Machine learning driven many‐objective moving horizon scheduling optimization

Abstract Industrial electrification can decarbonize chemical manufacturing, but it exposes operations to volatile electricity prices and carbon intensities. This work develops a machine learning‐enhanced many‐objective moving horizon scheduling framework that predicts objective correlation groupings from 48‐hour price and emission‐intensity profiles, avoiding repeated online dimensionality reduction and unnecessary Pareto frontier generation. Historical grid scenarios labeled by an objective dimensionality reduction algorithm are used to train classifiers, achieving 94% accuracy for five‐class grouping identification and nearly 99% accuracy for binary cost‐emission relationship classification. In flexible ammonia production, the random forest model enables efficient scheduling on unseen ISO New England data by identifying when compromise solutions are needed and when objectives are sufficiently correlated. The framework is also tested on chlor‐alkali electrolysis. Cross‐process results show that objective relationships depend on process constraints and operating flexibility, while transfer learning improves LSTM performance only under limited target‐process data.