Green machine learning assisted analytics for wastewater valorisation: a case study of Melbourne, Australia

Wastewater valorisation inherently benefits from predictive models that are accurate, generic, robust and environmentally sustainable. Benchmarked on a publicly available municipal wastewater dataset from Melbourne, Australia, characterised by temporal variability and multivariate complexity, we present a Green Machine Learning (GML) framework to predict Chemical Oxygen Demand (COD) and Total Nitrogen (TN), two variables directly linked to carbon and nutrient recovery pathways. Four widely used machine learning and deep learning models - FFNN, RNN, CNN, and XGBoost - were benchmarked to identify the most optimised approach, defined here as the model providing the best overall balance between predictive accuracy, generalisation, and computational environmental cost rather than the highest training performance alone. XGBoost and RNN emerged as the most optimised algorithmic pathways. RNN captured the temporal behaviour most effectively for COD, achieving an R 2 of 0.71 and an RMSE of 75.55 mg/L, while XGBoost showed the strongest generalisation, with validation/test R 2 values of 0.72/0/71, respectively. For TN, XGBoost was the best-performing model, with validation/test R 2 = 0.59/0.63, outperforming RNN (0.56/0.53) and the other neural architectures. Sustainability analysis using CodeCarbon showed that XGBoost required only 0.000018 kWh and emitted 0.0000129 kg CO 2 -eq for COD, compared to 0.001490 kWh and 0.001063 kg CO 2 -eq for RNN; for TN, XGBoost consumed 0.000028 kWh, compared to 0.002563 kWh for RNN. These results conclusively show that XGBoost delivers comparable or better predictions at 98%–99% lower computational carbon cost, while RNN remains the strongest neural network candidate for COD-related dynamics.

Wastewater Treatment Plants (WWTPs) are undergoing a paradigm shift towards becoming resource recovery facilities rather than solely pollution abatement facilities. Globally, around 380 billion m 3 of wastewater was generated annually as of 2015 (WWAP, 2017), while more recent estimates indicate that 332 billion m 3 of household wastewater alone was generated in 2024 (UN Statistics Division, 2024). This number is expected to double by 2050 due to rapid urbanisation, while only 56% of household wastewater is treated for safe discharge into the environment (United Nations Statistics Division, 2025). This raises serious concerns regarding the efficiency of wastewater treatment plants (WWTPs), indicating the existing gap in wastewater management. WWTPs are designed to eliminate pollutants that contribute to Biological Oxygen Demand (BOD), Chemical Oxygen Demand (COD), and other nutrients prior to discharge. Organic compounds that cause COD can be valorised to Biogas, Volatile Fatty Acids (VFAs), biopolymers, etc., which supports the concept of wastewater-enabled biorefineries (Romano et al., 2025).

A key nutrient in wastewater is nitrogen, which is often measured as Total Kjeldahl Nitrogen and can be recovered for use as fertiliser. Recent studies on wastewater valorisation have highlighted the need to recover nutrients from wastewater to prevent pollution while supporting sustainable agriculture (Zhou et al., 2023). Hence, nitrogen recovery from wastewater can both increase resource efficiency and reduce environmental impact, in line with circular economy concepts and the UN Sustainable Development Goals (SDGs) 6, 9 and 12 (Larsen et al., 2016; Wada et al., 2026). This information, combined, allows modern-day WWTPs to be appreciated as a component of the circular economy. WWTP can thus be treated as resource recovery facilities rather than just treatment facilities, analogous to the Metal Recovery from E-waste (MREW) facilities reported earlier (Debnath et al., 2018; Debnath et al., 2019).

In WWTPs, both COD and Nitrogen are important parameters that can be valorised rather than removed from wastewater. COD is the oxygen equivalent of the organic components of the wastewater that can be oxidised by strong chemical oxidants, especially potassium dichromate and potassium permanganate (Wojnárovits et al., 2024). This organic content can serve as a carbon source for the development of other products. A popular example of efficient resource management concerns the conversion of organic carbon in wastewater to biogas via anaerobic digestion (AD). AD has the potential to generate significant amounts of clean energy from wastewater and can substitute for 40%–76% of total electrical energy in WWTPs (Silvestre et al., 2015). COD can also be used to produce Volatile Fatty Acids (VFAs) and Polyhydroxyalkanoates (PHAs) (Sánchez-Ledesma et al., 2023). Researchers have emphasised that VFAs recovered from wastewater are emerging as critical pathway intermediates for biochemical and biofuel production (Owusu-Agyeman et al., 2023). PHAs have attracted significant attention in the last decade as the next-generation of sustainable materials, an alternative to petrochemical polymers. PHAs can also be synthesised from VFAs using them as a carbon source (Szacherska et al., 2021; Tao et al., 2025), thereby aligning with circular bioeconomy concepts (Ramos-Suárez et al., 2021). This has been further demonstrated by pilot-scale studies showing the high feasibility of PHA production from waste-activated sludge (Mineo et al., 2025). Sepúlveda et al. (2025) emphasised carbon recovery from WWTPs through the synthesis of biopolymers such as PHAs, while endorsing it as a key game-changer for establishing sustainable wastewater-based biorefineries. As Gracia et al. (2025) demonstrated using a statistical modelling approach, VFA production from municipal wastewater sludge can be optimised to increase resource efficiency. Wastewater also contains nitrogen, which is quite valuable. Instead of strict nitrogen removal, it is possible to treat it and convert it into forms such as struvite (magnesium ammonium phosphate) and ammonia-based fertilisers. Empirical research indicates that nutrient recovery technologies can capture up to 15%–20% of the world’s phosphorus and nitrogen requirements from wastewater streams (Cordell et al., 2011). On the other hand, the release of untreated nitrogen into water bodies may cause eutrophication, leading to excessive algal growth and, consequently, oxygen depletion (Smith and Schindler, 2009). In addition, nitrogen-removal processes can release nitrous oxide, a greenhouse gas with a global warming potential about 298 times that of CO2 (Intergovernmental Panel on Climate Change, 2014).

Machine Learning (ML) and Deep Learning (DL) have challenged conventional understandings across almost every sector while opening multiple avenues for understanding kinetics, process dynamics, patterns, and more at a greater granularity (Mowbray et al., 2022; Liu et al., 2024; Chowdhury et al., 2026). Tırınk (2025) developed a predictive model using XGBoost to predict azo dye removal from wastewater using boron-doped diamond electrodes, achieving 99.6% training efficiency and 95.4% test set efficiency. ML is also being implemented at a higher level to retrieve, understand, and predict the system behaviour of WWTPs (Kumar et al., 2025). The huge amount of operational data generated by WWTPs is an ideal feed for modern ML models to identify patterns, understand system dynamics, and then use them to enhance plant efficiency. Now, plant efficiency can be both treatment efficiency and resource recovery efficiency. While many mature ML models exist, Artificial Neural Networks (ANNs) and ensemble approaches such as gradient boosting are widely adopted for wastewater systems (Dantas et al., 2023; Wang et al., 2025). ANN has traditionally been the researchers’ favourite in Wastewater Treatment (WWT) analytics, primarily because of its ability to capture nonlinear relationships between water quality parameters and process variables (Nikhar et al., 2025; Dikmen et al., 2025).

The use of ML and DL models has proven to be effective in predicting key parameters, including Biochemical Oxygen Demand (BOD) (Ching et al., 2022), Chemical Oxygen Demand (COD) (Çimen Mesutoğlu and Gök, 2024), and Total Suspended Solids (TSS) (Gholizadeh et al., 2024), in the wastewater systems. Recently, Çimen Mesutoğlu and Gök (2024) reported that ANN models can efficiently forecast COD concentrations within the industrial WWTPs using historical operational data. Hamada et al. (2024) used Gaussian Process Regression (GPR), Random Forest (RF), eXtreme Gradient Boosting (XGB), and Light Gradient Boosting Machine (LightGBM) to predict effluent TSS, COD, and BOD. GPR was found to be the forecasting method with the best accuracy, as indicated by correlation coefficients (cc) of 0.964, 0.95, and 0.975 for TSS, COD, and BOD prediction, respectively. The Light Gradient Boosting Machine (LightGBM) framework can thus be considered a key DL approach for predictive analytics (Ma et al., 2025). Huang et al. (2025) employed a symmetry-inspired DL framework to accurately predict nitrous oxide emissions in wastewater treatment, revealing that water temperature is a critical factor, with dissolved oxygen and inlet flow rate also influential. Recently, Gulshin and Kuzina (2024) have highlighted the importance of historical operational data from WWTPs for data-driven modelling. Duarte et al. (2023) highlighted the use of ML and DL for anomaly detection and for predicting WWTP performance, including for system optimisation. Xiong et al. (2025) introduced a dual hybrid modelling framework that amalgamates LSTM and XGBoost models to predict COD, ammonia nitrogen, TN, and total phosphorus, achieving R2 values of 0.95–0.98. Yun et al. (2025) developed an XGBoost-based optimisation for urban WWTPs using ecological combination ponds (ECPs) in small towns across China to minimise Total Nitrogen (TN) discharge. The model achieved high predictive accuracy (R 2 = 0.91) while reducing TN discharge by 17.5%, COD dosage by 33.29%, and annual carbon emissions by 788.40 t CO 2 /y, calculated based on TN removal. Similarly, Carbureanu and Gheorghe (2026) developed a Histogram-Gradient Boosting Regression (HGBRCond) model to predict conductivity during biodegradation in industrial WWTPs. The model achieved R 2 = 0.877 and RMSE = 10.24 μS/cm across 424 datasets, with 6.8% better performance than standard gradient boosting models. These studies highlight the importance of collecting and preserving WWTP monitoring data. The growing trend of using public (epidemiological) datasets and open-platform datasets to train ML models is proof of this. Hence, ML models trained on such open-access datasets are both worthwhile and timely.

ML and DL models that are purely data-driven have shown significantly promising results in WWT systems. WWTPs are complex systems that are influenced not only by variables such as temperature, BOD, COD, hydraulic loading, etc., but also by kinetic parameters that govern the processes and their interactions. A key disadvantage of public datasets is that they do not include all the variables mentioned above, thereby reducing the predictive capabilities of data models (Maqsood et al., 2025; Duarte et al., 2023). This is a primary reason for the moderate prediction accuracy observed in such studies. State-of-the-art WWT studies focusing on data modelling using real plant data show that R2 values typically range from 0.6 to 0.8, emphasising the natural data variability and noise present in operational datasets (Khoshvaght et al., 2025; Dikmen et al., 2025). Another significant challenge for data-driven models is the risk of overfitting. In such cases, the models achieve very high training accuracy but fail to predict correctly, as they are more focused on pattern recognition than on interpreting the system behaviour (Dansawad et al., 2023). Data quality is another big issue that affects the reliability of the models. Sensor abnormality, missing data points, and noisy data are common characteristics of WWTP datasets, which directly affect the robustness of the models (Mantena et al., 2026). Moreover, many studies have focused more on prediction accuracy than on guidance frameworks for choosing suitable models for real-life WWTP datasets (Narayanan et al., 2024). While comparative ML/DL studies for wastewater prediction are increasingly common, model selection is still predominantly based on predictive accuracy metrics alone, and hence remains a major point of arbitrariness. Combining predictive performance, model generalisation, overfitting behaviour, and environmental sustainability within a unified decision-support framework could be the key to appropriate wastewater valorisation. The advantage is expected to be particularly relevant in the context of public operational datasets, where missing process-state variables and inherent data variability constrain predictive performance that may influence the suitability of different machine learning architectures. The present study is a major filler to this knowledge-gap.

Modern probabilistic models, e.g., Bayesian Inference, tend to use large amounts of data, which requires substantial electricity due to the embedded repetitive supervised training. According to de Vries-Gao (2026), AI systems could be responsible for 32.6 to 79.7 million tons of CO 2 emissions in 2025, primarily due to model training and deployment and the associated energy consumption. The rapid growth of data centres supporting AI loads is also noticeable. In 2023, about 4.4% of electricity was consumed by data centres in the United States alone, a figure that is increasing exponentially as AI applications expand (Kandemir, 2025). Even if it is speculated that, under the current trend, data centres could alone consume nearly 9% of global electricity demand by 2030 (Hankendi et al., 2025). For example, deep neural networks can significantly increase the energy footprint of AI systems, especially as models grow larger or more complex (Scala et al., 2025). Because of these concerns, the concept of Green Machine Learning (Green ML) has gained attention in recent years. Green ML aims to design models that achieve good predictive performance while being energy-efficient (Khan et al., 2026). The challenge, then, is to optimise between model accuracy and computational energy consumption when selecting machine learning models for environmental applications. In this study, Green ML is interpreted as the simultaneous optimisation of predictive accuracy and environmental footprint, rather than efficiency alone. This is a key difference to previous studies (Ching et al., 2022; Gholizadeh et al., 2024; Dikmen et al., 2025; Wang et al., 2025; Xiong et al., 2025) which uphelds the lineage of (Schwartz et al., 2020; Kaack et al., 2022) which emphasise the importance of treating efficiency as a primary evaluation criterion alongside accuracy, thereby framing Green AI as a multi-objective optimisation problem balancing model performance with resource consumption and environmental footprint.

Other studies abiding similar philosophy have also explored sustainability-aware evaluation strategies by integrating predictive performance with energy and carbon considerations (Castellanos-Nieves and García-Forte, 2024; Hasan et al., 2024). However, such approaches remain relatively underexplored in wastewater valorisation systems, particularly when model generalisation, overfitting, and environmental burden are assessed simultaneously. In this study, Green ML refers to a sustainability-aware model evaluation framework that extends beyond the traditional trade-off between predictive accuracy and computational efficiency. Rather than proposing a new machine learning algorithm, this study focuses on establishing a sustainability-aware bootstrapping model evaluation framework, a holistic undertaking, for realistic wastewater datasets. The framework jointly considers predictive accuracy, model generalisation, overfitting behaviour, and computational sustainability through sustainability-adjusted indicators, thereby providing practical guidance for selecting suitable ML models for wastewater valorisation applications. Consequently, the contribution of this work lies in the integrated evaluation methodology rather than the development of a new predictive architecture.

Study design, dataset, and target definition

A key aspect of this study is to assess how different machine learning and deep learning paradigms perform when applied to wastewater valorisation analytics using an open operational dataset. The dataset was obtained from the publicly available Kaggle repository (Kaggle Data set, 2026) and contains physicochemical and operational observations from a municipal wastewater treatment plant in Melbourne, Australia. Sixteen recorded variables define the raw dataset. Based on industrial input, two response variables were selected as prediction targets: Chemical Oxygen Demand (COD) and Total Nitrogen (TN). While COD represents the carbon-rich organic fraction relevant to biogas, volatile fatty acids, and other biochemical recovery pathways, TN represents the nutrient fraction relevant to ammonia and fertiliser-oriented recovery. In practice, COD forecasts can support feed allocation to anaerobic digestion, VFA-oriented fermentation, and carbon-routing strategies. In contrast, TN forecasts can inform nutrient recovery, aeration control, and compliance-risk management. Framing the targets this way links model evaluation directly to both resource recovery and day-to-day wastewater management decisions. The study also compares outcomes from different ML and DL architectures under an identical experimental design, then jointly assesses them on predictive quality and environmental burden using bootstrap standardisation.

Data preparation and feature screening

Data preprocessing was performed using a structured pipeline to improve data quality, reduce redundancy, and preserve interpretability. Data cleaning, variance screening, and scaling parameters were derived from the training subset only and then transferred unchanged to validation and test subsets.

First, the discrete, tabular dataset, spanning 1,382 records, was screened for structural inconsistencies, missing values, and variables with negligible information content. Each row represents an independent observation with operational parameters used as predictors. After preprocessing and feature selection, only the important input features have been retained for model development. Features showing near-zero variance across the observation window, namely, atmospheric pressure, total rainfall, and average visibility, were removed because they contributed little discriminatory value for the selected targets. The remaining numerical variables were standardised to a mean of 0 and a variance of 1 to prevent distance-based analyses and neural network training from being dominated by outliers (Elen and Avuçlu, 2021). This choice is crucial for wastewater systems because operational datasets often contain temporal drift, sensor noise, and occasional abnormal observations (Bellamoli et al., 2023); therefore, the preprocessing pipeline must mimic how a model would behave when deployed on unseen plant data rather than under idealised retrospective conditions.

Feature screening was then implemented in two complementary steps. In the first step, assuming a multivariate normal distribution, the Pearson correlation coefficient was used to analyse linear, monotonic dependencies among variables (van den Heuvel and Zhan, 2022). This was followed by the non-normal (and nonlinear) dependence assessment using Spearman and distance correlations (Spearman, 1904). In the second step, ANOVA-based relevance screening was applied to prioritise variables with the strongest statistical relationships with the target response (Pérez-Cova et al., 2022). This combined relevance-redundancy approach produced a compact predictor space that retained the dominant operational signal while limiting collinearity and unnecessary model complexity. The final reduced feature set contained seven predictors and was used consistently across all architectures to ensure a fair comparison. For each modelling task, the target variable itself was excluded from the predictor matrix.

The 1,382 observations were divided into a 70% training set, a 15% validation set, and a 15% test set. To ensure that the choice of these three sets–training, validation, test–were unbiased, a Gaussian random number generator was used to pick the first 70% training data, followed by the other two sets, both randomised. The training set was used for model fitting, the validation set for hyperparameter selection, and the test set for the final performance evaluation. The same preprocessing procedure and data split was maintained across all algorithms, e.g., FFNN, RNN, CNN, and XGBoost, ensuring uniform standardisation.

The feature-screening stage was designed to balance predictive power, interpretability, and deployability. Correlation analysis was first used to understand the joint contribution from variables as overlapping clusters. ANOVA-based screening was then used to rank variables by their statistical contribution to COD and TN estimation. This reduced predictor space is advantageous for WWTP applications because energy-thrifty models are easier to train into low-overhead decision-support workflows.

Comparative machine learning and deep learning framework

Four supervised architectures were implemented: Feedforward Neural Network (FFNN), Recurrent Neural Network (RNN), one-dimensional Convolutional Neural Network (CNN), and eXtreme Gradient Boosting (XGBoost). These models represent distinct learning logics for structured environmental data: dense nonlinear mapping (FFNN), sequence-aware learning (RNN), local pattern extraction (CNN), and boosted decision tree ensembles (XGBoost). The core methodological contribution of the study is the direct cross-paradigm comparison of these architectures under the same dataset, feature space, target variables, and train-validation-test split. This design allows observed performance differences to be attributed to model behaviour rather than to inconsistent experimental conditions.

The comparative framework, standardised and bootstrapped across multiple ML and DL regimes, was designed to reflect realistic choices available to water utilities by combining predictions from the seven protocols. FFNN provides a baseline nonlinear learner for tabular data, RNN tests whether ordered dependency information improves prediction, CNN probes local feature-pattern extraction, and XGBoost represents a strong, computationally efficient tree-based alternative. Evaluating these architectures side by side allows the study to determine not only which model predicts best, but which model is most appropriate for practical wastewater valorisation and management settings, a template for future such studies.

All models were trained using the same data partitioning strategy of 70% for training, 15% for validation, and 15% for testing. Preprocessing parameters were fitted on the training subset and then applied to the validation and test subsets to avoid data leakage. Hyperparameter selection was guided by validation performance, while regularisation and early stopping were used for the deep learning neural network models (FFNN, CNN, and RNN) to reduce overfitting. The standardised predictor vector was represented as a one-dimensional, ordered input to ensure identical information content across FFNN, RNN, and CNN implementations. From a methodological perspective, this harmonised design ensures bespoke standardisation, alleviating model bias that often leads to incorrect predictions due to inconsistent data preparation.

The predictor vector was represented in two formats depending on the model architecture. For FFNN and XGBoost, the seven predictors were passed as a flat feature vector. For RNNs and CNNs, the same information was arranged into a one-dimensional ordered representation of shape (7, 1). This arrangement enabled recurrent layers to process feature-space information via the hidden-state mechanism, while convolutional layers apply filters across feature positions to detect localised interaction patterns. Primarily, the RNN architecture was designed to capture feature dependencies. The CNN architecture was employed to identify localised interaction patterns among the parameters through its convolutional filter mechanism and hence was a complementary conduit. The motivation for including RNNs and CNNs alongside FFNNs and XGBoost lay in the cross-paradigm experimental design. Wastewater operational parameters tend to vary in ways that reflect underlying process states. By evaluating all four paradigms under identical experimental conditions, viz., same dataset, same features, same targets, same partition, the observed differences could be established within the same standardisation metric. The study thus provides a generic foundation for wastewater analytics.

Feedforward neural network (FFNN)

The FFNN was used as the baseline deep learning model for nonlinear regression of tabular wastewater data. It learns a direct mapping from the input vector to the target through stacked dense layers with nonlinear activation functions. In standard nomenclature, the predicted output can be written as: where x is the input feature vector, the W 1 matrix connects the input layer to the hidden layer, and the W 2 matrix connects the hidden layer to the output layer) and b’s are the respective weights or bias vectors, σ. denotes the hidden-layer activation, and f (.) is the output mapping (Ng et al., 2025). For wastewater management, the value of the FFNN baseline lies in its simplicity: it can approximate nonlinear process-response behaviour in WWTP data without imposing substantial computational overhead or requiring explicit sequential structure in the inputs.

Recurrent neural network (RNN)

The RNN was implemented to test whether sequential dependency modelling offers an advantage for the present dataset. At each step t, the hidden state is updated according to:

Moreover, the output is obtained as: where x t is the input at step t, h t is the hidden state, W xh is the weight matrix between the input and hidden layer, W hh is the weight matrix for the recurrent connection, W hy is the weight matrix between the hidden and the output layer, b h and b y are bias vectors, and σ h, σ y are the activation functions, typically the hyperbolic tangent function (tanh) or the rectified linear unit. By carrying information forward through the hidden state, the RNN can, in principle, capture temporal or ordered dependencies that are not explicitly available to a feedforward model (Wongburi and Park, 2023). In the wastewater valorisation context, RNNs are methodologically useful because wastewater datasets may contain ordered observations that implicitly reflect process inertia, delayed biological responses, or carryover effects between operating states.

eXtreme gradient boosting (XGBoost)

XGBoost was selected as the principal non-neural comparator because boosted tree ensembles are well-suited to nonlinear tabular data and building sequential trees with moderate-sized environmental datasets. The model prediction is expressed as: where f k denotes the k th regression tree, and K is the total number of trees. The input x i refers to the feature vector of sample i. XGBoost optimises the ensemble sequentially, so that each new tree corrects errors made by previous trees, while regularisation controls tree complexity and reduces the risk of overfitting (Xu et al., 2023). For valorisation studies, XGBoost is especially relevant because it can capture nonlinear interactions among operational variables while remaining fast to train and comparatively transparent to tune. These properties make it attractive for near-real-time wastewater management, where retraining frequency, energy use, and robustness often matter as much as marginal gains in predictive accuracy (Ching et al., 2022).

Convolutional neural network (CNN)

A one-dimensional CNN was included as a fourth comparator to test whether local feature-pattern extraction improves regression performance on structured wastewater data. For a convolutional layer, the activation of the j th filter at position r can be written as: where w ij are filter weights, b j is the bias term, k is the kernel size, and g (.) is a nonlinear activation function (Dai, 2021). In this work, the CNN acted as a controlled test of whether local ordered patterns among the retained predictors contain learnable structure beyond what can be captured by dense layers or tree ensembles. In this study, we used a 1D convolutional layer. The selected features were reshaped into a 1-D feature vector for learning spatial features. The CNN configuration used has a kernel size = 3, as this size efficiently captures local feature interactions while maintaining computational simplicity. The CNN was retained as methodological control rather than as an assumed best performer. A negative or weak result from CNN is still informative because it helps to rule out unnecessary architectural complexity for structured WWTP datasets.

Model implementation summary

The FFNN comprised two hidden layers containing 64 and 32 neurons, respectively, both with ReLU activation, and a single-neuron output layer with linear activation. The RNN employed a single recurrent layer with 32 units and ReLU activation, followed by a single-neuron dense output layer with linear activation. The CNN comprised a Conv1D layer with 32 filters, a kernel size of 3, and ReLU activation, followed by a MaxPooling1D with a pool size of 2, a Flatten layer, a dense layer with 64 neurons and ReLU activation, and a single-neuron output layer with linear activation. XGBoost was configured with 100 boosting trees, a maximum tree depth of 6, and a learning rate of 0.3.

All neural network models were trained across 400 epochs with a batch size of 50. Validation data were monitored throughout training to track generalisation behaviour. The identical architecture and training configuration were applied for both COD and Total Nitrogen targets. This was done to ensure that observed differences in performances between variable targets reflected data characteristics rather than model configuration differences.

Comparative evaluation protocol

Model comparison was performed through a unified evaluation protocol centred on accuracy, robustness, and generalisation across multiple DL and ML manifolds. Predictive performance was quantified using the coefficient of determination (R 2), Root Mean Square Error (RMSE), and Mean Absolute Error (MAE). These metrics were examined separately for COD and TN as risk assessment measures so that the architecture behaviour could be interpreted against two different wastewater valorisation objectives. In addition to scalar metrics, actual-versus-predicted plots, residual distributions, regression trendlines, and error histograms were analysed to identify systematic bias, heteroscedasticity, and sensitivity to extreme events.

The evaluation protocol was also designed to emphasise the usefulness of decisions. In addition to standard regression metrics, model behaviour was interpreted in terms of its suitability for supporting valorisation and wastewater management actions. For example, consistent prediction across typical operating ranges is important for routine control. In contrast, performance under high-load conditions is more relevant for anticipating recovery opportunities, shock loads, or nutrient-management interventions. Accordingly, the comparison does not treat all errors equally. Bias at elevated COD can affect carbon routing and estimates of biogas potential, while uncertainty in TN predictions can influence nutrient-recovery planning and nitrogen-removal control. This application-aware reading of the metrics strengthens the methodological contribution of the study beyond a generic benchmark exercise.

Green machine learning sustainability assessment

To complement predictive benchmarking, we embedded a Green ML audit into the experimental workflow. Training-time sustainability metrics were recorded using CodeCarbon, which estimates hardware-level electricity use and converts it to carbon dioxide equivalent emissions based on the relevant electricity mix (Courty et al., 2024). For each model-target combination, the following indicators were collected: RAM energy consumption, CPU energy consumption, total electricity consumption, estimated water consumption associated with data-centre cooling, and total CO2-equivalent emissions.

The sustainability assessment was included because a wastewater ML (or DL) model should be beneficial at the system level, not only accurate at the algorithm level. If a model delivers only small predictive gains while requiring substantially higher energy, water, and carbon inputs, its environmental justification becomes weak in a circular-economy context. Embedding CodeCarbon measurements into the workflow aligns the green analytics pipeline with the broader sustainability objectives of wastewater valorisation. These indicators were jointly interpreted as the algorithmic greenness profile for each model. Rather than treating sustainability as an auxiliary observation, the present study uses it as a coequal model-selection criterion alongside accuracy. Accordingly, the preferred model is not the one with the highest predictive score alone, but rather the one that offers the most credible balance among predictive utility, robustness, and computational cost. To ensure genericity, we have calculated the Environmental Cost Index (ECI), Green Performance Index (GPI), and Triple Green Sustainability Score (TGS), which highlight the environmental burden, the level of green performance, and the overall level of sustainability. These are defined below –

Where Ei, CO 2i and Wi are individual values of Energy Consumption, Carbon Dioxide Emission and Water Consumption, respectively. E ¯, C O 2 ¯, and W ¯, are the mean values of the respective parameters.

Here R t e s t 2 is the regression coefficient of the ML models during the test phase. The reason we use the test regression coefficient is that it represents the actual predictive capability on unseen data.

Here E n, C O 2 n, W n are the dimensionless Energy Consumption, Carbon Dioxide Emission and Water Consumption, respectively. This was estimated by dividing each green metrics by the arithmetic mean of the corresponding metric computed across all ML models. This is particularly helpful in defining environmental burden relative to the average model performance.

This integrated evaluation framework distinguishes the present work from most wastewater ML studies, which focus primarily on performance gains without justifying the additional carbon footprint. Green ML thus becomes a strategic selection principle for wastewater management rather than a purely academic evaluation.

Wastewater valorisation is a vast field of study. Among the many parameters, Chemical Oxygen Demand (COD) signifies the recoverable fractions of organic carbon for bioenergy, VFAs, and biochemicals. On the other hand, nitrogen facilitates nutrient recovery, including ammonia, struvite, and even fertiliser-grade nutrients. This enables the formation of a carbon-nitrogen valorisation axis. Simultaneously, both COD and Nitrogen control variables in secondary treatment influence aeration, sludge yield, and GHG emissions. ML/DL-guided predictive analytics can help identify green performance windows, thereby enabling process optimisation and energy balancing, leading to circular-economy benefits and climate co-benefits, and justifying the choice of COD and TN (Total Nitrogen) as our target variables.

From a wastewater management perspective, this carbon-nitrogen prediction axis is valuable because it connects monitoring to action. Reliable forecasts can help operators identify periods of elevated recovery potential, anticipate unstable loading conditions, prioritise process adjustments, and support a transition from compliance-driven treatment to resource-aware plant operation.

Chemical oxygen demand (COD) as target parameter

FFNN vs. RNN for COD as a parameter

Considering COD as a target wastewater valorisation variable, we employed FFNN, RNN, and XGBoost as the primary modelling comparators, with CNN results incorporated into the cross-model performance summary in Section 4.2. This sub-section compares FFNNs and RNNs to isolate the contribution of temporal dependency modelling. The training vs. validation loss curves for FFNN and RNN are depicted in Supplementary Figures 1a,b, respectively.

Both FFNN and RNN exhibit a similar sharp drop in loss, within 30–50 epochs for FFNN and within 10–20 epochs for RNN, respectively. This means that both models learn the dominant COD patterns well. The validation and training loss curves overlap closely for both the RNN and the FFNN. This not only confirms that there is no severe overfitting, but also that the regularisation and dataset sizes are adequate. It is also evident that the RNN curve reaches a lower, flatter validation loss and stabilises slightly more smoothly than the FFNN, which may be data-specific.

The actual vs. predicted plots for the FFNN and RNN using COD as a parameter are shown in Figures 1a,b. The R 2 values for the FFNN and RNN are 0.65 and 0.71, respectively, which are reasonable for real data. The MSE and RMSE values for FFNN are 6,810.41 (mg/L) 2 and 82.63 mg/L, while the same for RNN are 5,707.73 (mg/L) 2 and 75.55 mg/L. The FFNN actual vs. predicted plot (Figure 1a) follows a general linear ca. 45-degree profile, and the model captures the overall COD pattern. At higher COD levels, several outliers indicate underprediction in the high-value (>1,000 mg/L) regime. On the other hand, the RNN plot (Figure 1b) shows a better profile, with points much closer along the diagonal. Unlike the FFNN case, the outliers are less pronounced, show better alignment at higher COD values, and exhibit reduced dispersion of data points. Overall, RNNs can capture temporal dependencies marginally better for COD events.

Figures 1c,d represent the residual plots for FFNN and RNN, respectively. Both models show randomised residual plots and hence are fair predictors. The FFNN residual plot (Figure 1c) depicts a random cloud of points mostly structured around the zero line. There are several outliers at higher predicted COD values, with slight heteroscedasticity (i.e., the variance increases with magnitude), as expected in systems with non-linear kinetics. In real-world scenarios, errors are higher during high-load conditions, which is quite common in wastewater valorisation. The RNN residual plot (Figure 1d) is more symmetrically distributed with a little tighter clustering around zero, although some outliers are visible at higher COD values. We can conclude that RNN reduces systematic bias, while predicting at higher COD values remains challenging.

The trendlines obtained from the different phases of the FFNN, with COD as the parameter, are shown in Figure 2. The regression analysis suggests that the model successfully translates the data to a validation point, with a regression coefficient (R 2) of 0.54, while the same for training, test, and overall is 0.73, 0.72, and 0.71, respectively. This highlights the algorithm’s efficiency, which captures ∼70% of the dataset’s dimensionality and reproduces it. Visually, the outliers indicate the model’s underperformance, which is corroborated by the lower R 2 values. The error histogram (Figure 3a) shows a bell-shaped curve centred near zero, with a moderate spread of ±150–200 mg/L. The slight tails at both ends indicate some extreme under- or overpredictions, as identified previously.

The trendlines obtained from the different phases of the RNN, with COD as the parameter, are depicted in the (Supplementary Figure S2). The regression analysis suggests that the model can translate the data to a validation point with a regression coefficient (R 2) of 0.47, while the same for training, test, and overall is 0.72, 0.70, and 0.69, respectively. This highlights the algorithm’s efficiency, which captures ca. 70% of the dataset’s dimensionality and reproduces it. Visually, the outliers represent the underperformance of the model, which is in corroboration with the lower R 2 values. The error histogram (Figure 3b) shows a slightly skewed bell-shaped curve with tighter central clustering than in the FFNN. The distribution is still centred close to zero but with a mild right skewness. This highlights that the system is free of bias but susceptible to deviations from its equilibrium with moderate to strong perturbations, as identified in previous studies (Ng et al., 2025).

XGBoost for COD as a parameter

The actual vs. predicted plot for XGBoost using COD as a parameter is shown in Figure 4a, where we observe linear growth. Although there is slightly more dispersion of datapoints than in RNN, we observe noticeable clustering around mid-COD values of 800–950 mg/L. We also find that at very high COD values (i.e., greater than 1,100 mg/L), there are some underpredictions, which is quite typical for tree-based models when extreme values are sparse. The R 2 value for XGBoost is 0.71, as shown in these visuals. The residual plot is highlighted in Figure 4b. We observe that there are no strong biases present as the residuals are centred around the zero line. There is a mild funnel shape in the residual plot, indicating that, as predicted, COD values increase, the residual spread widens. There is a slight asymmetry in the high-value region. This means the model is more accurate at low-to-moderate COD levels. This is known as heteroscedasticity, and it is quite common in wastewater systems demonstrating shock loads, industrial discharges, and hydraulic fluctuations.

Figure 5 depicts the regression performance of XGBoost with COD as a parameter across the training, validation, test, and overall datasets. The regression analysis suggests that the model can translate the data to a validation point with a regression coefficient (R 2) of 0.73, while the same for training, test, and overall process are 0.99, 0.72, and 0.95, respectively. XGBoost fits the training data with an R 2 value close to 1. This specifies high model capacity and potential risk of overfitting, particularly with structured datasets. The lower R 2 value in validation, a significant drop from the training values, indicates slight underprediction at higher COD values. This confirms mild overfitting. The high, near-perfect training coefficient may inflate the overall regression coefficient. Visually, the outliers are fewer than in the FFNN and RNN cases, indicating the model’s underperformance, which is corroborated by the lower R 2 values. The error histogram (Figure 3c) is mostly concentrated around zero with a surprisingly sharp central peak. The spread is much narrower compared to NN’s. XGBoost produces more consistent errors with fewer large deviations across all the ML and DL algorithms used.

Total nitrogen as target parameter

FFNN vs. RNN for total nitrogen as a parameter

Considering total nitrogen as the target variable, we used a Feedforward Neural Network (FFNN) and a Recurrent Neural Network (RNN). Training vs. validation loss for both FFNN and RNN, are depicted in Supplementary Figures 1c,d. Both FFNN and RNN exhibit sharp drops: within 20–40 epochs for FFNN and within 10–20 epochs for RNN. This means that both models conveniently learn the dominant patterns for total nitrogen while the signal-to-noise ratio for total nitrogen data is reasonably strong. As shown in Figure 6 and Supplementary Figure 10, the validation and training loss curves overlap closely for both the RNN and the FFNN.

The actual vs. predicted plots for the FFNN and RNN using total nitrogen as a parameter are shown in Figures 7a,b. The R 2 values for FFNN and RNN are 0.54 and 0.51, respectively. The MSE and RMSE values for FFNN are 6.80 (mg/L) 2 and 2.61 mg/L, while the same for RNN are 7.25 (mg/L) 2 and 2.69 mg/L. The FFNN actual vs. predicted plot (Figure 6a) follows a positive linear trend. The RNN plot (Figure 7b) also exhibits a similar trend with mildly tighter clustering in the mid-range. Predictions are comparable between FFNN and RNN, as indicated by the R 2 values. The residual plots in Figures 3c, 7d show that the residuals are centred around the zero line, indicating no strong systematic biases. The analysis confirms that total nitrogen dynamics are likely to be influenced by missing process state variables, such as Dissolved Oxygen (DO), Solid Retention Time (SRT), and temperature. It should be noted that the RNN failed to improve prediction, suggesting that total nitrogen variability might be more inherent to nonlinear feature interactions than to sequence memory.

The trendlines for the different phases of the FFNN, with Total Nitrogen as the parameter, are shown in Figure 8. The regression analysis suggests that the model translates the data to a validation point with a regression coefficient (R 2) of 0.44, while the same values for training, test, and overall process are 0.62, 0.59, and 0.60, respectively. This highlights the algorithm’s efficiency, which captures ca. 60% of the dataset’s dimensionality and reproduces it. Visually, the outliers indicate the model’s underperformance, which aligns with the lower R 2 values. The error histogram (Figure 3c) shows a roughly bell-shaped curve centred near zero, with a moderate spread of ±5–10 mg/L. This indicates no strong systematic bias.

The trendlines obtained from the different phases of the RNN, with total nitrogen as the parameter, are depicted in Supplementary Figure S3. The regression analysis suggests that the model can barely translate the data to a validation point, with a regression coefficient (R 2) of 0.59, while the same for training, test, and overall is 0.61, 0.58, and 0.60, respectively. This highlights that the algorithm captures ca. 60% of the dataset’s dimensionality and reproduces it. Visually, the outliers represent the underperformance of the model, which is in corroboration with the lower R 2 values. The error histogram (Figure 3d) shows a bell-shaped curve with slightly tighter central clustering than for the FFNN. The distribution remains symmetrical around zero, and the extreme tails are narrower. RNN exhibits slightly narrower error dispersion than FFNN, marginally improving the consistency of the prediction errors. However, the overall predictive performance remains comparable, highlighting a slightly more uniform spread. Therefore, this temporal model adds little value for total nitrogen.

XGBoost for total nitrogen as a parameter

The actual vs. predicted plot for XGBoost using total nitrogen as a parameter is shown in Figure 4c, where we observe a clear, positive linear trend. We observe that the points are more tightly clustered than in FFNN and RNN. There is better alignment of the data points around the mid-total nitrogen values, ranging from 58 to 65 mg/L. We also find mild dispersion at lower total nitrogen levels, whereas there is a slight underprediction at higher levels. The R 2 value for XGBoost is 0.62, indicating it improves total nitrogen prediction over RNN and FFNN, as it captures non-linear feature interactions better than temporal dependencies. The residual plot is highlighted in Figure 4d. We observe that there are no strong biases present as the residuals are centred around the zero line. There is slight clustering around the predicted 60–65 mg/L, however.

Figure 7 depicts the regression performance of XGBoost with total nitrogen as a parameter across the training, validation, test, and overall datasets. The regression analysis suggests that the model can translate the data to a validation point with a regression coefficient (R 2) of 0.60, while the same for training, test, and overall process are 0.99, 0.64, and 0.93, respectively. XGBoost fits the training data with an R 2 value approaching unity. This indicates high model flexibility and highlights clear evidence of overfitting, particularly with structured datasets. The lower R 2 indicates that the training fit does not fully translate the unseen data. The test plot shows that the model generalises the data moderately, while highlighting only moderate overfitting. As it appears, the outliers are fewer than in the FFNN and RNN cases, indicating improved robustness and lower prediction variance, particularly in handling nonlinear operational variability. The error histogram (Figure 3f) is concentrated around zero with an unusually sharp central peak. The spread is much narrower than that of the NN architectures. XGBoost produces the most stable and consistent predictions for total nitrogen, which aligns with the R 2 value of 0.62.

Sustainability or green performance of the machine learning (ML) models

Computational advancement, particularly artificial intelligence (AI), has opened unprecedented avenues for delving into granular processes, kinetics, systems, and more. However, these come with a price, the price of environmental sustainability. With the help of ‘ codecarbon ’, a specific Python library that calculates the sustainability metrics, including energy consumption, electricity, water consumption and carbon dioxide emissions (Debnath et al., 2025). The overall findings for both COD and Total Nitrogen are detailed in Table 1 below.

The evaluation of sustainability metrics highlights details relating computational proficiency to environmental sustainability. XGBoost successfully identifies the minimum-energy regime and produces the minimum carbon emissions, while the RNN maps to the regime with the highest computational energy requirement. Clearly, greater model complexity entails a larger carbon footprint. Among the neural networks, FFNN consumes the least energy, followed by CNN and RNN. When compared with models using COD and TN as parameters, the COD models consume less energy than the Total Nitrogen models. One key reason is that these evaluations are purely data-driven and do not incorporate process kinetics. This also means that state process variables are missing. RAM consumption varies significantly across architectures, corroborating the energy consumption portfolio. The CO 2 emissions are directly proportional to electricity usage. Similarly, the water footprint also increases with computational intensity, as evidenced by the least for XGBoost and the most for RNN. From an environmental sustainability perspective, boosting models outperform NNs while achieving only marginal accuracy at a higher environmental cost. Note, the absolute values of the environmental metrics are small and may appear negligible, but they reveal key differences between the implemented ML models. The central objective here is not to quantify a single model training cycle; rather, it is to provide a relative basis for sustainability-aware model selection, towards which a combined bootstrapping approach may be the way forward, the first step of which is outlines in this study. Since ML models are trained and deployed repeatedly across different facilities, cumulative environmental burdens can become significant. This reinforces the strategy of achieving high predictive accuracy with higher sustainability metrics.

The proposed performance metrics indicate that environmental burden (ECI = 1.62–1.68) is highest for RNN, while FFNN and CNN are more optimal. As expected, XGBoost exhibits the best balance between sustainability footprint and predictive efficiency withthe highest GPI (0.63–0.70) and TGS (0.64–0.72) values. Beyond the traditional comparison of efficiency and cost, the developed framework evaluates model suitability by considering prediction performance, sustainability metrics and model generalisation. This enables the deployment of sustainability-aware decision-making. Recent studies have evaluated the sustainability aspects of ML models, including energy consumption and carbon-centric metrics (Castellanos-Nieves and García-Forte, 2024; Hasan et al., 2024). The present study transcends these approaches by unifying environmental costs with predictive accuracy in a generic way, creating a decision-support perspective for wastewater resource recovery applications. These findings highlight the Green ML paradigm, in which model selection is guided by both accuracy and computational sustainability.

Performance evaluation

The overall evaluation of the various ML and DL (AI) models used in our analysis is comprehensively presented in Table 2.

RNN captures COD temporal dynamics in the best possible manner, while FFNN shows stable but smoother predictions. XGBoost is the most efficient algorithm with nonlinear feature interactions and exhibits fewer extreme residual outliers. While the training R 2 ∼ 1 for XGBoost indicates an overfitting tendency, it is also a critical thing to consider while choosing an AI model. Our results suggest realistic validation and test R 2 scores within the limit of 0.65–0.70. Total nitrogen is inherently harder to predict, as process kinetic variables were not incorporated. COD is more predictable than multi-stage nitrogen removal, and public datasets notably lack SRT, DO, and explicit kinetic parameters. The moderate R 2 range observed is therefore realistic and expected for operational wastewater data. Among the architectures evaluated, CNN showed limited added value over FFNN, confirming that local convolutional filtering does not benefit low-dimensional tabular WWTP data. The tabular data structure consistently favours boosted tree ensembles over neural architectures. It can be inferred that efficient ML models can significantly reduce the operational carbon footprint associated with AI-assisted wastewater analytics; hence, Green ML architectures should be selected based on the trade-off between predictive accuracy and computational sustainability before field deployment.

Combined, the comparative results show that no single architecture is universally optimal unless predictive performance is considered alongside overfitting behaviour and computational cost. For COD, XGBoost and RNN deliver the strongest predictive profiles, but XGBoost does so with a markedly lower environmental burden. For TN, the tree-based model is the front-runner, indicating that nonlinear feature interactions are more informative than sequential memory for this target under the available public data. CNNs do not offer a meaningful advantage over FFNNs, suggesting that local convolutional filtering adds limited value to this low-dimensional tabular problem. The moderate R 2 values observed across all models are realistic for operational WWTP datasets, where missing kinetic variables, sensor noise, and plant variability constrain predictive ceilings. A later outcome of this study is a comparative decision framework showing when additional model complexity is justified and when it is not.

The near-unity training R 2 values achieved by XGBoost indicate that the model effectively captures the relationships present in the training dataset. However, the lower validation and test R 2 values suggest that some of these relationships are less consistent in unseen observations. This indicates that, although the seven selected predictors capture substantial predictive information, they do not fully represent the complex process variability governing COD and Total Nitrogen under real-world WWTP operating conditions. Wastewater treatment performance is affected by additional process-state variables, biological dynamics, operational fluctuations, and measurement uncertainty, which are often industrial IP-protected coordinates and hence difficult to measure and thus not provided in the given dataset. Therefore, the validation and test results provide the most realistic assessment of predictive capability under practical operating conditions that may differ from specific outlets. These findings demonstrate that apparent gains in training performance do not necessarily translate into improved generalisation, underscoring the need to prioritise predictive performance on unseen data when making deployment decisions.

This study developed and tested a Green ML benchmarking framework for wastewater valorisation using a publicly available Melbourne WWTP dataset and two strategically selected prediction targets, COD and Total Nitrogen. The principal novelty lies in combining predictive accuracy with environmental footprint, that is the Green ML corridor, in the process identifying the SPE vs. energy-efficient algorithm across 4 ML and DL models: FFNN, RNN, CNN, and XGBoost. For simpler kinetic pathways, XGBoost emerged as the most optimised algorithm, followed by RNN for more complex multidimensional kinetics across both COD and TN pathways. This comparative design addresses a clear and persistent gap in the wastewater ML literature, where single-model studies still dominate, and sustainability-aware model selection criteria remain largely absent. A related novelty is in the outlining of a bootstrapping mechanism which we have not incorporated explicitly, that being beyond the scope of this study, rather laying a foundation that can combine performances of multiple ML and DL algorithms to present holistic scores that remain technically unbiased to the user interface.

From a predictive performance standpoint, architecture choice matters, but its value is target-dependent. For COD, both RNN and XGBoost outperform the other architectures, with RNN benefiting from temporal dependency modelling and XGBoost exploiting nonlinear feature interactions. However, XGBoost achieves comparable or superior accuracy while requiring substantially less computational resources, making it especially attractive for deployment in resource-constrained or sustainability-conscious operational settings. For Total Nitrogen, XGBoost is the most reliable model across all metrics, while neural architectures converge to moderate, closely grouped R 2 values. This outcome indicates that TN variability in the present dataset is governed primarily by latent biochemical process-state variables absent from the public record, most notably dissolved oxygen and sludge retention time, rather than by temporal sequence structure. CNN offers no clear benefit over FFNN, confirming that local convolutional filtering adds limited value for low-dimensional, tabular wastewater data. The moderate R 2 range of 0.54–0.71 observed across all models is consistent with established benchmarks for operational WWTP datasets and reflects inherent noise and missing kinetic variables, rather than model inadequacy.

The Green ML sustainability audit yields the study’s most striking quantitative finding. XGBoost consumes two to three orders of magnitude less energy than any neural architecture evaluated, with CO 2 equivalent emissions approximately 65–100 times lower than RNN across both target variables. This dramatic difference in computational footprint occurs despite XGBoost delivering the most stable and consistent predictions for both COD and TN. The practical implication is unambiguous: in wastewater analytics contexts where XGBoost achieves comparable or superior predictive quality, selecting it over deep-learning alternatives reduces the algorithmic carbon footprint of routine WWTP analytics by more than 98%. The conventional instinct towards architectural complexity is therefore demonstrably unjustified on both performance and sustainability grounds for structured tabular wastewater data. The framework proposed here reframes AI model selection for environmental engineering as a multi-criteria decision problem in which algorithmic greenness is a core technical consideration, not an afterthought.

From a wastewater valorisation perspective, the study reinforces the strategic importance of predicting COD and TN as key valorisation variables rather than treating them entirely as compliance parameters. Accurate anticipation of COD levels can directly support carbon recovery pathways, including anaerobic digestion, VFA generation, and biopolymer production. However, improved TN prediction can inform nitrogen recovery strategies linked to struvite crystallisation and circular fertiliser production. This study, therefore, positions ML not only as a monitoring instrument but also as a decision-support architecture for circular wastewater system design and monitoring. The use of an open-access dataset further ensures transparency and reproducibility, making the comparative framework transferable to future benchmarking studies across diverse plant types, geographic contexts, and recovery pathways.

Several limitations of the present work motivate future research. The reliance on a public operational dataset means that key biochemical and process-state descriptors, dissolved oxygen, hydraulic and sludge retention times, reactor temperature profiles, and explicit kinetic parameters were unavailable. This constrains the achievable predictive ceiling, particularly for nitrogen dynamics, where multi-stage biological processes govern behaviour. Future studies may aim to embed such process-state features within the proposed evaluation framework to assess whether the XGBoost sustainability advantage persists under richer, kinetics-informed feature spaces. Physics-informed neural networks, which incorporate mechanistic kinetics as structural model constraints, represent a compelling frontier for bridging data-driven accuracy with mechanistic interpretability.

Finally, the comparative Green ML framework developed here is explicitly designed for generalisation: its systematic application to additional WWTP parameters, diverse geographic and climatic contexts, and broader circular bioeconomy recovery pathways is strongly encouraged as part of an open-science agenda for sustainable water systems management.