Scholarlink

<p>The growing use of renewable energy sources, distributed energy assets, and smart grid technologies has immensely contributed to the complexity and susceptibility of the modern power system and has resulted in a range of anomalies, including equipment failures, cyberattacks, and nature-induced disruptions. The traditional threshold-based and single-model anomaly detection methods fail to represent the temporal dependencies, sparsity, and non-stationary behavior of grid data. In order to tackle these issues, this paper will suggest a role-conscious hybrid anomaly detection architecture that combines the use of long short-term memory (LSTM) networks, Isolation Forests, and Autoencoders in a correlative fashion. The LSTM block identifies long-term time behavior, the Isolation Forest detects sparse and infrequent deviations, and the Autoencoder measures inconsistencies in the reconstruction, which makes it possible to detect anomalies at a multistage level. The structure is tested on a 50,000 sample IEEE DataPort dataset enhanced with realistically injured artificial anomalies of various fault types, fault severity levels, and temporal distributions. Experimental findings show that the suggested methodology results in 95.8 percent accuracy, 94.6 percent precision, 96.2 percent recall, and a 95.4 percent F1-score, which is better than traditional threshold-based detection (78.5 percent) and standalone Isolation Forest models (85.3 percent). The model can identify 98 percentage of cyberattack cases and 93 percentage of faults involving equipment with a low false alarm rate when everything is functioning as expected. By analysis, it has been shown that computational training occurs offline, and lightweight inference can be used in a near-real-time monitoring scenario. Furthermore, the trends of anomaly scores and feature-level reconstruction errors are also offered to give initial interpretability to assist operations-level decision-making. The suggested framework is an excellent and realistic way to improve the reliability of the power system in intelligent grid settings through scaling.</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

Enhancing Power System Reliability with a Hybrid LSTM-Isolation Forest-Autoencoder Model

  • 期刊论文

  • 发表日期：

    2026年3月27日

查看全文
  • Manjot Kaur Sidhu,
  • Cindhe Ramesh,
  • Geetika Sharma,
  • Anupam Mittal,
  • Neera Batra

摘要

The growing use of renewable energy sources, distributed energy assets, and smart grid technologies has immensely contributed to the complexity and susceptibility of the modern power system and has resulted in a range of anomalies, including equipment failures, cyberattacks, and nature-induced disruptions. The traditional threshold-based and single-model anomaly detection methods fail to represent the temporal dependencies, sparsity, and non-stationary behavior of grid data. In order to tackle these issues, this paper will suggest a role-conscious hybrid anomaly detection architecture that combines the use of long short-term memory (LSTM) networks, Isolation Forests, and Autoencoders in a correlative fashion. The LSTM block identifies long-term time behavior, the Isolation Forest detects sparse and infrequent deviations, and the Autoencoder measures inconsistencies in the reconstruction, which makes it possible to detect anomalies at a multistage level. The structure is tested on a 50,000 sample IEEE DataPort dataset enhanced with realistically injured artificial anomalies of various fault types, fault severity levels, and temporal distributions. Experimental findings show that the suggested methodology results in 95.8 percent accuracy, 94.6 percent precision, 96.2 percent recall, and a 95.4 percent F1-score, which is better than traditional threshold-based detection (78.5 percent) and standalone Isolation Forest models (85.3 percent). The model can identify 98 percentage of cyberattack cases and 93 percentage of faults involving equipment with a low false alarm rate when everything is functioning as expected. By analysis, it has been shown that computational training occurs offline, and lightweight inference can be used in a near-real-time monitoring scenario. Furthermore, the trends of anomaly scores and feature-level reconstruction errors are also offered to give initial interpretability to assist operations-level decision-making. The suggested framework is an excellent and realistic way to improve the reliability of the power system in intelligent grid settings through scaling.