Abstract
Transient Stability Analysis (TSA) is critical for power grid reliability, yet identifying specific stability "insights" within massive simulation datasets remains a manual, labor-intensive process. This paper presents DUHiV (DenseU-Hierarchical Variational Autoencoder), a framework that combines a novel deep learning architecture with an interactive visual interface. By leveraging hierarchical latent variables and latent space arithmetic, DUHiV enables domain experts to discover and annotate complex grid behaviors in an unsupervised or semi-supervised manner, significantly reducing the reliance on large-scale labeled datasets.
Introduction
Power grid analysts increasingly rely on Power Grid Pixel Maps (PGPMs) to visualize complex simulation data. In these maps, time and bus locations define the axes, while pixel intensity represents critical variables like voltage. Despite their utility, the sheer volume of data—often exceeding hundreds of thousands of maps—makes manual identification of stability insights a major bottleneck. Traditional supervised learning is hindered by the lack of labeled data, while standard unsupervised models often fail to capture the intricate spatial and semantic structures of high-resolution PGPMs.
The DUHiV Architecture
To address these challenges, the DUHiV framework introduces a hybrid architecture designed for high-fidelity reconstruction and interpretable feature learning.
- DenseU Structure: The inference network utilizes Dense blocks to facilitate feature reuse and gradient flow. The generative network employs a symmetric U-Net-like expanding path, ensuring that multi-scale spatial details are preserved during reconstruction.
- Hierarchical Latent Variables: Unlike standard VAEs that use a single flat latent vector, DUHiV implements a hierarchy. The top layer is designed to capture high-level semantic "insights," while the bottom layer focuses on local pixel-level "shapes." This separation is crucial for making the latent space navigable for human experts.
Interactive Insight Discovery
The core innovation of DUHiV lies in its "human-in-the-loop" approach. Analysts interact with the learned representations through a visual interface, performing latent space arithmetic to define and discover insights. For instance, by subtracting a "stable" state vector from an "anomalous" one, experts can isolate a "difference vector" that represents a specific fault type. This vector can then be used as a template to search the entire dataset (e.g., the PGPM-250k set) for similar occurrences.
Experimental Evaluation
The framework was tested on two industrial datasets from Northern China: PGPM-3K (labeled for ground truth) and PGPM-250k (unlabeled).
- Reconstruction and ELBO: DUHiV achieved a significantly tighter Evidence Lower Bound (ELBO) of -2416.5, vastly outperforming baseline models like LVAE (-13624.3) and DLGM (-25520.0). This indicates a much more accurate approximation of the underlying data distribution.
- Annotation Performance: In unsupervised mode, the framework reached a mean Average Precision (mAP) of 64.4%. This performance is comparable to semi-supervised methods requiring 1% labeled data and far exceeds existing unsupervised benchmarks like NPID (25.7%).
Conclusion
DUHiV represents a significant advancement in power grid data analysis. By combining spatial awareness through DenseU blocks with hierarchical latent modeling, it provides a powerful tool for unsupervised insight discovery. While VAE-based reconstructions still exhibit some characteristic blurriness, the framework's ability to facilitate expert-driven discovery through latent arithmetic offers a scalable solution for managing massive, unlabeled industrial datasets.
Summary updated