Predictive Modeling of Polyphenol Concentration After Sequencing Batch Reactor Winery Wastewater Treatment ^†

Abstract

Winery wastewater contains recalcitrant pollutants, such as phenolic compounds, which can hinder biological treatment processes. While monitoring these systems is essential to prevent treatment failure, quantifying recalcitrant compounds through conventional methods can be time-consuming and costly due to complex analytical procedures and chemical disposal. In this study, machine learning (ML) was used to predict polyphenol concentration after the biological treatment of winery wastewater using a sequencing batch reactor (SBR). ML models, including ElasticNet (ENet), Multi-Layer Perceptron Regressor (MLPR), and Support Vector Regressor (SVR), were developed and tested using a small, high-dimensional dataset and leave-one-out cross-validation (LOOCV). Feature selection and hyperparameter tuning were applied to improve model performance. After optimization, the SVR model achieved the best performance, with MAE, MAPE, and R² of 0.88 mg/L, 9.3%, and 0.75, respectively.

1. Introduction

Large volumes of wastewater are generated by the winery industry, which must be treated before discharge to protect surface waters or enable water reuse. Biological treatment systems, such as constructed wetlands and activated sludge processes, offer a flexible and cost-effective approach to treat these effluents, remaining the most widely used option for winery wastewater treatment [1,2]. However, winery wastewater contains several recalcitrant and toxic compounds that can hinder biological treatment processes, with phenolic compounds being the most representative of these effluents [3,4]. In fact, metabolic inhibition and changes in microbial community composition have been reported as negative impacts of polyphenols during biological treatment [5].

Polyphenol concentrations in winery wastewater vary widely due to production-related factors and differences in characterization methodologies [6]. The concentration of phenolic compounds in winery wastewater reported in the literature ranges from 10 to 700 mg/L [7]. However, frequent assessment of these compounds can be time-consuming, labor-intensive, and economically expensive due to complex analytical methods that often require specialized equipment and chemical reagents, which demand suitable disposal. Therefore, new approaches are needed to reduce reliance on traditional methods for monitoring these compounds.

Given the complexity of wastewater treatment processes, the application of machine learning (ML) models has been explored in recent years as powerful new tools for process control and optimization [8]. For instance, distinct ML models have been successfully applied to monitor the effluent quality of biological treatment processes [9,10]. Nevertheless, data-driven models such as ML often rely on large amounts of high-quality data [11]. For instance, four effluent quality parameters were effectively predicted using 1665 samples from a wastewater treatment plant [10]. However, the limited amount of available data remains a constraining factor for the widespread application of modeling approaches to monitor biological treatment systems [12].

The aim of this work was to develop a robust ML model to predict polyphenol concentration after the biological treatment of winery wastewater, using a small but high-dimensional dataset. The dataset encompassed key operational parameters collected during the continuous operation of a sequencing batch reactor (SBR) treating winery wastewater. Three distinct ML models, namely ElasticNet (ENet), Multi-Layer Perceptron Regressor (MLPR), and Support Vector Regressor (SVR), were developed and evaluated. Leave-one-out cross-validation (LOOCV) was used to assess model generalization, as it is preferable for small datasets [13]. Two distinct pipelines were implemented for each model: a base model pipeline used as a benchmark, and an optimized pipeline designed to improve model performance and generalization.

2. Materials and Methods

2.1. Experimental Setup

A lab-scale SBR with a 4 L working volume was inoculated with activated sludge from a municipal wastewater treatment plant in Vila Real, Portugal. The SBR was operated in 12 h cycles comprising 60 min of feeding, 60 min of idling, 480 min of aeration, 90 min of sludge settling, and 30 min of effluent withdrawal. During the feeding phase, two liters of winery wastewater were supplied to the SBR, corresponding to a volume exchange ratio (VER) of 50% and a hydraulic retention time (HRT) of one day. Sludge retention time (SRT) was monitored but not controlled.

Winery wastewater was collected from a cellar located in the Douro Valley, North of Portugal. The winery wastewater was characterized as follows: 1.8 ± 0.3 g/L chemical oxygen demand (COD), 23 ± 12 mg/L polyphenols, 18.9 ± 1.1 mg/L total nitrogen (TN), 2.4 ± 0.4 mg/L total phosphorus (TP), and pH 4. Prior to reactor feeding, winery wastewater was supplemented with NH₄Cl (0.17 g/L), K₂HPO₄ (0.028 g/L), and KH₂PO₄ (0.022 g/L), resulting in final concentrations of 62 ± 9 mg/L for TN and 12.4 ± 0.4 mg/L for TP. Additionally, the pH was adjusted to 7.4 ± 0.2 using NaHCO₃ (0.58 g/L).

2.2. Analytical Methods

The influent COD and effluent COD were quantified spectrophotometrically using a Hach COD reactor and a DR 2400 spectrophotometer (Hach Co., Loveland, CO, USA). Influent and effluent samples were filtered using a syringe membrane filter (0.22 μm pore size) prior to the determination of total dissolved carbon (TDC), dissolved inorganic carbon (DIC), dissolved organic carbon (DOC), and total dissolved nitrogen (TDN) using a Shimadzu TOC–L CSH analyzer (Shimadzu, Kyoto, Japan). After sludge settling, the turbidity of the bulk medium was measured using a 2100N IS turbidimeter (Hach Co., Loveland, CO, USA). Polyphenols were determined using a spectrophotometer (DR 2400, Hach Co., Loveland, CO, USA), following the Folin–Ciocalteu method [14]. Total suspended solids (TSS), fixed suspended solids (FSS), volatile suspended solids (VSS), and sludge volume index at 30 min (SVI₃₀) were determined according to Standard Methods [15]. COD, DOC, TDN, and polyphenol removal efficiencies were calculated according to Equation (1):

$$\mathrm{Removal}\left(\%\right)={\displaystyle \frac{{\mathrm{C}}_{\mathrm{i}}-{\mathrm{C}}_{\mathrm{f}}}{{\mathrm{C}}_{\mathrm{i}}}}\times 100$$

(1)

where ${\mathrm{C}}_{\mathrm{i}}$ and ${\mathrm{C}}_{\mathrm{f}}$ are the influent and effluent concentrations, respectively.

2.3. Data Collection

The dataset used in this study was obtained from the continuous operation of the SBR treating winery wastewater. The dataset comprised 36 observations (n = 36) and 39 physico-chemical and operational parameters collected over 140 days of operation. These parameters were grouped into three categories—influent, reactor, and effluent—resulting in a total of 38 input features (X = 38) used by the ML models, along with one target variable (y) (

Table 1. Influent, reactor, and effluent features from reactor operation.

			Features
Parameter	Units	Influent	Reactor	Effluent
Temperature	°C	T_in	T_sbr	T_out
pH	-	pH_in	pH_sbr	pH_out
Chemical oxygen demand	mg/L	COD_in	-	COD_out
Total dissolved carbon	mg/L	TDC_in	-	TDC_out
Dissolved inorganic carbon	mg/L	DIC_in	-	DIC_out
Dissolved organic carbon	mg/L	DOC_in	-	DOC_out
Total dissolved nitrogen	mg/L	TDN_in	-	TDN_out
Polyphenol concentration	mg/L	PPh_in	-	PPh_out *
Flow rate	L/d	Flow	-	-
Organic loading rate	gCOD/L/d	OLR	-	-
Total suspended solids	mg/L	-	TSS_sbr	TSS_out
Fixed suspended solids	mg/L	-	FSS_sbr	FSS_out
Volatile suspended solids	mg/L	-	VSS_sbr	VSS_out
VSS/TSS ratio	-	-	VSS/TSS_sbr	VSS/TSS_out
Sludge volume index	mL/g	-	SVI30	-
Turbidity	NTU	-	Turbidity	-
Operation time	days	-	Op_t	-
Volume	L	-	Vol_sbr	-
Volume exchange ratio	%	-	VER	-
Hydraulic retention time	days	-	HRT	-
Food-to-microorganism ratio	gCOD/gVSS/d	-	FM	-
Sludge retention time	days	-	SRT	-
COD removal	%	-	-	COD_rmv
DOC removal	%	-	-	DOC_rmv
TDN removal	%	-	-	TDN_rmv

* Target variable.

).

2.4. Machine Learning Model Selection and Optimization

In this work, three ML models (i.e., ENet, MLPR, and SVR) with distinct architectures were developed to predict polyphenols. ENet has a linear structure and reduces overfitting in high-dimensional datasets through combined L1 (Lasso) and L2 (Ridge) regularization. MLPR captures non-linear relationships among features using a neural network trained via backpropagation to approximate functional dependencies in the data. SVR applies the structural risk minimization principle to balance model complexity and training error, effectively handling small and complex datasets through kernel functions.

Two distinct pipelines were developed for each model. The base pipeline included data pre-processing steps, such as missing-value imputation, data transformation, and normalization. The optimized pipeline incorporated feature selection and hyperparameter tuning steps (Figure 1). Due to the small number of observations (n = 36), LOOCV was used as the cross-validation method, allowing for maximization of the training dataset and minimization of bias [13]. In LOOCV, the number of folds (k) is equal to the number of observations, ensuring that in each fold, the model is trained on 35 observations and tested on the single unseen observation left out.

The overall model performance and generalization were evaluated using the mean absolute error (MAE), mean absolute percentage error (MAPE), and coefficient of determination (R²) [16]. All models were implemented using the scikit-learn library in Python 3.12.4.

3. Results and Discussion

3.1. Reactor Performance

An SBR was operated for 140 days to treat winery wastewater. The average COD, DOC, TDN, and polyphenol removal efficiencies throughout the operation are shown in

Table 2. SBR removal efficiencies for COD, DOC, TDN, and polyphenols.

	COD	DOC	TDN	Polyphenols
Removal (%)	92.0 ± 10.3	93.3 ± 6.0	89.0 ± 14.8	49.8 ± 14.4

Average ± standard deviation.

. Overall, the SBR exhibited good performance in terms of COD, DOC, and TDN removal, achieving average efficiencies of approximately 90%. In contrast, the average polyphenol removal remained around 50%, suggesting limited sludge ability to remove these compounds even at relatively low influent concentrations (i.e., 23 ± 12 mg/L). In fact, phenolic compounds are not easily removed through biological processes, often requiring additional treatment to achieve higher effluent quality [3,17,18,19].

3.2. Model Development and Prediction

The base ENet, MLPR, and SVR models were developed using all 38 features included in the dataset and default hyperparameters. On the other hand, optimized models were developed by selecting a subset of features to reduce model complexity and applying hyperparameter tuning to maximize model performance. The relationship between observed and predicted effluent polyphenol concentrations (i.e., PPh_out) for each model is shown in Figure 2.

All the base models exhibited low prediction performance, with R² values of 0.58, 0.40, and 0.26 for ENet, MLPR, and SVR, respectively (Figure 2a,c,e). While the base ENet model explained 58% of the variance in effluent polyphenol concentration, the base MLPR and SVR models only explained 40% and 26% of the variance, respectively. After model optimization, prediction performance improved considerably for all models, with R² values increasing to 0.71, 0.73, and 0.75 for ENet, MLPR, and SVR, respectively (Figure 2b,d,f). Interestingly, all three optimized models achieved comparable prediction performance. The optimized ENet and MLPR models explained 71% and 73% of the variance in effluent polyphenol concentration, respectively, while the optimized SVR model exhibited the highest improvement, explaining 75% of the variance in the target variable.

The MAE and MAPE scores for all models are presented in

Table 3. Summary of the overall performance metrics.

Model	Model Pipeline	MAE * (mg/L)	MAPE * (%)	CoV for MAE	CoV for MAPE
ENet	Base	1.08 ± 0.94	11.7 ± 12.5	0.87	1.07
MLPR	Base	1.44 ± 0.93	15.0 ± 11.0	0.65	0.73
SVR	Base	1.45 ± 1.26	15.6 ± 15.3	0.87	0.98
ENet	Optimized	0.94 ± 0.72	10.1 ± 8.7	0.77	0.86
MLPR	Optimized	0.90 ± 0.73	9.4 ± 8.7	0.81	0.93
SVR	Optimized	0.88 ± 0.68	9.3 ± 8.3	0.77	0.89

* MAE and MAPE are presented as the average ± standard deviation from the 36 individual test scores from each LOOCV fold.

. The relatively high standard deviations obtained in MAE and MAPE indicate that the prediction errors were inconsistent across the 36 observations. Such variability may be expected when using LOOCV, particularly when applied to small datasets, reflecting the model’s sensitivity to individual data points. The exclusion of a single observation for testing, especially extreme values, may significantly change the prediction error in each LOOCV fold. Hence, the coefficient of variation (CoV) for MAE and MAPE was calculated to evaluate model stability (Table 3).

Among the base models, ENet exhibited the lowest errors, with MAE and MAPE scores of 1.08 ± 0.94 mg/L and 11.7 ± 12.5%, respectively. In contrast, the base SVR model revealed the highest prediction errors, with a MAE score of 1.45 ± 1.26 mg/L and a MAPE score of 15.6 ± 15.3%. While the MAE and MAPE scores for the base MLPR were similar to the ones observed for the base SVR, the lower CoVs for both metrics evidence greater stability across all LOOCV folds.

The MAE and MAPE scores decreased considerably across all optimized models. Moreover, the CoVs for MAE and MAPE decreased for ENet and SVR models, indicating an improvement in the model’s stability. In contrast, although the optimized MLPR showed lower MAE and MAPE scores than the base MLPR, the higher CoV values suggest a decrease in the model’s stability. The lowest error values were obtained with the optimized SVR model, achieving MAE and MAPE scores of 0.88 ± 0.68 mg/L and 9.3 ± 8.3%, respectively. Overall, the optimized models demonstrated robust and generalizable performance to predict polyphenol concentration after biological treatment. Although prediction performance may vary under transient or unstable operational conditions, these robust ML models can be used to provide real-time assessment of polyphenol levels.

4. Conclusions

In this study, ML models were successfully applied to challenging environmental monitoring tasks, specifically the prediction of polyphenol concentration after the biological treatment of winery wastewater. The development of an optimized model pipeline was essential to achieve stable and generalizable models despite the limitation of the small, high-dimensional dataset used in this work. The optimized SVR achieved the highest performance, with the lowest MAE (0.88 mg/L) and MAPE (9.3%) scores, explaining 75% (R² = 0.75) of the variance in effluent polyphenol concentration. Overall, this work provides a data-driven tool that can be integrated into continuous process monitoring and control systems, enabling the timely assessment of effluent quality in terms of polyphenol content for safe wastewater discharge or potential water reuse.