<![CDATA[Purpose – This study develops and evaluates a day-ahead forecasting framework for sulfuric acid production in a continuous chemical manufacturing process and examines whether machine-learning models provide meaningful predictive value beyond simple persistence forecasting. Methods – Five years of operational production data were processed through a validated ETL pipeline and transformed using calendar, lag, and rolling-window features. Six regression algorithms were screened using chronological time-series validation, followed by equal-budget Optuna hyperparameter optimization of selected candidates. The final model was evaluated on an untouched holdout period and through repeated walk-forward validation against persistence, seasonal-naive, and rolling-mean baselines. Permutation importance was used to examine predictor contributions. Findings – Random Forest achieved the strongest development-stage performance and produced accurate day-ahead forecasts. However, its advantage over persistence was marginal on the final holdout and inconsistent across repeated temporal evaluations. Current-day production overwhelmingly dominated predictor importance, indicating strong persistence in the underlying industrial process. Performance also deteriorated during maintenance-related shutdown conditions. Research Implications – Industrial forecasting systems should prioritize robust temporal validation and comparison with simple operational baselines before adopting more complex machine-learning models. Originality – The study provides a leakage-aware forecasting and evaluation pipeline that combines baseline benchmarking, model interpretation, and deployment within a prototype decision-support system for continuous chemical production.]]>
Copyrights © 2026
