Continuous thermal monitoring of aluminum electrolysis cells—which operate at ~950 °C under strong magnetic fields in a corrosive fluoride melt—remains an unsolved process monitoring challenge. This paper presents a fiber-optic Raman distributed temperature sensing (DTS) deployment for high-temperature cathode steel bar monitoring in a 380 kA industrial cell supplied by a grid with over 30% wind and solar penetration. A custom fiber ring packaging scheme, pressed against the underside of the cathode bar (temperature >300 °C), achieved >1 m thermal contact length within the confined space beneath the cell. Features encoding supply-side renewable power periodicity, thermal inertia, and local fluctuation intensity were engineered to anchor the learning task in the process physics of the electrolysis cell. CatBoost, LightGBM, and Random Forest were combined in a stacking ensemble with a linear regression meta-learner, attaining RMSE = 0.7451 °C and R2 = 0.9944 over two months of continuous industrial operation across seven cathode bars. Frequency-domain residual decomposition revealed why LightGBM—the aggregate-weakest base learner—received the dominant meta-learner weight (+2.15) while CatBoost—the aggregate-strongest—received a negative weight (−1.85): LightGBM uniquely minimized high-frequency error (1.06 vs. 1.29 °C for CatBoost). The ensemble advantage was spatially robust across all seven bars. The 0.75 °C RMSE establishes a noise floor for residual-based monitoring, demonstrating that process-informed feature engineering and frequency-resolved stacking ensemble learning deliver predictive accuracy suitable as a process monitoring baseline in high-temperature industrial environments under increasing renewable power penetration.

Continuous thermal monitoring of aluminum electrolysis cells—which operate at ~950 °C under strong magnetic fields in a corrosive fluoride melt—remains an unsolved process monitoring challenge. This paper presents a fiber-optic Raman distributed temperature sensing (DTS) deployment for high-temperature cathode steel bar monitoring in a 380 kA industrial cell supplied by a grid with over 30% wind and solar penetration. A custom fiber ring packaging scheme, pressed against the underside of the cathode bar (temperature > 300 °C), achieved >1 m thermal contact length within the confined space beneath the cell. Features encoding supply-side renewable power periodicity, thermal inertia, and local fluctuation intensity were engineered to anchor the learning task in the process physics of the electrolysis cell. CatBoost, LightGBM, and Random Forest were combined in a stacking ensemble with a linear regression meta-learner, attaining RMSE = 0.7451 °C and R2 = 0.9944 over two months of continuous industrial operation across seven cathode bars. Frequency-domain residual decomposition revealed why LightGBM—the aggregate-weakest base learner—received the dominant meta-learner weight (+2.15) while CatBoost—the aggregate-strongest—received a negative weight (−1.85): LightGBM uniquely minimized high-frequency error (1.06 vs. 1.29 °C for CatBoost). The ensemble advantage was spatially robust across all seven bars. The 0.75 °C RMSE establishes a noise floor for residual-based monitoring, demonstrating that process-informed feature engineering and frequency-resolved stacking ensemble learning deliver predictive accuracy suitable as a process monitoring baseline in high-temperature industrial environments under increasing renewable power penetration.