimPowerSystems toolbox. The simulations generate voltage waveforms for four types of disturbances: sag, swell, oscillatory transient, and harmonics. These simulated datasets are used for both training and testing the proposed methods.

Key Insights and Ideas

  • Supervised Anomaly Detection for PQDs: The key insight is that by incorporating knowledge of known disturbance types into the training phase of an anomaly detection framework (like ODIT), the performance in terms of speed and accuracy for PQD detection and classification can be significantly improved.
  • Multi-Meter Cooperation: The multivariate nature of the proposed detector allows for the joint processing of data from multiple meters, which can enhance detection accuracy, especially in noisy environments.
  • Unified Detection and Classification: The Vector-ODIT framework elegantly combines detection and classification into a single sequential process, eliminating the need for separate detection and classification stages and thus improving overall speed.
  • Data-Driven Approach: The method is data-driven, meaning it learns patterns from data rather than relying on pre-defined parametric models, making it more robust to the complexities and uncertainties of real-world power systems.

Key Results

  • Improved Detection Delay: The proposed supervised ODIT detector consistently outperforms the semi-supervised ODIT and GLLR methods in terms of average detection delay across various disturbance types, especially for sag and transient disturbances.
  • High Classification Accuracy: The Vector-ODIT achieves high classification accuracy, reaching 98.38% with a 0.38% false alarm rate under 20 dB SNR, and 100% accuracy with a 0% false alarm rate under 30 dB SNR.
  • Real-Time Performance: The method achieves detection and classification within, on average, 0.61 waveform cycles (approximately 39.46 samples or 0.0123 seconds), which is significantly faster than many existing methods that require multiple cycles.
  • Scalability with Meters: The performance of the supervised ODIT detector improves notably as the number of meters employed in the system increases, demonstrating its effectiveness in cooperative sensing scenarios.

Limitations and Future Work

  • Simulation-Based Evaluation: The evaluation is primarily based on simulated data. While the simulations are designed to be realistic, real-world testing on actual power system data would be a valuable next step to validate the method's performance in diverse and complex scenarios.
  • Threshold Selection: The paper mentions that setting proper thresholds is crucial for balancing false alarm rates, classification accuracy, and delay. While empirical values were used for simulations, a more systematic or adaptive approach to threshold selection could be explored for greater robustness.
  • Computational Cost for Classification: While the detection is fast, the classification aspect, involving multiple decision statistics, might introduce a slight increase in computational overhead compared to detection-only methods. The paper acknowledges this, stating that additional classification capability comes with "some degree of larger delays."
  • Future Work: The authors suggest further research into more advanced feature extraction techniques and exploring adaptive thresholding mechanisms. They also imply that the framework could be extended to handle more complex or novel types of disturbances.

Summary updated

BibTeX
@article{mozaffari2022real,
  author = {Mahsa Mozaffari and Keval Doshi and Yasin Yilmaz},
  title = {{Real-Time Detection and Classification of Power Quality Disturbances}},
  journal = {Sensors},
  year = {2022},
  doi = {10.3390/s22207958},
}

Loading related content...

Figure