Abstract
Deep learning based transient stability assessment (TSA) has achieved great success, yet the lack of interpretability hinders its industrial application. Although a great number of studies have tried to explore the interpretability of network solutions, many problems still remain unsolved: (1) the difference between the widely accepted power system knowledge and the generated interpretive rules is large, (2) the probability characteristics of the neural network have not been fully considered during generating the interpretive rules, (3) the cost of the trade-off between accuracy and interpretability is too heavy to take. To address these issues, an interpretable power system Transient Stability Assessment method with Expert guiding Neural-Regression-Tree (TSA-ENRT) is proposed. TSA-ENRT utilizes an expert guiding nonlinear regression tree to approximate the neural network prediction and the neural network can be explained by the interpretive rules generated by the tree model. The nonlinearity of the expert guiding nonlinear regression tree is endowed with the extracted knowledge from a simple two-machine three-bus power system, which forms an expert knowledge base and thus the generated interpretive rules are more consistent with human cognition. Besides, the expert guiding tree model can build a bridge between the interpretive rules and the probability prediction of neural network in a regression way. By regularizing the neural network with the average decision length of ENRT, the association of the neural network and tree model is constructed in the model training level which provides a better trade-off between accuracy and interpretability. Extensive experiments indicate the interpretive rules generated by the proposed TSA-ENRT are highly consistent with the neural network prediction and more agreed with human expert cognition.
Introduction
Transient Stability Assessment (TSA) is critical for determining if a power system can remain synchronized following major disturbances, such as lightning strikes or line trips. While deep learning (DL) models have demonstrated exceptional predictive accuracy in this domain, their "black box" nature limits their adoption in high-stakes grid operations. Engineers require transparent reasoning before initiating emergency controls.
Existing attempts to explain DL models often employ "Neural Trees," where a decision tree mimics the neural network's behavior. However, these methods frequently suffer from a significant cognition gap (linear splits vs. nonlinear physics), information loss (binary classification vs. probability nuances), and a severe accuracy-interpretability trade-off. This paper introduces TSA-ENRT, a framework designed to bridge these gaps by integrating physics-based expert knowledge with a differentiable tree-regularized neural network.
Methodology
The TSA-ENRT framework consists of four integrated modules designed to align neural network performance with human-interpretable logic.
1. Expert Knowledge Base
Standard decision trees typically use linear splits (e.g., $V > 0.9$). However, power system dynamics are governed by nonlinear physics. TSA-ENRT extracts nonlinear terms—such as $\sin\delta_G$, $\cos\delta_G$, and quadratic voltage transforms—derived from the swing equations of a two-machine three-bus system. These terms are added to the tree's feature set, allowing the resulting rules to align with fundamental physical principles.
2. Neural Network Evaluation Model (NNEM)
The framework utilizes Gated Recurrent Units (GRUs) for prediction but recasts the task as a regression of stability probability ($Y_{NN,soft}$) rather than a hard binary classification. This preserves the nuanced "soft" information necessary for high-fidelity explanations.
3. Nonlinear Regression Tree (NRT)
A regression tree is trained to approximate the NN's probability predictions using the expert-guided nonlinear features. By mimicking the continuous output of the NN, the tree can explain why a sample is "barely stable" versus "highly stable," leading to more granular interpretive paths.
4. Tree Regularization via Surrogate Model
To mitigate the accuracy-interpretability trade-off, the NN is trained with a penalty for complexity. Since tree depth is non-differentiable, a Multilayer Perceptron (MLP) acts as a surrogate model to estimate tree depth, providing a differentiable path for backpropagation. This forces the NN to develop decision boundaries that are naturally easier for a shallow tree to approximate.
Experimental Setup
The method was evaluated using the CEPRI-TAS system, a 15-machine, 85-bus regional power system model. Researchers generated 2,000 samples via Monte-Carlo simulations of three-phase faults, utilizing a 3:1 training-to-testing split to validate the model's performance and interpretability.
Key Results
TSA-ENRT was compared against standard regularized GRUs and previous state-of-the-art interpretable models (GRU-TR):
- Accuracy: TSA-ENRT maintained an accuracy of 94.3%, significantly outperforming previous tree-regularized methods that often sacrifice accuracy for interpretability.
- Fidelity: The model achieved a fidelity score of 0.971, indicating that the tree's explanations match the neural network's predictions with high precision.
- Interpretability: The generated rules explicitly utilize physical terms like $\sin(\delta_G)$, which allows engineers to validate the model's logic against established physical intuition.
Conclusion
TSA-ENRT successfully alleviates the trade-off between accuracy and interpretability in power system TSA. By incorporating physics-informed nonlinearity and a differentiable regularization scheme, the model provides high-performance predictions alongside rules that are consistent with human expert cognition. Future work will explore expanding the expert knowledge base for more complex dynamics and testing the framework on larger, more diverse grid topologies.
Summary updated