Scholarlink

<p>In this work, we developed a new Physics-Informed Long Short-Term Memory (PI-LSTM)–based method for online voltage stability assessment (VSA) in power systems with significant penetration of renewable energy sources (RESs). Unlike conventional hybrid model- and data-driven approaches, our method integrates physical laws directly into the neural network structure, enabling smooth physics-data integration. By embedding the Thevenin equivalent model and voltage-current dynamics into the network’s loss function, the learning process is guided by physical constraints. This integration improves the model’s generalization to unseen faults and new RES configurations, simplifies the VSA pipeline, and reduces computational overhead. Thanks to its reliance on both basic system equations and measured data patterns, the PI-LSTM framework demonstrates robustness against noise and missing phasor measurement unit (PMU) data. Extensive case studies on the IEEE 39-bus system verify that the proposed method significantly outperforms conventional Bidirectional Long Short-Term Memory (BiLSTM)-based and model-driven approaches in terms of accuracy, robustness, and interpretability. Moreover, the approach accelerated and guaranteed online stability decisions by coordinating with the dynamics of RESs. With complicated, dynamic behaviors brought by renewable integration, the proposed PI-LSTM approach presents a viable path for improving the resilience of next 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

Physics-Informed LSTM for Real-Time Voltage Stability Assessment of Power Systems with High Renewable Energy Integration

  • 期刊论文

  • 发表日期：

    2025年11月27日

查看全文
  • Jie Yao,
  • Huashi Zhao,
  • Yaping Hu,
  • Kaiting Lai,
  • Wen Zhu

摘要

In this work, we developed a new Physics-Informed Long Short-Term Memory (PI-LSTM)–based method for online voltage stability assessment (VSA) in power systems with significant penetration of renewable energy sources (RESs). Unlike conventional hybrid model- and data-driven approaches, our method integrates physical laws directly into the neural network structure, enabling smooth physics-data integration. By embedding the Thevenin equivalent model and voltage-current dynamics into the network’s loss function, the learning process is guided by physical constraints. This integration improves the model’s generalization to unseen faults and new RES configurations, simplifies the VSA pipeline, and reduces computational overhead. Thanks to its reliance on both basic system equations and measured data patterns, the PI-LSTM framework demonstrates robustness against noise and missing phasor measurement unit (PMU) data. Extensive case studies on the IEEE 39-bus system verify that the proposed method significantly outperforms conventional Bidirectional Long Short-Term Memory (BiLSTM)-based and model-driven approaches in terms of accuracy, robustness, and interpretability. Moreover, the approach accelerated and guaranteed online stability decisions by coordinating with the dynamics of RESs. With complicated, dynamic behaviors brought by renewable integration, the proposed PI-LSTM approach presents a viable path for improving the resilience of next power systems.