Soft-computing-driven Uncertainty Propagation in Finite Element Thermal Models of Underground Cable Networks: a Comparative Study of ANFIS, Deep Learning, and Probabilistic Fuzzy Approaches

Proper forecasting of conductor temperature and ampacity within an underground cable network is crucial for safe operation and for proper network capacity analysis. A finite element (FE) thermal solver ensures high-fidelity answers. Nevertheless, when uncertainty in soil thermal resistivity, ambient temperature, loading, burial geometry, and sheath losses must be propagated through the model, the use of FE becomes impractical. Most existing soft-computing surrogates have been proposed as individual replacements, but few have been examined under the same uncertainty-quantification (UQ) microscope. This paper proposes a framework in which an FE truth model is combined with three surrogate families: an Adaptive Neuro-Fuzzy Inference System (ANFIS); a Bayesian deep neural network (Bayesian DNN) with Monte Carlo Dropout and heteroscedastic loss; and a probabilistic fuzzy ensemble with α-cut Monte Carlo propagation. Identical datasets are used to train and evaluate all three. Latin Hypercube Sampling is used to create 5000 FE designs of a 132 kV trefoil cable circuit. Models are assessed based on point-prediction error, predictive interval calibration, robustness to input noise, and inference cost. The Bayesian DNN obtains the highest R^2 (0.994), the lowest RMSE (0.94 °C), and well-calibrated 95