Source-linked AI summary
A Deep-Learning Intelligent System Incorporating Data Augmentation for Short-Term Voltage Stability Assessment of Power Systems
Yang Li, Meng Zhang, Chen Chen
TL;DR
The paper addresses how to make deep-learning STVSA work with small training datasets when labeled data are costly and difficult to obtain. It combines semi-supervised labeling, LSGAN-based augmentation, and BiGRU-Attention modeling, with tests reporting better accuracy, faster response, and robustness to noise. Future work targets missing PMU information, class imbalance from rare instability events, and automated model optimization.
Problem
Deep-learning STVSA models require data that are costly and difficult to collect and accurately label, creating a challenge for small training datasets.
Method
The proposed system uses semi-supervised cluster learning, conditional LSGAN data augmentation, and BiGRU-Attention to assess post-disturbance voltage stability.
Results
The approach outperforms other state-of-the-art alternatives with better accuracy and faster response time, while statistical indicators further validate its performance.
Takeaways & Limitations
LSGAN-based augmentation significantly improves training and testing accuracies for the studied small-scale dataset and remains robust under noisy environments.
Takeaways & Limitations
Future work must address missing PMU information and class imbalance caused by rare instability events.
Abstract
from arXiv · showhide
Facing the difficulty of expensive and trivial data collection and annotation, how to make a deep learning-based short-term voltage stability assessment (STVSA) model work well on a small training dataset is a challenging and urgent problem. Although a big enough dataset can be directly generated by contingency simulation, this data generation process is usually cumbersome and inefficient; while data augmentation provides a low-cost and efficient way to artificially inflate the representative and diversified training datasets with label preserving transformations. In this respect, this paper proposes a novel deep-learning intelligent system incorporating data augmentation for STVSA of power systems. First, due to the unavailability of reliable quantitative criteria to judge the stability status for a specific power system, semi-supervised cluster learning is leveraged to obtain labeled samples in an original small dataset. Second, to make deep learning applicable to the small dataset, conditional least squares generative adversarial networks (LSGAN)-based data augmentation is introduced to expand the original dataset via artificially creating additional valid samples. Third, to extract temporal dependencies from the post-disturbance dynamic trajectories of a system, a bi-directional gated recurrent unit with attention mechanism based assessment model is established, which bi-directionally learns the significant time dependencies and automatically allocates attention weights. The test results demonstrate the presented approach manages to achieve better accuracy and a faster response time with original small datasets. Besides classification accuracy, this work employs statistical measures to comprehensively examine the performance of the proposal.
NOMENCLATURE
STVSA is framed as an urgent binary-classification problem requiring accurate and timely detection of post-disturbance voltage stability. The paper addresses limited labeled data by combining data augmentation with sequence modeling and attention.
- STVSA evaluates whether power-system voltage can quickly return to an acceptable level after a large disturbance.
- Existing STVSA approaches include physical-model methods, hybrid monitoring, and shallow machine-learning techniques.
- Deep learning is attractive for sequential post-disturbance data, while GRU offers fewer parameters and higher training efficiency than LSTM without sacrificing accuracy.
- Reliable large-scale labeled datasets are difficult and expensive to obtain because no unified criterion consistently identifies stability status.
- Data augmentation is presented as a low-cost, efficient alternative to cumbersome contingency simulation for expanding representative training data.
- The related-work comparison organizes whether prior methods adopt the proposed approach’s corresponding techniques.
III. DLBAN Framework
The DLBAN framework combines conditional LSGAN-based data augmentation with a BiGRU-Attention assessment model to support STVSA from small datasets. LSGAN expands labeled training data, while bidirectional recurrence and attention extract influential temporal information for classification.
- DLBAN combines LSGAN and BiGRU-Attention for short-term voltage stability assessment.
- Least Squares Generative Adversarial Network: LSGAN uses least-squares discriminator loss to penalize samples far from the decision boundary and address vanishing gradients.
- Least Squares Generative Adversarial Network: Conditional LSGAN receives one-hot labels with noise for generation and appends labels to real or generated data for discrimination.
- BiGRU-Attention: BiGRU captures temporal dependencies in both forward and backward directions by concatenating two GRU streams.
- DLBAN Framework: The framework first generates large, balanced labeled data with LSGAN, then trains BiGRU-Attention on the expanded dataset.
- BiGRU-Attention: The attention layer assigns higher weights to strongly correlated hidden features and extracts the most significant temporal information.
- Performance Evaluation: LSGAN augmentation improves WD by 21.9%, MMD by 3.23%, and FID by 8.75% compared with CGAN.
- Performance Evaluation: BiGRU-Attention achieves the highest accuracy among compared approaches and exceeds GRU and LSTM in MCC by 4.83% and 6.49%, respectively.
OTW (ms)
The proposed STVSA system balances rapid assessment with accuracy by selecting a 30 ms observation window, while maintaining strong performance under noisy PMU measurements. Its data augmentation and BiGRU-attention components support small-dataset learning and temporal feature extraction.
- OTW (ms): 30 ms is selected as the observation window because longer windows provide only marginal accuracy gains while reducing rapidness.The BiGRU-Attention model reportedly outperforms existing methods with shorter response time and no accuracy sacrifice.
- OTW (ms): 0.08 second is the maximum response time after accounting for typical PMU-to-PDC communication delays.The communication delay is reported as 0.02 to 0.05 second.
- OTW (ms): 99.32% accuracy is retained at 30 dB SNR, with MCC 0.9865, F1-score 0.9939, and AUC 0.9905.Accuracy decreases only slightly as noise increases from 50 dB to 30 dB.
- OTW (ms): The DLBAN framework enables deep-learning assessment with small training datasets through data augmentation.The framework is presented as a tool for handling limited data availability.
- OTW (ms): The BiGRU-attention model captures temporal dependence in both directions and improves feature extraction by emphasizing important information.Attention assigns greater weight to system-dynamics information from sampling times that contribute more strongly to STVSA.