TL;DR
This paper introduces a novel framework for precisely identifying power quality disturbances (PQDs) by transforming 1D electrical signals into 2D images using the Relative Position Matrix (RPM) and then classifying these images with an enhanced deep learning model, SGE-MobileViT. This approach achieves state-of-the-art accuracy, even in noisy conditions, offering a robust solution for managing power quality in grids with increasing renewable energy integration.
What Problem Does This Paper Address?
The increasing integration of renewable energy sources like wind and solar power into electrical grids, while beneficial for sustainability, introduces significant challenges to power quality. The intermittent nature of these sources can lead to various power quality disturbances (PQDs), such as voltage sags, swells, harmonics, and flickers. These disturbances can cause equipment malfunction, economic losses, and grid instability. Effectively managing these issues requires rapid and accurate detection and classification of PQDs. Existing methods often struggle with complex disturbances, noise, and the need for manual feature engineering. This paper aims to address this by developing a more precise and robust PQD identification system.
Methods
The proposed framework consists of two main components:
- Two-Dimensional Characterization Feature Enhancement (RPM): Instead of directly analyzing the raw 1D time-series PQD signals, the authors employ the Relative Position Matrix (RPM) technique. RPM transforms the 1D signal into a 2D image. This process involves:
- Z-score normalization: Standardizing the signal.
- Piecewise Aggregate Approximation (PAA): Reducing dimensionality while preserving trends.
- Matrix Construction: Calculating the relative position differences between time points to form a matrix.
-
Normalization: Converting the matrix into a grayscale image. This 2D representation is designed to better capture temporal dependencies and inter-point relationships, enhancing feature extraction.
-
Deep Learning Classification (SGE-MobileViT): The generated 2D images are then fed into a novel deep learning model called Spatial Group-wise Enhance (SGE)-MobileViT. This model is an adaptation of the MobileViT architecture, which combines convolutional neural networks (CNNs) with Vision Transformers (ViTs). The key enhancement is the integration of the SGE attention mechanism. SGE adaptively adjusts the importance of different spatial locations within feature maps, allowing the model to:
- Focus on critical semantic features.
- Suppress noise.
- Improve the handling of both local and global contextual information. This attention mechanism is embedded within the convolutional and transformer modules of MobileViT, aiming to boost its feature expression capabilities and robustness.
Datasets
The paper utilizes a simulated PQD dataset generated using MATLAB.
- Signal Generation: A total of 14 PQD types were simulated, including 8 single disturbance types and 6 composite types (4 double-composite and 2 triple-composite).
- Dataset Size: 1000 randomly selected samples were generated for each disturbance type, resulting in a total of 14,000 samples.
- Training/Testing Split: The dataset was divided into training (70%) and testing (30%) sets, with 9800 samples for training and 4200 for testing.
- Hardware Platform Data: Additionally, data was collected from a physical experimental platform, generating 100 sets of PQD data for each of the 14 types, totaling 1400 samples for validation.
Key Insights and Ideas
- Transforming Time-Series to Images for PQD: The core idea of converting 1D electrical signals into 2D images using RPM is a powerful way to leverage the strengths of image-based deep learning models for signal analysis. This approach effectively visualizes complex temporal relationships.
- SGE Attention for Enhanced Feature Representation: The SGE attention mechanism is crucial for making the MobileViT model more effective. By adaptively weighting spatial features, it allows the model to better discern subtle but important characteristics of PQDs, especially in the presence of noise.
- Robustness to Noise: The RPM + SGE-MobileViT framework demonstrates remarkable resilience to noise, maintaining high accuracy even at low signal-to-noise ratios (SNRs). This is a critical advantage for real-world power grid applications where noise is prevalent.
- Comprehensive Validation: The study validates its approach through both extensive simulations and a physical hardware platform, including an embedded system demonstration, lending strong credibility to its practical applicability.
Key Results
- High Accuracy: The proposed SGE-MobileViT model achieved a peak classification accuracy of 99.17% in a noiseless environment.
- Noise Performance: Even with noise, the model maintained high accuracy: 95.13% at 20 dB SNR, 97.00% at 30 dB SNR, and 97.50% at 50 dB SNR.
- Superiority over Baselines: Compared to other methods like GTI, RP, and MTF for image generation, RPM consistently yielded better accuracy. When compared to other deep learning models (CNN, VggNet16, DenseNet, ShuffleNetV2, MobileViT, Swin Transformer), SGE-MobileViT consistently outperformed them across various noise conditions.
- Embedded System Viability: The model demonstrated strong performance on an embedded platform (Raspberry Pi 4B), achieving a total recognition rate of 97.50% for various PQD types.
Limitations and Future Work
- Limitations:
- Performance Degradation in Extreme Noise: The authors acknowledge that the model's performance may degrade in extremely noisy environments or with complex multi-source disturbances.
- Computational Complexity: The computational complexity of SGE-MobileViT might be a concern for large-scale power grids requiring real-time processing.
- Future Work:
- Efficient Feature Extraction: Explore more efficient techniques for PQD signal processing.
- Lightweight Architectures & Model Compression: Optimize computational efficiency for real-time applications using lightweight networks or compression techniques.
- Multimodal Data Fusion: Integrate electrical signals with environmental or operational data to improve overall system accuracy.
- Quantitative Analysis Integration: Incorporate methods like THD, PFC, and load voltage deviation for enhanced accuracy and reliability in PQD management.
Summary updated