Scholarlink

<p>Single-phase, non-linear traction loads, regenerative braking, and low-frequency harmonics with frequent inter-harmonic side lobes diminish power quality (PQ) in electrified railway networks. Conventional proportional–integral controllers utilizing instantaneous reactive-power theory (PI-IRP) and fuzzy logic controllers (FLC) for Unified Power Quality Conditioners (UPQCs) exhibit sluggishness, lack adaptability to swiftly fluctuating traction loads, and fail to guarantee stability. Nonetheless, deep learning applications for railway UPQC control remain inadequately investigated. A deep learning-based control framework is proposed, utilizing two feedforward artificial neural networks (FF-ANNs): a 3–15–10–2 network for the shunt active filter and a 3–12–8–2 network for the series active filter, along with a quadratic Lyapunov stability constraint to ensure bounded closed-loop behaviour. The Levenberg–Marquardt algorithm trains the networks offline using 48,000 samples encompassing normal operation, voltage sags and swells, harmonic distortion, load transients, and single-phase-to-ground failures. An adaptation rule restricted by Lyapunov optimizes them in real-time. A framework for a 2 × 25&#xa0;kV, 50&#xa0;Hz autotransformer-fed traction power supply was validated by numerical simulations in steady-state, dynamic, and fault conditions. The suggested controller reduces source-current total harmonic distortion (THD) from 21.74% to 1.46% (a 93.3% decrease), voltage imbalance factor (VUF) from 3.65% to 0.82% (a 77.5% decrease), and achieves a displacement power factor of 0.99. The dynamic response during a 30% voltage sag is 61.1% superior to PI-IRP. The statistical significance of all enhancements is assessed using one-way ANOVA with Tukey’s honest significant difference (HSD) post-hoc test at a 95% confidence level, alongside Shapiro–Wilk normality testing over 30 distinct trials. The design of next-generation railway power management is scalable, addressing current and future requirements for renewable energy integration.</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-based control framework using dual feedforward artificial neural networks with Lyapunov stability optimization for unified power quality conditioners in electrified railway systems

  • 期刊论文

  • 发表日期：

    2026年7月9日

查看全文
  • Dinesh Kumar Nishad,
  • A. N. Tiwari,
  • Saifullah Khalid

摘要

Single-phase, non-linear traction loads, regenerative braking, and low-frequency harmonics with frequent inter-harmonic side lobes diminish power quality (PQ) in electrified railway networks. Conventional proportional–integral controllers utilizing instantaneous reactive-power theory (PI-IRP) and fuzzy logic controllers (FLC) for Unified Power Quality Conditioners (UPQCs) exhibit sluggishness, lack adaptability to swiftly fluctuating traction loads, and fail to guarantee stability. Nonetheless, deep learning applications for railway UPQC control remain inadequately investigated. A deep learning-based control framework is proposed, utilizing two feedforward artificial neural networks (FF-ANNs): a 3–15–10–2 network for the shunt active filter and a 3–12–8–2 network for the series active filter, along with a quadratic Lyapunov stability constraint to ensure bounded closed-loop behaviour. The Levenberg–Marquardt algorithm trains the networks offline using 48,000 samples encompassing normal operation, voltage sags and swells, harmonic distortion, load transients, and single-phase-to-ground failures. An adaptation rule restricted by Lyapunov optimizes them in real-time. A framework for a 2 × 25 kV, 50 Hz autotransformer-fed traction power supply was validated by numerical simulations in steady-state, dynamic, and fault conditions. The suggested controller reduces source-current total harmonic distortion (THD) from 21.74% to 1.46% (a 93.3% decrease), voltage imbalance factor (VUF) from 3.65% to 0.82% (a 77.5% decrease), and achieves a displacement power factor of 0.99. The dynamic response during a 30% voltage sag is 61.1% superior to PI-IRP. The statistical significance of all enhancements is assessed using one-way ANOVA with Tukey’s honest significant difference (HSD) post-hoc test at a 95% confidence level, alongside Shapiro–Wilk normality testing over 30 distinct trials. The design of next-generation railway power management is scalable, addressing current and future requirements for renewable energy integration.