两阶段解耦与多教师蒸馏的电力系统主导失稳模式识别

Identification of dominant instability mode in power systems via two-stage decoupling and multi-teacher distillation

关键词:  两阶段训练策略  多教师模型  特征蒸馏  主导失稳模式  电力系统仿真分析  样本不平衡

Key words:  two-stage training strategy  multi-teacher model  feature distillation  dominant instability mode  power system simulation analysis  sample imbalance

摘要:

电力系统的暂态稳定性对保障电网安全运行和持续供电至关重要,而准确识别电力系统暂态主导失稳模式(DIM)是制定有效应急控制策略的关键。文中针对电力系统暂态特征数据分布不平衡问题,提出了一种两阶段解耦学习与多教师蒸馏框架(TSDM)。该框架采用两阶段解耦训练策略,实现了表征学习和分类器训练的协同优化。首先,采用实例采样训练多个教师模型,以学习电力系统暂态特征数据的全局特征分布。其次,采用类平衡采样训练学生模型,通过特征蒸馏从教师模型中迁移高阶特征表示,而非直接复用其分类器权重,从而缓解偏差传递问题。同时对特征向量和分类器权重分别进行归一化处理,有效消除了因特征尺度差异而导致的预测偏差。最后,以可分离Transformer模块作为骨干网络,该模块通过参数共享机制和注意力优化设计,能够准确捕捉长时间序列的时空关联特征,使特征提取性能不受序列长度影响。基于CEPRI-36节点系统算例的仿真结果表明,所提方法在电力系统暂态主导失稳模式识别任务中达到98.61%的分类准确率,尤其在少数类样本识别上展现出显著优势,为电力系统暂态稳定分析提供了有效的解决方案。

Abstract:

The transient stability of power systems is crucial for ensuring secure grid operation and continuous power supply, and accurately identifying the transient dominant instability mode (DIM) of power systems is key to formulating effective emergency control strategies. To address the problem of imbalanced distribution of power system transient characteristic data, this paper proposed a two-stage decoupling learning and multi-teacher distillation (TSDM) framework. This framework employed a two-stage decoupling training strategy to achieve the collaborative optimization of representation learning and classifier training. First, instance sampling was used to train multiple teacher models to learn the global feature distribution of the power system transient characteristic data. Second, class-balanced sampling was adopted to train the student model, which transferred high-order feature representations from the teacher models through feature distillation rather than directly reusing their classifier weights, thereby mitigating the problem of bias propagation. Simultaneously, normalization was applied to the feature vectors and classifier weights, respectively, effectively eliminating the prediction biases caused by differences in feature scales. Finally, a separable Transformer module served as the backbone network; through a parameter sharing mechanism and attention optimization design, this module could accurately capture the spatiotemporal correlation features of long time sequences, ensuring that the feature extraction performance was not affected by sequence length. Simulation results based on the CEPRI-36 node system case show that the proposed method achieves a classification accuracy of 98.61% in the recognition of DIM of power systems, particularly demonstrating a significant advantage in the recognition rate of minority class samples, and it provides an effective solution for power system transient stability analysis.