Predictive Modelling of Polyphenol Concentration after Sequencing Batch Reactor Winery Wastewater Treatment

The treatment of winery wastewater (WWW) is crucial for surface water protection and/or enabling water reuse. While biological treatment offers a flexible and cost-effective approach, monitoring and process control remain quite complex. Machine learning (ML) models have gained attention as effective tools to address these challenges. However, extensive operational data from such systems is inherently limited; monitoring specific recalcitrant compounds (e.g. polyphenols) is time-consuming and costly due to complex analytical methodologies and chemical reagent disposal. This study aimed to develop and evaluate ML models to predict polyphenol concentration after biological treatment using a small and high-dimensional dataset. A Sequencing Batch Reactor (SBR), fed with WWW, was monitored for 140 days, generating a small yet comprehensive dataset (38 features and 36 observations) that captured different operational conditions. These features served as input to predict polyphenol concentration (target). Three ML algorithms were evaluated: ElasticNet (EN), Support Vector Regression (SVR) and Multi-Layer Perceptron Regressor (MLPR). To maximize training data and ensure robust evaluation given the limited dataset, Leave-One-Out Cross-Validation (LOOCV) was used for model performance assessment. Mean Absolute Error (MAE) and Mean Absolute Percentage Error (MAPE) were employed as evaluation metrics. For models with default parameters, the EN model demonstrated the best initial predictive capacity for polyphenol concentration (MAE: 1.08 ± 0.94 mg/L; MAPE: 11.7 ± 12.5%). However, after feature selection and hyperparameter tuning, the SVR model achieved superior predictive capability (MAE: 0.88 ± 0.68 mg/L; MAPE: 9.3 ± 8.3%) via LOOCV. This study established a robust SVR predictive model for polyphenol concentration in SBR winery wastewater treatment. Despite the small dataset, LOOCV, feature selection and hyperparameter tuning ensured the development of a robust model with promising performance and generalization. This predictive tool offers significant potential for real-time monitoring, SBR optimization, and enhancing polyphenol removal, contributing to environmental sustainability in winery wastewater.