Machine learning models for predicting sucrose extraction as a basis for advanced control of sugarcane milling

Sugarcane milling efficiency is governed by a complex interaction between raw material properties, mechanical energy input and process dynamics. This study proposes a data-driven framework to predict sucrose extraction efficiency in an industrial milling tandem using Near-Infrared (NIR) spectroscopy and operational process data. Two complementary models were developed: XGBoost focusing on steady-state relationships, and a LSTM network designed to capture temporal dynamics. To enhance physical interpretability and robustness, feature engineering was guided by process knowledge, including the introduction of cane millability (Pol/fibre ratio), aggregated mechanical power indices derived from roll torque and rotational speed, and condensed roll displacement indicators. Instead of relying on autoregressive target lags, a baseline – residual decomposition was adopted, allowing the models to focus on deviations from the slow-moving extraction trend. Results show that both models consistently identify cane millability, fibre content, juice purity, imbibition-related ratios and mechanical power indices as the most influential predictors. The LSTM model captures short-term dynamic responses, while XGBoost emphasizes physicochemical and mechanical steady-state effects. This convergence across models demonstrates improved robustness, reduced dependence on persistence effects, and stronger alignment with milling physics.