Scholarlink

<p>Real-time anomaly detection in power grids is essential to maintain system reliability and prevent cascading failures. This paper introduces the deep-learning-enhanced solid-state sensor arrays with gallium nitride high-electron-mobility transistors (GaN HEMTs) and silicon carbide (SiC) Schottky diodes to ensure extensive grid coverage during anomaly detection. Metal–Organic Chemical Vapor Deposition-fabricated GaN HEMTs on semi-insulating SiC substrates exhibit a two-dimensional electron gas mobility of 1850&#xa0;cm<sup>2</sup>/V·s, enabling high-bandwidth voltage sensitivity of up to 100&#xa0;MHz. The 4H-SiC Schottky diodes enable current measurement at a 1.2&#xa0;kV breakdown voltage. A neural processor, Spatio-Temporal Graph Transformer with Adaptive Dual-stream, is used for on-chip anomaly detection. It achieves 99.2% classification accuracy across nine fault types: line-to-ground faults, line-to-line faults, three-phase faults, generator trips, load shedding, capacitor switching, voltage sags/swells, frequency deviations, and cyberattacks. Material characterization proves its excellent crystalline quality with an X-ray Diffraction, FWHM (Full Width at Half Maximum) of 285 arcsec and surface roughness of 0.32&#xa0;nm. The system’s inference latency of 14.8&#xa0;ms fulfils the protection relay requirements. It has an overall end-to-end response time of 28.3&#xa0;ms. Testing over a temperature range of − 40&#xa0;°C to + 125&#xa0;°C demonstrates an accuracy degradation of less than 0.5%. The module's low form factor (50 × 50 × 15&#xa0;mm<sup>3</sup>) and power (450&#xa0;mW) demonstrate its potential for deployment in a space-limited substation without external computing resources. This research shows that wide-bandgap semiconductors and transformer deep learning allow for real-time monitoring of smart grids.</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

Deep learning-enhanced solid-state sensor arrays for real-time power grid anomaly detection and classification

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Zhixin Li

摘要

Real-time anomaly detection in power grids is essential to maintain system reliability and prevent cascading failures. This paper introduces the deep-learning-enhanced solid-state sensor arrays with gallium nitride high-electron-mobility transistors (GaN HEMTs) and silicon carbide (SiC) Schottky diodes to ensure extensive grid coverage during anomaly detection. Metal–Organic Chemical Vapor Deposition-fabricated GaN HEMTs on semi-insulating SiC substrates exhibit a two-dimensional electron gas mobility of 1850 cm2/V·s, enabling high-bandwidth voltage sensitivity of up to 100 MHz. The 4H-SiC Schottky diodes enable current measurement at a 1.2 kV breakdown voltage. A neural processor, Spatio-Temporal Graph Transformer with Adaptive Dual-stream, is used for on-chip anomaly detection. It achieves 99.2% classification accuracy across nine fault types: line-to-ground faults, line-to-line faults, three-phase faults, generator trips, load shedding, capacitor switching, voltage sags/swells, frequency deviations, and cyberattacks. Material characterization proves its excellent crystalline quality with an X-ray Diffraction, FWHM (Full Width at Half Maximum) of 285 arcsec and surface roughness of 0.32 nm. The system’s inference latency of 14.8 ms fulfils the protection relay requirements. It has an overall end-to-end response time of 28.3 ms. Testing over a temperature range of − 40 °C to + 125 °C demonstrates an accuracy degradation of less than 0.5%. The module's low form factor (50 × 50 × 15 mm3) and power (450 mW) demonstrate its potential for deployment in a space-limited substation without external computing resources. This research shows that wide-bandgap semiconductors and transformer deep learning allow for real-time monitoring of smart grids.