.png){kind=icon}
.png){kind=icon}
Return
Power system fault diagnosis and cascading failure early warning method based on rule knowledge embedding
.png){kind=icon}
PRE
AI
S
Z
Z
Z
DOI：10.1016/j.epsr.2026.113016.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Power system fault diagnosis and cascading failure early warning face unprecedented challenges as grid complexity increases. Traditional methods are confronted by critical limitations, including the inadequate handling of incomplete measurement data, an inability to predict complex cascading failure propagation paths, and insufficient identification of near-overload states that often precede catastrophic failures. Most critically, these conventional approaches rely on preset static fault sets, which cannot address the exponentially growing combinations of possible fault scenarios in large-scale power grids. This renders them ineffective for occasional minor faults or unknown failure patterns emerging from new equipment and operational conditions. To address these challenges, this paper proposes an intelligent fault diagnosis and cascading failure early warning method based on rule knowledge embedding, which aims to advance the early warning paradigm. The method explicitly embeds power system domain knowledge as differentiable rules into a deep learning model. This model achieves multi-scale spatial feature extraction through parallel convolutional neural networks and captures temporal dependencies via a dual-layer bidirectional long short-term memory network. The rule embedding mechanism enables an active search for cascading development paths without requiring manual fault enumeration, while a multi-strategy missing data processing mechanism enhances adaptability to realistic incomplete measurement conditions. Experimental validation on the IEEE 118-bus system demonstrates that the proposed EnhancedPCDB model achieves 98.00% fault diagnosis accuracy with a 97.41% F1-score, representing a 5.75 percentage points improvement over the basic PCDB model. For cascading detection, the model attains 96.39% accuracy with 97.70% precision and 96.40% recall, while near-overload warning reaches 98.67% accuracy with a critical 99.81% recall. The model maintains robust performance even under 13.45% data missing conditions. Compared to traditional preset fault set methods, the proposed approach eliminates exponential fault enumeration requirements while achieving superior cascading failure prediction coverage and accuracy, providing a trans-formative solution for real-time power system early warning.
Keywords:
Cascading failure warning
Fault diagnosis
Knowledge embedding
Near-overload warning
Risk analysis
Journal
IF:
4.2
Papers:
1.1W
Citations:
2.2W
Organization
Cited Papers