Scholarlink

<p>As the multi-faceted interactive characteristics among sources, grids, loads, and storage in new-type power systems become increasingly prominent, the time-series stability assessment (TSA) of power grids faces the triple challenges of balancing high precision, high timeliness, and strong interpretability. Although existing deep neural networks demonstrate excellent assessment accuracy, their “black-box” nature restricts their engineering applications in critical decision-making fields. To address this issue, this study proposes an interpretable assessment framework that integrates an improved multivariate decision tree (IMDT) with the Informer model. The core strategy is a “divide-and-conquer” approach: leveraging the Informer’s powerful time-series perception capabilities as a feature extractor rather than an end-to-end decision-maker, and then submitting these highly condensed dynamic features to a cost-sensitive IMDT—inherently a “white-box”—for logically clear discrimination. The framework first employs a pre-trained Informer model to efficiently encode long-term dependency features in the grid’s dynamic response, generating highly condensed transient feature vectors. Subsequently, these feature vectors, along with the system’s static features after attribute reduction, are used as inputs for an multivariate decision tree (MDT) model optimized with a Boosting-like approach and cost-sensitive learning to perform the final stability state discrimination. Validation results on the IEEE 39-bus standard system demonstrate that, compared to traditional MDT, long short-term memory (LSTM), and standalone Informer models, this proposed model achieves highly competitive performance while providing full transparency in decision paths. Under a 1.1&#xa0;s decision window, the model achieves a 99.1% recognition rate for unstable samples (with a miss rate of &lt; 1%) and a 98.8% recognition rate for stable samples (with a false alarm rate (FAR) of 1.2%), resulting in an overall accuracy of 99.0%. More importantly, the number of decision rules is reduced by 25% compared to traditional MDT, and the decision boundaries are clearer, fully proving that this hybrid model is an advanced technical solution that combines accuracy, interpretability, and engineering applicability for addressing TSA challenges in new-type power systems.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Interpretable evaluation model for power grid temporal stability integrating informer and cost sensitive decision tree

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月22日

查看全文
  • Zheng Li,
  • Yongyong Jia,
  • Xinyao Zhu,
  • Xiaobo Wang

摘要

As the multi-faceted interactive characteristics among sources, grids, loads, and storage in new-type power systems become increasingly prominent, the time-series stability assessment (TSA) of power grids faces the triple challenges of balancing high precision, high timeliness, and strong interpretability. Although existing deep neural networks demonstrate excellent assessment accuracy, their “black-box” nature restricts their engineering applications in critical decision-making fields. To address this issue, this study proposes an interpretable assessment framework that integrates an improved multivariate decision tree (IMDT) with the Informer model. The core strategy is a “divide-and-conquer” approach: leveraging the Informer’s powerful time-series perception capabilities as a feature extractor rather than an end-to-end decision-maker, and then submitting these highly condensed dynamic features to a cost-sensitive IMDT—inherently a “white-box”—for logically clear discrimination. The framework first employs a pre-trained Informer model to efficiently encode long-term dependency features in the grid’s dynamic response, generating highly condensed transient feature vectors. Subsequently, these feature vectors, along with the system’s static features after attribute reduction, are used as inputs for an multivariate decision tree (MDT) model optimized with a Boosting-like approach and cost-sensitive learning to perform the final stability state discrimination. Validation results on the IEEE 39-bus standard system demonstrate that, compared to traditional MDT, long short-term memory (LSTM), and standalone Informer models, this proposed model achieves highly competitive performance while providing full transparency in decision paths. Under a 1.1 s decision window, the model achieves a 99.1% recognition rate for unstable samples (with a miss rate of < 1%) and a 98.8% recognition rate for stable samples (with a false alarm rate (FAR) of 1.2%), resulting in an overall accuracy of 99.0%. More importantly, the number of decision rules is reduced by 25% compared to traditional MDT, and the decision boundaries are clearer, fully proving that this hybrid model is an advanced technical solution that combines accuracy, interpretability, and engineering applicability for addressing TSA challenges in new-type power systems.