Scholarlink

<p>An adequate model of predicting the thermal behaviour of underground power cable systems is needed to prevent insulation failure and maximize current-carrying capacity (ampacity). The existing approach is spatial, whether it is through computationally intensive physics-based models that cannot be operated in real time to track data, or data-based surrogates that take the soil to be a stationary medium that neglects the influence of moisture-sensitive conductivity, cyclic thermal charging, and convection processes. A hierarchical hybrid neuro-fuzzy model with three integrated steps, (A) PSO-optimized ANFIS soil thermal-state estimator in the format of predicting effective soil thermal conductivity as a factor of geotechnical properties, is presented in this paper; (B) GA-optimized ANFIS cable transient thermal predictor, which uses the output of Stage A predictors to predict the temperature of the conductor, sheath, and thermal charging rate, and regularized by a physics-informed loss criterion based on the transient heat conduction equation; and three data sets are used: a 257 case database of soil thermal conductivity to be used in the Stage A calibration, 7560 FEM-simulated cable conditions under the dry, moist, and wet soil regime to be used in the Stage B training, and published experimental benchmarks to be used to independently validate those experiments. The mean RMSE of the proposed framework is 0.92 ± 0.04&#xa0;°C and a mean <i>R</i><sup>2</sup> of 0.983 ± 0.002 on heldout test data which is better than plain ANFIS (RMSE 1.85 ± 0.08&#xa0;°C), ANN-BP (RMSE 1.45 ± 0.06&#xa0;°C) and IEC 60287 solutions (RMSE 4.31&#xa0;°C). Ablation analysis reveals ability to reduce RMSE by 40.0% with optimization by PSO-GA, physics-informed loss by 14.6%, and the hierarchical cable coupling of soils by 5.3%. The inference latency of an average workstation (9.4&#xa0;ms) means that it can be used with real-time dynamic thermal rating systems. It is, as far as we know, the first hierarchical neuro-fuzzy structure, which jointly develops the estimation of the moisture-sensitive soil thermal state and the transient cable thermal prediction, under variable, cyclic loading, through physics-based regularization and interpretable fuzzy decision support.</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

Hierarchical hybrid neuro-fuzzy framework for predictive thermal behaviour analysis of underground electrical cables under soil moisture and load variability

  • 期刊论文

  • 发表日期：

    2026年6月23日

查看全文
  • Girish M. Dhote,
  • S. Gayathri Devi,
  • K. Sujatha,
  • B. Natarajan,
  • Mylavarapu Kalyan Ram,
  • Elangovan Muniyandy

摘要

An adequate model of predicting the thermal behaviour of underground power cable systems is needed to prevent insulation failure and maximize current-carrying capacity (ampacity). The existing approach is spatial, whether it is through computationally intensive physics-based models that cannot be operated in real time to track data, or data-based surrogates that take the soil to be a stationary medium that neglects the influence of moisture-sensitive conductivity, cyclic thermal charging, and convection processes. A hierarchical hybrid neuro-fuzzy model with three integrated steps, (A) PSO-optimized ANFIS soil thermal-state estimator in the format of predicting effective soil thermal conductivity as a factor of geotechnical properties, is presented in this paper; (B) GA-optimized ANFIS cable transient thermal predictor, which uses the output of Stage A predictors to predict the temperature of the conductor, sheath, and thermal charging rate, and regularized by a physics-informed loss criterion based on the transient heat conduction equation; and three data sets are used: a 257 case database of soil thermal conductivity to be used in the Stage A calibration, 7560 FEM-simulated cable conditions under the dry, moist, and wet soil regime to be used in the Stage B training, and published experimental benchmarks to be used to independently validate those experiments. The mean RMSE of the proposed framework is 0.92 ± 0.04 °C and a mean R2 of 0.983 ± 0.002 on heldout test data which is better than plain ANFIS (RMSE 1.85 ± 0.08 °C), ANN-BP (RMSE 1.45 ± 0.06 °C) and IEC 60287 solutions (RMSE 4.31 °C). Ablation analysis reveals ability to reduce RMSE by 40.0% with optimization by PSO-GA, physics-informed loss by 14.6%, and the hierarchical cable coupling of soils by 5.3%. The inference latency of an average workstation (9.4 ms) means that it can be used with real-time dynamic thermal rating systems. It is, as far as we know, the first hierarchical neuro-fuzzy structure, which jointly develops the estimation of the moisture-sensitive soil thermal state and the transient cable thermal prediction, under variable, cyclic loading, through physics-based regularization and interpretable fuzzy decision support.