What question did this study set out to answer?

The research aims to enhance temperature predictions for hollow conductors in hydrogenerators using a physics-guided approach.

June 20, 2026Thermal Science0 citationsOpen Access

Physics-guided feature selection for temperature prediction of stator winding’s hollow conductors in evaporative cooling hydrogenerators

YDYiwei DingInstitute of Process Engineering JAJian AiJiangnan University JCJ Y ChenXi'an University of Science and Technology

Key Points

The research aims to enhance temperature predictions for hollow conductors in hydrogenerators using a physics-guided approach.
Developed a surrogate modeling framework combining experimental data and Gaussian Process Regression.
Optimized monitoring and design models using 5-fold cross-validation and Leave-One-Group-Out validation techniques.
Identified two distinct feature subsets to improve model accuracy.
The GPR model reduced RMSE by 16.2% compared to Random Forest and 9.7% compared to XGBoost.
Achieved an R2 of 0.893 and RMSE of 1.770 °C with a six-feature monitoring model.
A four-feature design model resulted in a pooled R2 of 0.616 and RMSE of 3.200 °C.

Abstract

Accurate temperature prediction for hollow conductors in evaporative cooling hydrogenerators is critical for design optimization but hampered by the high cost of experiments and simulations. To address this, this paper proposes a physics-guided surrogate modeling framework using experimental data from 12 conductor designs. An optimized Gaussian Process Regression (GPR) model is shown to outperform Random Forest and XGBoost, reducing RMSE by 16.2% and 9.7%, respectively. The framework identifies two physically distinct feature subsets for complementary use cases. A six-feature (6D) monitoring model that achieves 𝑅2=0.893 and RMSE =1.770 ∘𝐶 under random 5-fold cross-validation, and a four-feature (4D) design model that obtains pooled 𝑅2=0.616 and RMSE =3.200 ∘𝐶 under rigorous Leave-One-Group-Out (LOGO) validation, sufficient for ranking candidate designs. The analysis further identifies outlet measurements as information-leaking features that inflate within-design accuracy but degrade extrapolation to unseen geometries, highlighting the importance of causal feature selection for robust design-stage surrogate models.

Ask AI

Helpful

Bookmark

View Full Paper