Machine learning based sugar recovery prediction model in sugarcane agroindustry

Abstract

Importance of the work: Sugar recovery is crucial for mill efficiency. In Indonesia,
recovery rates declined by 1.91% from 2019 to 2023, hitting a decade low of 6.6% in 2022.
These fluctuations indicate inefficiencies. This work introduced a machine learning-based
model for early prediction and process optimization to achieve better production planning.
Objectives: To develop a model for predicting final sugar recovery with machine learning,
using a multistage process and related variables in processing terms.
Materials and Methods: Day-to-day data from Sugar Mill XYZ (2020–2024) in West Java,
Indonesia were used, including Brix, purity and pol values from multiple processing stages.
Random forest, extreme gradient boosting (XGBoost) and artificial neural network (ANN)
methods were used to develop a model and its evaluation using mean squared error (MSE)
and mean absolute error (MAE). mean absolute percentage error (MAPE), coefficient of
determination (R²) and feature importance analysis.
Results: The model was developed using 699 daily milling records from 2020–2024,
comprising 18 initial features. XGBoost outperformed random forest and ANN, achieving an
MAE of 0.116, MSE of 0.03852, MAPE of 1.81% and a coefficient of determination (R²) of
86.84% on the testing set. Feature significance analysis, which combines machine learning
insights with empirical plant data, identified the key variables that had the greatest impact on
sugar recovery, such as boiling house recovery, winter recovery, Pol in cane, milling potential
efficiency and Pol in bagasse. The model correctly predicted the daily sugar recovery for
production in 2024.
Main finding: This work provides a decision-support tool for sugar mill optimization.
It illustrates how well XGBoost and random search optimization work together to predict
sugar recovery depending on process variables.