Deep learning in wastewater treatment: a critical review

Recognizing the state of aerobic granular sludge over its life-cycle in a continuous-flow membrane bioreactor with an artificial intelligence approach

The continuous-flow aerobic granular sludge-membrane bioreactor (AGS-MBR) system represents an efficient and sustainable technology for wastewater treatment. AGS, a spherical or ellipsoidal granular sludge formed through microbial self-aggregation under aerobic conditions, progresses through four distinct life-cycle stages in the AGS-MBR system: initial, growth, mature, and cleaved. Accurate identification and classification of these stages are crucial for optimizing AGS-MBR operations and maintaining system stability; however, traditional monitoring methods are labor-intensive and error-prone. This study utilized Artificial Intelligence (AI) to develop a machine learning model based on the You Only Look Once (YOLOv8) algorithm for automated AGS monitoring and classification. Trained on 862 annotated images, the model achieved average precision of 0.985 at an Intersection over Union (IoU) threshold of 0.5 (mAP50), and the mAP50-95 of 0.837, demonstrating high accuracy in AGS classification. The t-distributed Stochastic Neighbor Embedding (t-SNE) revealed distinct clusters of AGS features across life-cycle stages, while SHapley Additive exPlanations (SHAP) demonstrated that the model focused on global features of small-grained images and edge features of large-grained images, both confirming the robustness of classification. The model's statistical functionality, supported by global variables, enabled real-time AGS monitoring in MBR system. This study provides a powerful tool for detecting and classifying the AGS life-cycle, offering guidance for the operation and maintenance of AGS-MBR system and demonstrating the potential applications of AI in wastewater treatment.

Heterogeneous graph neural networks enhance pressure estimation in water distribution networks

Pressure estimation is crucial for efficient operation and management of water distribution networks (WDNs). However, it is often challenged by limited sensor observations. While graph neural networks (GNNs) have been used to improve hydraulic and water quality predictions of WDNs, their reliance on homogeneous graphs oversimplifies the diverse roles and interactions of hydraulic components, resulting in lower performance under dynamic system states. This research introduces a novel heterogeneous graph neural network (HGNN) framework, which models control units such as pumps and valves as distinct nodes while preserving their interactions through additional edge types. Experimental results using C-Town as a benchmark demonstrate that HGNN outperforms GNN in terms of accuracy, robustness, and adaptability, achieving a mean absolute percentage error (MAPE) of 1.88 % and a mean absolute error (MAE) of 1.70 m under a 95 % masking rate. Additionally, this study shows that optimal sensor placement reduces MAE by up to 15 %, and the proposed HGNN framework achieves high computational efficiency, highlighting its effectiveness in WDN analysis and management. This research offers an advanced and transferable approach for WDN pressure estimation, serving as a superior alternative to traditional pressure evaluation models.

Data-driven differentiable model for dynamic prediction and control in wastewater treatment

Wastewater treatment plants, while critical for environmental protection, face mounting challenges in operational efficiency and sustainability due to increasing urbanization and stricter environmental standards. In this study, we introduce an innovative continuous-time neural framework based on Neural Ordinary Differential Equations (Neural ODEs) to enhance the modeling of sewage treatment processes. Addressing the dual challenges of operational efficiency and sustainable development in urban wastewater treatment plants (WWTPs), our methodology marks a significant departure from traditional approaches by implementing a continuous-time neural framework that captures the inherent dynamics of wastewater treatment processes while reducing computational demands by 95 % compared to discrete-time models. We analyzed operational data from three full-scale WWTPs over a year, demonstrating that our model not only achieves superior prediction accuracy (R² > 0.95) with various input window sizes but also significantly reduces memory usage-from 111.88-12,484.59 MB to just 17.74-50.92 MB. Notably, our framework exhibits robust performance even with up to 30 % missing data, uncovering new process insights through interpretable feature attribution. Further integration with reinforcement learning has led to a 21.9 % reduction in aeration energy consumption compared to conventional open-loop control strategies while adhering to effluent quality standards. This research establishes a novel paradigm for intelligent wastewater management that optimizes operational efficiency and promotes environmental sustainability.

Biokinetic soft-sensing using Thiothrix and Ca. Microthrix bacteria to calibrate secondary settling, aeration and N2O emission digital twins

Climate resilience in water resource recovery facilities (WRRFs) necessitates improved adaptation to shock-loading conditions and mitigating greenhouse gas emission. Data-driven learning methods are widely utilised in soft-sensors for decision support and process optimization due to their simplicity and high predictive accuracy. However, unlike for mechanistic models, transferring machine-learning-based insights across systems is largely infeasible, which limits communication and knowledge sharing. To harness the benefits of both approaches, this study introduces a mechanistic online soft-sensor (MOSS) developed to calibrate digital twins of secondary settling tanks (hydraulic shock), aeration systems and nitrous oxide (N2O) greenhouse gas emission. MOSS integrates biokinetic models of filamentous microbial predictors to calibrate digital twins through meta-models (data-driven part), updated using offline settling column tests and amplicon sequencing data for microbial analysis. For the first time, this approach employs multi-filamentous-community predictors for dynamic calibration, i.e., Thiothrix and Ca. Microthrix. The calibration and early-warning capabilities of MOSS are demonstrated using experimental data from a laboratory-scale WRRF.

Data assimilation for prediction of ammonium in wastewater treatment plant: From physical to data driven models

This study compares various modeling approaches to predict ammonium concentration in wastewater treatment plants (WWTPs), with a focus on integrating data assimilation techniques. It explores white-box, grey-box, and black-box models, evaluating their ability to capture the complex dynamics of WWTPs and manage uncertainties associated with limited data and sensor noise. The article highlights the importance of data assimilation for simultaneously calibrating model parameters, latent variables (such as unmeasured species concentrations), and quantifying prediction uncertainty. Simulation results demonstrate that the non-parametric black box model outperforms all other models in terms of predictive accuracy and uncertainty estimation. This finding underscores the effectiveness of machine learning when integrated with data assimilation techniques to extract insights from training datasets, even in the presence of limited data. Interestingly, the addition of an extra sensor, such as an oxygen sensor, did not enhance model performance. Experiments conducted in a real system showed that the non-parametric black box model could effectively capture the general dynamics of ammonium concentration in an actual wastewater treatment plant. However, its performance was somewhat diminished compared to simulation results, likely due to variability in input concentrations that were not accounted for in the model.

Data-driven fault detection and diagnosis methods in wastewater treatment systems: A comprehensive review

Wastewater treatment systems are essential for sustainable water resource management but face challenges such as equipment and sensor malfunctions, fluctuating influent conditions, and operational disturbances that compromise process stability and detection accuracy. To address these challenges, this paper systematically reviews data-driven fault detection and diagnosis (FDD) methods applied in wastewater treatment systems from 2014 to 2024, focusing on their applications, advancements, and limitations. Main contributions include an overview of key treatment processes, a detailed evaluation of fault types (process and sensor faults), advancements in multivariate statistical methods, machine learning (ML), and hybrid FDD techniques, as well as their effectiveness in anomaly detection, managing complex data distributions, and enabling real-time monitoring. Furthermore, the paper highlights critical challenges such as data quality and model interpretability, proposing actionable future directions, including the development of explainable artificial intelligence, adaptive real-time processing, and cross-system generalizability. These insights are intended to guide the development of robust, scalable, and interpretable FDD solutions, ultimately improving the efficiency, reliability, and sustainability of wastewater treatment systems.

Wastewater quality prediction based on channel attention and TCN-BiGRU model

Water quality prediction is crucial for water resource management, as accurate forecasting can help identify potential issues in advance and provide a scientific basis for sustainable management. To predict key water quality indicators, including chemical oxygen demand (COD), suspended solids (SS), total phosphorus (TP), pH, total nitrogen (TN), and ammonia nitrogen (NH₃-N), we propose a novel model, CA-TCN-BiGRU, which combines channel attention mechanisms with temporal convolutional networks (TCN) and bidirectional gated recurrent units (BiGRU). The model, which uses a multi-input multi-output (MIMO) architecture, is capable of simultaneously predicting multiple water quality indicators. It is trained and tested using data from a wastewater treatment plant in Huizhou, China. This study investigates the impact of data preprocessing and the channel attention mechanism on model performance and compares the predictive capabilities of various deep learning models. The results demonstrate that data preprocessing significantly improves prediction accuracy, while the channel attention mechanism enhances the model's focus on key features. The CA-TCN-BiGRU model outperforms others in predicting multiple water quality indicators, particularly for COD, TP, and SS, where MAE and RMSE decrease by approximately 23% and 26%, respectively, and R2 improves by 5.85%. Moreover, the model shows strong robustness and real-time performance across different wastewater treatment plants, making it suitable for short-term (1-3 days) water quality prediction. The study concludes that the CA-TCN-BiGRU model not only achieves high accuracy but also offers low computational overhead and fast inference speed, making it an ideal solution for real-time water quality monitoring.

Semi-supervised learning-based identification of the attachment between sludge and microparticles in wastewater treatment

Monitoring the microparticle transfer process in wastewater treatment systems is crucial for improving treatment performance. Supervised deep learning methods show high performance to automatically detect particles, but they rely on vast amounts of labeled data for training. To overcome this issue, we proposed a semi-supervised learning (SSL) method based on the Simple framework for Contrastive Learning of visual Representations (SimCLR), to detect microparticles free from sludge and attached to sludge. First, we pre-trained a ResNet50 backbone by SimCLR, to extract features from much unlabeled data (1,000 images). Then, we constructed a Mask R-CNN architecture based on the pre-trained ResNet50, and fine-tuned it on a small quantity of labeled data (≈200 images with ≈600 annotated particles) in supervised learning fashion. We showcased its performance and practical applicability for microscopy images obtained from the water lab of TU Delft. The results demonstrate that the SSL methods obtain a significant improvement in mean average precision of up to 5% compared to the conventional supervised learning method, when a limited amount of labeled data is available (e.g., 91 labeled images). Furthermore, these methods improve the average precision for detecting attached particles by over 12%. With the detection results from the SSL methods, we measured the attachment efficiency of microparticles to sludge under varying mixed liquor suspended solids concentration and aeration intensity. The precise measurements demonstrate the effectiveness and practical applicability of the SSL method in facilitating long-term monitoring of particle transfer processes in biological wastewater treatment systems.

A probabilistic deep learning approach to enhance the prediction of wastewater treatment plant effluent quality under shocking load events

Sudden shocking load events featuring significant increases in inflow quantities or concentrations of wastewater treatment plants (WWTPs), are a major threat to the attainment of treated effluents to discharge quality standards. To aid in real-time decision-making for stable WWTP operations, this study developed a probabilistic deep learning model that comprises encoder-decoder long short-term memory (LSTM) networks with added capacity of producing probability predictions, to enhance the robustness of real-time WWTP effluent quality prediction under such events. The developed probabilistic encoder-decoder LSTM (P-ED-LSTM) model was tested in an actual WWTP, where bihourly effluent quality prediction of total nitrogen was performed and compared with classical deep learning models, including LSTM, gated recurrent unit (GRU) and Transformer. It was found that under shocking load events, the P-ED-LSTM could achieve a 49.7% improvement in prediction accuracy for bihourly real-time predictions of effluent concentration compared to the LSTM, GRU, and Transformer. A higher quantile of the probability data from the P-ED-LSTM model output, indicated a prediction value more approximate to real effluent quality. The P-ED-LSTM model also exhibited higher predictive power for the next multiple time steps with shocking load scenarios. It captured approximately 90% of the actual over-limit discharges up to 6 hours ahead, significantly outperforming other deep learning models. Therefore, the P-ED-LSTM model, with its robust adaptability to significant fluctuations, has the potential for broader applications across WWTPs with different processes, as well as providing strategies for wastewater system regulation under emergency conditions.

A calibration framework toward model generalization for bacteria concentration estimation in water resource recovery facilities

Reduced bacteria concentrations in wastewater is a key indicator of the efficacy of water resource recovery facilities (WRRFs). However, monitoring the presence of bacterial concentrations in real time at each stage of the WRRF is challenging as it requires taking and processing water samples offline. Although few studies have been proposed to predict bacterial concentrations using data-driven models, generalizing these models to unseen data from different WRRFs remains challenging. This paper proposes a calibration approach based on neural networks to adapt the optimal models across various WRRFs in Saudi Arabia for bacterial estimation at the influent and effluent stages. The calibration relies on the out-of-distribution (OOD) framework of the physiochemical water parameters (e.g., pH, COD, TDS, turbidity, conductivity) with a design threshold chosen based on the data distribution of the received unseen samples. We propose a calibration framework that continues updating the trained neural network model for accurate bacterial concentration estimation upon receiving new samples. We tested the effectiveness of the proposed calibration scheme on four WRRF datasets in Saudi Arabia, comparing the results with before and after calibration without the OOD. Before calibration model was based on a traditional and optimal neural network approach, typically considered the conventional method for building neural networks. After calibration without OOD, the model continued retraining without explicitly checking for OOD condition. The results showed that the proposed calibration framework of the selected baseline WRRF with the OOD scheme improved [Formula: see text] and [Formula: see text] of the worst-case influent bacteria concentration before calibration and after calibration without OOD, respectively. Similarly, the worst-case effluent bacteria concentration estimation was enhanced by [Formula: see text] before calibration and [Formula: see text] after calibration without the OOD. Our findings highlight the importance of integrating the calibration framework with neural network approaches to achieve model generalization.

Machine learning revealing overlooked conjunction of working volume and mixing intensity in anammox optimization

Extensive studies on improving anammox performance have taken place for decades with particular focuses on its operational and environmental factors, but such parameter-based optimization is difficult, because of the sheer number of possible combinations and multidimensional arrays of these factors. Utilizing machine-learning algorithm and published anammox data, Bayesian nonparametric general regression (BNGR) was applied to identify the possible governing variable(s) from among 11 operating and environmental parameters: reactor type, mixing type, working volume, hydraulic retention time, temperature, influent pH, nitrite, ammonium, nitrate concentration, nitrogen loading rate, and organic concentration. The results showed that working volume is a key but oft-overlooked governing parameter. By integrating the BNGR findings with computational fluid dynamics simulation, which assessed mixing properties, it became feasible to conclude that working volume and mixing intensity co-regulated flow fields in reactors and had a significant influence on anammox performance. Furthermore, this study experimentally validated how mixing intensity affected performance, and specific input power (x), a parameter that conjugates both working volume and mixing intensity, was used to establish the relationship with ammonium removal rate (NH4+-N RR, y) y = 49.90x+1.97 (R2 = 0.94). With specific input power increased from 3.4 × 10-4 to 2.6 × 10-2 kW m-3, the ammonium removal rate exhibited a rise from 1.8 to 3.2 mg L-1h-1. Following, a relationship among input power-working volume-nitrogen removal rate was also established with a view to determining the design variables for anammox reactor. Consequently, the study highlighted the necessity to consider the working volume-mixing intensity correlation when optimizing the anammox process.

Enhanced biodegradation of glyphosate by Chlorella sorokiniana engineered with exogenous purple acid phosphatase

Organophosphate pesticides, particularly glyphosate, persist in aquatic environments due to widespread agricultural usage, posing substantial environmental and health risks. This study explores the bioremediation potential of genetically engineered Chlorella sorokiniana, expressing purple acid phosphatase (PAP) from Phaeodactylum tricornutum, for glyphosate biodegradation. The engineered strain (OE line) demonstrated complete glyphosate biodegradation at concentrations below 10 ppm within 4-6 days, surpassing the wild type (WT). Enhanced biodegradation in the OE line was attributed to increased growth and ATP levels due to the release of inorganic phosphate, indicating enhanced metabolic efficiency. Photosynthetic parameters, as well as chlorophyll, and carotenoid contents, were significantly improved, driving higher biomass accumulation. Metabolic shifts toward lipogenesis were observed, supported by the upregulation of triacylglycerol-related genes. Additionally, antioxidant enzyme activities (GPx, SOD, CAT) were elevated in the OE line, mitigating oxidative stress. Importantly, the overexpression of PAP activated and upregulated the level of endogenous CsPAP18, which displayed stable binding with glyphosate and its metabolite aminomethylphosphonic acid, highlighting the synergistic role of PAP and CsPAP18 in glyphosate biodegradation. The OE line effectively treated glyphosate-contaminated real wastewater, confirming the feasibility of engineered strain for environmental remediation. This study provides valuable insights into the potential of engineered microalgae for effective and sustainable wastewater treatment, specifically targeting the removal of organophosphate contaminants in freshwater environments.

A novel approach for multivariate time series interval prediction of water quality at wastewater treatment plants

This study proposes a novel approach for predicting variations in water quality at wastewater treatment plants (WWTPs), which is crucial for optimizing process management and pollution control. The model combines convolutional bi-directional gated recursive units (CBGRUs) with adaptive bandwidth kernel function density estimation (ABKDE) to address the challenge of multivariate time series interval prediction of WWTP water quality. Initially, wavelet transform (WT) was employed to smooth the water quality data, reducing noise and fluctuations. Linear correlation coefficient (CC) and non-linear mutual information (MI) techniques were then utilized to select input variables. The CBGRU model was applied to capture temporal correlations in the time series, integrating the Multiple Heads of Attention (MHA) mechanism to enhance the model's ability to comprehend complex relationships within the data. ABKDE was employed, supplemented by bootstrap to establish upper and lower bounds of the prediction intervals. Ablation experiments and comparative analyses with benchmark models confirmed the superior performance of the model in point prediction, interval prediction, the analysis of forecast period, and fluctuation detection for water quality data. Also, this study verifies the model's broad applicability and robustness to anomalous data. This study contributes significantly to improved effluent treatment efficiency and water quality control in WWTPs.

Perspectives on Advancing Multimodal Learning in Environmental Science and Engineering Studies

The environment faces increasing anthropogenic impacts, resulting in a rapid increase in environmental issues that undermine the natural capital essential for human wellbeing. These issues are complex and often influenced by various factors represented by data with different modalities. While machine learning (ML) provides data-driven tools for addressing the environmental issues, the current ML models in environmental science and engineering (ES&E) often neglect the utilization of multimodal data. With the advancement in deep learning, multimodal learning (MML) holds promise for comprehensive descriptions of the environmental issues by harnessing data from diverse modalities. This advancement has the potential to significantly elevate the accuracy and robustness of prediction models in ES&E studies, providing enhanced solutions for various environmental modeling tasks. This perspective summarizes MML methodologies and proposes potential applications of MML models in ES&E studies, including environmental quality assessment, prediction of chemical hazards, and optimization of pollution control techniques. Additionally, we discuss the challenges associated with implementing MML in ES&E and propose future research directions in this domain.

Application of machine learning for antibiotic resistance in water and wastewater: A systematic review

Antibiotic resistance (AR) is considered one of the greatest global threats in the current century, which can only be overcome if all interconnected areas of humans, animals and the environment are taken into account as part of the One Health concept proposed by the World Health Organization (WHO). Water and wastewater are among the most important environmental media of AR sources, where the phenomena are generally non-linear. Therefore, the aim of this study was to investigate the application of machine learning-based methods (MLMs) to solve AR-induced problems in water and wastewater. For this purpose, most relevant databases were searched in the period between 1987 and 2023 to systematically analyze and categorize the applications. Accordingly, the results showed that out of 12 applications, 11 (91.6%) were for shallow learning and 1 (8.3%) for deep learning. In shallow learning category, n = 6, 50% of the applications were regression and n = 4, 33.3% were classification, mainly using artificial neural networks, decision trees and Bayesian methods for the following objectives: Predicting the survival of antibiotic-resistant bacteria (ARB), determining the order of influencing parameters on AR-based scores, and identifying the major sources of antibiotic resistance genes (ARGs). In addition, only one study (8.3%) was found for clustering and no study for association. Surprisingly, deep learning had been used in only one study (8.3%) to predict ARGs sequences. Therefore, working on the knowledge gaps of AR, especially using clustering, association and deep learning methods, would be a promising option to analyze more aspects of the related problems. However, there is still a long way to go to consider and apply MLMs as unique approaches to study different aspects of AR in water and wastewater.

Spatiotemporal estimation of groundwater and surface water conditions by integrating deep learning and physics-based watershed models

The impacts of climate change on hydrology underscore the urgency of understanding watershed hydrological patterns for sustainable water resource management. The conventional physics-based fully distributed hydrological models are limited due to computational demands, particularly in the case of large-scale watersheds. Deep learning (DL) offers a promising solution for handling large datasets and extracting intricate data relationships. Here, we propose a DL modeling framework, incorporating convolutional neural networks (CNNs) to efficiently replicate physics-based model outputs at high spatial resolution. The goal was to estimate groundwater head and surface water depth in the Sabgyo Stream Watershed, South Korea. The model datasets consisted of input variables, including elevation, land cover, soil type, evapotranspiration, rainfall, and initial hydrological conditions. The initial conditions and target data were obtained from the fully distributed hydrological model HydroGeoSphere (HGS), whereas the other inputs were actual measurements in the field. By optimizing the training sample size, input design, CNN structure, and hyperparameters, we found that CNNs with residual architectures (ResNets) yielded superior performance. The optimal DL model reduces computation time by 45 times compared to the HGS model for monthly hydrological estimations over five years (RMSE 2.35 and 0.29 m for groundwater and surface water, respectively). In addition, our DL framework explored the predictive capabilities of hydrological responses to future climate scenarios. Although the proposed model is cost-effective for hydrological simulations, further enhancements are needed to improve the accuracy of long-term predictions. Ultimately, the proposed DL framework has the potential to facilitate decision-making, particularly in large-scale and complex watersheds.

Applying machine learning to anaerobic fermentation of waste sludge using two targeted modeling strategies

Anaerobic fermentation is an effective method to harvest volatile fatty acids (VFAs) from waste activated sludge (WAS). Accurately predicting and optimizing VFAs production is crucial for anaerobic fermentation engineering. In this study, we developed machine learning models using two innovative strategies to precisely predict the daily yield of VFAs in a laboratory anaerobic fermenter. Strategy-1 focuses on model interpretability to comprehend the influence of variables of interest on VFAs production, while Strategy-2 takes into account the cost of variable acquisition, making it more suitable for practical applications in prediction and optimization. The results showed that Support Vector Regression emerged as the most effective model in this study, with testing R2 values of 0.949 and 0.939 for the two strategies, respectively. We conducted feature importance analysis to identify the critical factors that influence VFAs production. Detailed explanations were provided using partial dependence plots and Shepley Additive Explanations analyses. To optimize VFAs production, we integrated the developed model with optimization algorithms, resulting in a maximum yield of 2997.282 mg/L. This value was 45.2 % higher than the average VFAs level in the operated fermenter. Our study offers valuable insights for predicting and optimizing VFAs production in sludge anaerobic fermentation, and it facilitates engineering practice in VFAs harvesting from WAS.

Deep learning in water protection of resources, environment, and ecology: achievement and challenges

The breathtaking economic development put a heavy toll on ecology, especially on water pollution. Efficient water resource management has a long-term influence on the sustainable development of the economy and society. Economic development and ecology preservation are tangled together, and the growth of one is not possible without the other. Deep learning (DL) is ubiquitous in autonomous driving, medical imaging, speech recognition, etc. The spectacular success of deep learning comes from its power of richer representation of data. In view of the bright prospects of DL, this review comprehensively focuses on the development of DL applications in water resources management, water environment protection, and water ecology. First, the concept and modeling steps of DL are briefly introduced, including data preparation, algorithm selection, and model evaluation. Finally, the advantages and disadvantages of commonly used algorithms are analyzed according to their structures and mechanisms, and recommendations on the selection of DL algorithms for different studies, as well as prospects for the application and development of DL in water science are proposed. This review provides references for solving a wider range of water-related problems and brings further insights into the intelligent development of water science.