Adaptive Identification of Power System Operational States Based on Spatio-Temporal Dynamic Graph Neural Networks

The state identification of the operational status of power systems plays a crucial role in ensuring the safety, efficiency, and reliable operation of power systems, forming the foundation for the management and control of power system operations. In the practical operation of power systems, the duration of fault states is short, while the system remains in a safe and stable state for a significantly longer period, leading to an imbalance in the training samples available. This imbalance can cause traditional machine learning classification algorithms to favor the majority categories of power system operational states, resulting in inaccurate evaluation metrics and reduced model generalization capability. To address this issue, this paper introduces a spatiotemporal dynamic graph neural network model for the identification of power system operational states. By incorporating class cost coefficients, focal loss, and cross-entropy in a joint loss function, the model adaptively focuses on the less represented operational state categories during training, improving classification accuracy on imbalanced and challenging samples. Moreover, the integration of temporal information and spatial topology information within the graph neural network enhances the model’s robustness and generalization ability in state identification. Experimental results on the standard IEEE 68-bus system demonstrate that, compared to traditional machine learning methods, the proposed joint loss with the spatiotemporal dynamic graph neural network model achieves higher accuracy in the identification of power system operational states, significantly improving classification performance on imbalanced samples with a 4.8% increase in F1-score, reaching a peak F1-score of 98.6%.