ABSTRACT
Artificial intelligence (AI) is increasingly being applied to wastewater treatment to enhance efficiency, improve processes, and optimize resource utilization. This review focuses on objectives, advantages, outputs, and major findings of various AI models in the three key aspects: the prediction of removal efficiency for both organic and inorganic pollutants, real-time monitoring of essential water quality parameters (such as pH, COD, BOD, turbidity, TDS, and conductivity), and fault detection in the processes and equipment integral to wastewater treatment. The prediction accuracy (R2 value) of AI technologies for pollutant removal has been reported to vary between 0.64 and 1.00. A critical aspect explored in this review is the cost-effectiveness of implementing AI systems in wastewater treatment. Numerous countries and municipalities are actively engaging in pilot projects and demonstrations to assess the feasibility and effectiveness of AI applications in wastewater treatment. Notably, the review highlights successful outcomes from these initiatives across diverse geographical contexts, showcasing the adaptability and positive impact of AI in revolutionizing wastewater treatment on a global scale. Further, insights on the ethical considerations and potential future directions for the use of AI in wastewater treatment plants have also been provided.
HIGHLIGHTS
Artificial intelligence (AI) has immense potential to optimize the wastewater treatment processes.
Prediction, real-time monitoring, and fault detection using AI have been reviewed.
Cost-effectiveness of implementing AI in wastewater treatment has been discussed.
Adaptability and impact of AI in wastewater treatment globally has been highlighted.
INTRODUCTION
Wastewater treatment plays an important role in protecting public health and the environment by reducing the impact of pollutants discharged into water bodies. The challenges faced by wastewater treatment plants (WWTPs) are becoming increasingly complex due to rapidly increasing urbanization and industrialization (Ullah et al. 2020). The current traditional methods, while effective to a certain extent, face limitations in addressing emerging challenges effectively (Rout et al. 2021a). The current technology lacks adequate parametric quality metrics, causing increased nutrient load in aquatic systems (Alprol et al. 2024). Further, measuring all parameters in the influent is time-taking, requiring complex tests and the use of hazardous materials, as detailed in the Standard Methods for the Examination of Water and Wastewater (Rice et al. 2012). To solve this, recent developments in electrical sensors allow for real-time measurement of quality parameters of the influent. However, key parameters like BOD5 and COD remain challenging and costly to measure with sensors. Wastewater treatment is also complicated because of natural events, human activities, and the treatment process itself. This complexity leads to uncertainties in how well wastewater treatment systems work. These uncertainties can change randomly based on factors like the amount of wastewater, its quality, and how effectively it's treated (Long et al. 2019). The presence of diverse pollutants, coupled with the need for optimal resource management, necessitates innovative alternate solutions to enhance the efficiency of wastewater treatment processes (Nourani et al. 2018). These values can be estimated using historical data by developing mathematical predictive models. The primary benefit of employing statistical and mathematical methods for predicting wastewater characteristics and the performance of WWTPs is the comprehensive understanding they provide of the underlying physicochemical and biological processes. Additionally, integrating these models with chemical reactions offers a deeper understanding of result interpretations (Aghdam et al. 2023).
AI models can be utilized throughout the wastewater treatment process to simplify various tasks and achieve more accurate results. Various single models like support vector machines (SVMs), artificial neural networks (ANNs), genetic algorithms (GAs), fuzzy logic (FL), and decision trees (DTs) have been employed for modeling processes in WWTPs. Further, hybridization is also done which involves combining two separate models, either in series or parallel, to create a more accurate model than either of the individual models alone. Commonly employed hybrid models include ML-ML models which involve combining ML models like SVM, ANN, GA, DT, and FL in several ways to enhance the prediction's efficiency and performance, ML-ML-ML models, ML-metaheuristic models and ML-ML-metaheuristic models. The model's accuracy depends on numerous factors, including the quality of data, the method of cleaning data, and the choice of suitable model structure and type. The nature and complexity of the problem of research and the study's objectives are crucial factors in choosing the method. Additionally, the quality and availability of the essential data significantly influence the choice of method (Bahramian et al. 2023). The model's performance is generally assessed using various metrics, including R (coefficient of correlation), root mean square error (RMSE), R2 (coefficient of determination), absolute average deviation (AAD%), mean square error (MSE), index of agreement (IA), sum of squared error (SSE), mean absolute percentage error (MAPE), factorial variance (FV), Nash–Sutcliffe error (NSE), and percent bias (PBIAS). Among these, R, R2, RMSE, and MAPE were the most frequently used statistical indicators (Rajaee et al. 2020). Among single models, ANN appears to outperform other models in terms of accuracy, with an R-value reported between 0.6289 and 0.998 and an R2 value between 0.616 and 0.997 (Ghaedi & Vafaei 2017; Ju et al. 2019; Najafzadeh & Zeinolabedini 2019). It has gained wider acceptance within the scientific community. Further, ANNs are widely used due to their effectiveness in handling complex, multivariate, and nonlinear problems. This capability is crucial in wastewater treatment, where the relationships between many parameters like chemical oxygen demand (COD), pH, biological oxygen demand (BOD), and others are often nonlinear and complex. Moreover, the adaptability of ANNs allows them to model various processes within WWTPs, such as biological, chemical, and physical treatment methods. This helps in simulating and optimizing different stages of wastewater treatment, from predicting influent quality to controlling effluent quality. Their scalability and potential for cost savings further justify their widespread adoption in WWTPs (Wang et al. 2023). However, when dealing with incomplete data in wastewater modeling, FL is the most commonly used method because it employs linguistic expert rules to approximate missing data (Ansari et al. 2020).
In terms of accuracy and precision, several studies have reported that hybrid models have shown relatively better performance compared with single models (Asadi et al. 2020). The performance of an ANN model can be enhanced by combining it with some other model like GA. ANN-GA combines the strengths of both ANNs and GAs. Genetic algorithms enhance the ANN training by optimizing the weights and architecture of the ANN model (Zhang et al. 2019). Traditional gradient-based methods like backpropagation can get stuck in local minima, but GAs, being evolutionary algorithms, can explore a larger solution space and find better weight configurations. Neural fuzzy (NF) and adaptive neural fuzzy intelligent system (ANFIS) are types of hybrid models which combine the learning capabilities of ANNs with the interpretability of FL, providing more understandable models (Mahshidnia & Jafarian 2016; Fu et al. 2018). This interpretability is necessary in WWTPs, where understanding the system's behavior is essential for operators.
In this article, the recent applications of the various AI models in different aspects of wastewater treatment like pollutant removal, real-time monitoring of parameters indicating water quality and fault detection in processes and equipment have been discussed in detail. In further sections, we have discussed the economic viability of AI tools and the future projects and policies undertaken globally for the application of AI models in the treatment of wastewater. Further, the ethical considerations and potential future directions for the use of AI in WWTPs have also been discussed.
DATA PREPROCESSING AND INTERPRETABILITY OF AI MODELS
Data preprocessing
This preprocessing is categorized into five main classes (Bahramian et al. 2023):
1. Data cleaning aims to detect and correct errors within the data. Techniques under data cleaning include statistical imputation, outlier removal, and exclusion of low-variance data.
2. Feature selection identifies input variables which are most appropriate for the modeling job. Methods of feature selection include wrapper, filter, and intrinsic methods.
3. Data transformation applies new probability distributions to input variables to convert raw variables into Gaussian variables. Common methods in this category are standardization and normalization. Normalization scales data to a range between 0 and 1, while standardization adjusts data to have a mean value 0 and a standard deviation value 1.
4. Feature engineering involves creating new variables not present in the original dataset to capture key features. The goal is to extract new variables from the raw data that highlight the important factors in the process of learning.
5. Dimensionality reduction compresses raw data to create a smaller set of projections. This process significantly decreases the input variables required for data training. All feature selection methods fall under the category of dimensionality reduction.
Interpretability of AI models
Nowadays, machines automatically learn, discover, and carry out the extraction of hierarchical representations of data required for classification or detection tasks. This increasing level of complexity, coupled with the huge data required for training and developing such systems, often enhances their predictive power. However, it also reduces the ability of the systems to describe their internal mechanisms and workings. As a result, the basis for the decisions becomes difficult to comprehend, leading to difficult interpretability of the predictions (Linardatos et al. 2020).
There exists a trade-off between the ML model's performance and its capability to generate interpretable and explainable predictions. On one side are the black box models, like SVMs and neural networks. On the other side are the white-box models, which generate results which are easily explainable, such as linear models and DT-based models (Silva et al. 2019). Systems with decisions that are hard to interpret are challenging to trust, especially in critical sectors like healthcare, where fairness and moral issues are significant concerns. This requirement for fair, trustworthy, high-performing and robust models for real-world applications has led to the resurgence of explainable artificial intelligence (XAI), which focuses on interpreting and understanding the behavior of AI systems (Gunning & Aha 2019). One widely accepted definition of interpretability is ‘the ability to explain or present in understandable terms to a human’ (Doshi-Velez & Kim 2017). Interpretability has also been defined as ‘the degree to which a human can understand the cause of a decision’ (Miller 2019). Therefore, interpretability is largely related to the instinct behind the output of the models. The concept is that the interpretability of an ML system is directly related to the identification of cause-and-effect relationships between the inputs and outputs of the systems.
Further, various interpretability techniques have been introduced and employed for complex black box AI models (Table 1).
Tool . | Local vs global . | Model specific vs model agnostic . | Data type . | Ref . |
---|---|---|---|---|
Lime Eli5 InterpretML AIX360 Skater | L | Agnostic | img txt tab | Ribeiro et al. (2016) |
PDPbox InterpretML Skater | G | Agnostic | tab | Friedman (2001) |
shap alibi AIX360 InterpretML | L and G | Agnostic | img txt tab | Lundberg & Lee (2017) |
alibi Anchor | L | Agnostic | img txt tab | Ribeiro et al. (2018) |
alibi | L | Agnostic | tab img | Wachter et al. (2017) |
PyCEbox | L and G | Agnostic | tab | Goldstein et al. (2015) |
L2X | L | Agnostic | img txt tab | Chen et al. (2018) |
Eli5 | G | Agnostic | tab | Altmann et al. (2010) |
alibi AIX360 | L | Agnostic | tab img | Dhurandhar et al. (2018) |
Alibi | G | Agnostic | tab | Apley & Zhu (2020) |
Alibi | L | Agnostic | tab img | Van Looveren & Klaise (2021) |
pyBreakDown | L | Agnostic | tab | Staniak & Biecek (2018) |
pyBreakDown | G | Agnostic | tab | Staniak & Biecek (2018) |
DLIME | L | Agnostic | img txt tab | Zafar & Khan (2019) |
AIX360 | L | Agnostic | tab | Gurumoorthy et al. (2019) |
AIX360 | L | Agnostic | tab img | Luss et al. (2019) |
Tool . | Local vs global . | Model specific vs model agnostic . | Data type . | Ref . |
---|---|---|---|---|
Lime Eli5 InterpretML AIX360 Skater | L | Agnostic | img txt tab | Ribeiro et al. (2016) |
PDPbox InterpretML Skater | G | Agnostic | tab | Friedman (2001) |
shap alibi AIX360 InterpretML | L and G | Agnostic | img txt tab | Lundberg & Lee (2017) |
alibi Anchor | L | Agnostic | img txt tab | Ribeiro et al. (2018) |
alibi | L | Agnostic | tab img | Wachter et al. (2017) |
PyCEbox | L and G | Agnostic | tab | Goldstein et al. (2015) |
L2X | L | Agnostic | img txt tab | Chen et al. (2018) |
Eli5 | G | Agnostic | tab | Altmann et al. (2010) |
alibi AIX360 | L | Agnostic | tab img | Dhurandhar et al. (2018) |
Alibi | G | Agnostic | tab | Apley & Zhu (2020) |
Alibi | L | Agnostic | tab img | Van Looveren & Klaise (2021) |
pyBreakDown | L | Agnostic | tab | Staniak & Biecek (2018) |
pyBreakDown | G | Agnostic | tab | Staniak & Biecek (2018) |
DLIME | L | Agnostic | img txt tab | Zafar & Khan (2019) |
AIX360 | L | Agnostic | tab | Gurumoorthy et al. (2019) |
AIX360 | L | Agnostic | tab img | Luss et al. (2019) |
Source: Reproduced from Linardatos et al. (2020). This article is published under a Creative Commons Attribution 4.0 International License.
INTRODUCTION OF AI IN WASTEWATER TREATMENT
AI can be applied in wastewater treatment for process optimization which involves the use of AI algorithms for the real-time analysis of data from sensors and instruments to optimize the various stages of wastewater treatment. This includes monitoring parameters such as BOD, pH, and COD levels, turbidity, chemical concentrations, and microbial activity (Saboe et al. 2021). Further, AI can help in the optimization and prediction of removal efficiency of different pollutants from wastewater. This has been summarized in Table 2. In addition, anomaly detection is another useful application where AI can identify abnormal conditions or deviations from expected performance in the treatment process. This includes detecting unusual variations in water quality parameters that may indicate system malfunctions or the presence of pollutants (Cheng et al. 2019). The following sections will discuss the applications of AI in wastewater treatment in detail.
Pollutant . | Treatment process . | Model . | Remarks . | R2 value . | Reference . |
---|---|---|---|---|---|
Bromophenol blue | Electro-oxidation | ANN |
| 0.946 | Picos-Benítez et al. (2020) |
Triamide | Adsorption on multi-walled and single-walled carbon nanotubes | MLR and ANN |
| 0.980 and 0.986 (ANN) 0.926 and 0.751 (MLR) | Ghaedi et al. (2016) |
2-Nitrophenol | Solar-induced degradation | RSM and ANN |
| 0.9954 (ANN) 0.9289 (RSM) | Tou et al. (2018) |
Pb2+ | Adsorption on thiosemicarbazide modified chitosan | MLP-ANN and RSM |
| 0.995 (MLP-ANN) 0.9864 (RSM) | Zaferani et al. (2019) |
Cadmium, lead, arsenic, nickel, zinc, and copper | Adsorption on 44 biochars | ANN and RF |
| 0.973 (RF) 0.948 (ANN) | Zhu et al. (2019) |
Cd2+ and Ni2+ | Adsorption on Typha domingensis | ANFIS |
| – | Fawzy et al. (2016) |
NH4+ and total nitrogen (TN) | Anammox and partial nitritation process | BP-ANN (Backpropogation ANN) | Removal efficiencies of NH4+ and total nitrogen in the range of 80–85% were observed | 0.989–0.997 | Antwi et al. (2019) |
Total Organic Carbon (TOC) and TN | combined anaerobic and aerobic processes | RSM | Maximum TOC removal = 85.03% Maximum TN removal = 72.10% Minimum residual TSS = 19.54 mg/L Maximum biogas yield = 116.56 mL/min | 0.9722–0.9910 | Bustillo-Lecompte & Mehrvar (2017) |
Nitrogen | Combination of anammox bacteria and biochar | ANN | Nitrogen removal efficiency of 90.9% | 0.990 | Mojiri et al. (2020) |
Pollutant . | Treatment process . | Model . | Remarks . | R2 value . | Reference . |
---|---|---|---|---|---|
Bromophenol blue | Electro-oxidation | ANN |
| 0.946 | Picos-Benítez et al. (2020) |
Triamide | Adsorption on multi-walled and single-walled carbon nanotubes | MLR and ANN |
| 0.980 and 0.986 (ANN) 0.926 and 0.751 (MLR) | Ghaedi et al. (2016) |
2-Nitrophenol | Solar-induced degradation | RSM and ANN |
| 0.9954 (ANN) 0.9289 (RSM) | Tou et al. (2018) |
Pb2+ | Adsorption on thiosemicarbazide modified chitosan | MLP-ANN and RSM |
| 0.995 (MLP-ANN) 0.9864 (RSM) | Zaferani et al. (2019) |
Cadmium, lead, arsenic, nickel, zinc, and copper | Adsorption on 44 biochars | ANN and RF |
| 0.973 (RF) 0.948 (ANN) | Zhu et al. (2019) |
Cd2+ and Ni2+ | Adsorption on Typha domingensis | ANFIS |
| – | Fawzy et al. (2016) |
NH4+ and total nitrogen (TN) | Anammox and partial nitritation process | BP-ANN (Backpropogation ANN) | Removal efficiencies of NH4+ and total nitrogen in the range of 80–85% were observed | 0.989–0.997 | Antwi et al. (2019) |
Total Organic Carbon (TOC) and TN | combined anaerobic and aerobic processes | RSM | Maximum TOC removal = 85.03% Maximum TN removal = 72.10% Minimum residual TSS = 19.54 mg/L Maximum biogas yield = 116.56 mL/min | 0.9722–0.9910 | Bustillo-Lecompte & Mehrvar (2017) |
Nitrogen | Combination of anammox bacteria and biochar | ANN | Nitrogen removal efficiency of 90.9% | 0.990 | Mojiri et al. (2020) |
Optimization and prediction of removal efficiency of pollutants from wastewater
Organic pollutants
Organic pollutants in wastewater are chemical compounds and substances that come from organic matter, such as carbon-based compounds, that pollute water sources. They can pose significant environmental and health risks. They can have adverse effects on human health for example; exposure to polycyclic aromatic hydrocarbons leads to an increased risk of lung cancer (Schell & Rousham 2022).
Among the pollutants of concern, synthetic dyes have garnered significant attention recently. When these synthetic dyes find their way into water bodies, they can lead to various environmental issues, including esthetic problems and reduced sunlight penetration into the water (Asad et al. 2007; Mendoza-Mendoza et al. 2018). Picos-Benítez and group evaluated the efficacy of an AI model employing GA and ANN to predict and optimize the treatment of wastewater containing bromophenol blue dye through electro-oxidation (EO). The ANN model demonstrated performance with an RMSE of 10.73% and MAPE of 8.81% (Picos-Benítez et al. 2020). The results were further validated experimentally and were found to be consistent. The conditions for the experiment identified by the AI model were successfully demonstrated to effectively remove other dyes sharing a similar structure like thymol blue and bromothymol blue. It was also shown how RBF-ANN (radial basis function artificial neural network) can be employed for the simulation and optimization of the removal efficiency of malachite green and methylene blue from water using hydrogels made of TiO2 nanoparticles and gum tragacanth biopolymer (GT or TG) (Ranjbar-Mohammadi et al. 2019). Multilayer perceptron (MLP)-ANN is preferred for highly nonlinear and complex problems, while RBF-ANN is favored for its speed and robustness in applications with localized patterns. MLP-ANN can model complex nonlinear relationships between input and output parameters but it requires significant computational resources and time for training, especially with large datasets whereas RBF-ANN is suitable for real-time monitoring and control due to faster training and inference times but can struggle with high-dimensional data typical in complex wastewater treatment systems (Bagheri et al. 2015; Ranjbar-Mohammadi et al. 2019). MLP-ANN and RBF-ANN were used for the simulation and optimization of the removal efficiency of malachite green and methylene blue from water. It was discovered that the MLP-ANN model provides better predictions than the RBF-ANN model (Zhao et al. 2020).
In another study, a single-layer neural network analysis was conducted to determine the removal efficiency of reactive dye (Red 195) using natural seed gum extracted from Cassia fistula Linn. The ANN model took five inputs, namely reaction time, initial pH, agitation speed, initial dye concentration, and gum dosage gave two outputs in the form of COD and color removal efficiencies. The ANN model was proficient in accurately predicting the coagulation process, achieving a determination coefficient (R2) of 0.924 and RMSE of 3.759 (Bui et al. 2016). Using Garson's algorithm and connection weights algorithm, it was discovered that time was the most influential factor for the decolorization process and the efficiency of the coagulation process was highly dependent on agitation speed, reaction time, and concentration of gum.
For removal of triamide, multiple linear regression (MLR) and ANN-GA models were used. These models predict the triamide adsorption on multi-walled and single-walled carbon nanotubes. When the proposed models were compared, the ANN model was found to be more suitable for predicting the adsorption efficiency of the process than the MLR model (Ghaedi et al. 2016). They obtained R2 of 0.980 and MSE of 0.002 using the ANN model while the MLR model gave R2 of 0.7828 and MSE of 0.0165. Vakili and group demonstrated the potential of ANN in optimizing the elimination of organic micropollutants like carbamazepine, bisphenol A, tonalide, and ketoprofen from wastewater. They achieved this through an adsorption technique utilizing a chitosan/zeolite fixed-bed column (Vakili et al. 2019).
Elmolla and group conducted a study to explore the use of ANN in predicting and simulating the degradation of antibiotics such as amoxicillin, ampicillin, and cloxacillin through the Fenton process in aqueous solutions. The results generated by the ANN demonstrated a high R2 of 0.997, confirming the accuracy and suitability of the model (Elmolla et al. 2010).
Utilizing response surface methodology (RSM) and ANN, Tou et al. (2018) examined the impacts of two potent oxidants, K2S2O8 and H2O2 (two powerful oxidants), on the solar-induced degradation of 2-nitrophenol in water. The findings highlighted that the ANN model exhibited better predictive capabilities compared with the RSM model, evidenced by a notably higher R² value of 0.9954. Jing et al. (2014) were able to develop an ANN model which could predict the removal of naphthalene, a polycyclic aromatic hydrocarbon (PAH), from marine oily wastewater using UV irradiation. To identify emerging microplastic contaminants in water, a convolution neural network (CNN) was employed for the classification of microbeads within wastewater using microscopic images. The CNN achieved a classification performance of 89% in accurately categorizing the microbeads (Yurtsever & Yurtsever 2019).
Inorganic pollutants
The quality of water is influenced by numerous inorganic components found in the effluent (Altowayti et al. 2019). Inorganic contaminants include inorganic salts, nitrates, mineral acids, sulfates, trace elements, phosphates, chlorides, fluorides, cyanides, oxalates, etc. (Fu & Wang 2011; Wasewar 2021). These pollutants can eliminate the microorganisms required for biological treatment in the process of treatment of wastewater (Altowayti et al. 2022).
The RBF-ANN model was used to forecast copper removal effectiveness in the emulsion liquid membrane method. RBF-ANN can be trained faster than other neural networks. With an R2 (determination coefficient) value of 0.997, the predicted values closely matched the experimental data (Messikh et al. 2015). Based on the ANFIS, it was determined that the pH had a significant impact on the Cd2+ ions ability to adsorb from an aqueous solution to Typha domingensis (Fawzy et al. 2016). When modeling the impact of thiosemicarbazide modified chitosan on Pb2+ removal, MLP-ANN shows superior accuracy of prediction (R2 = 0.990) than that of RSM (Zaferani et al. 2019).
Based on 353 datasets of adsorption studies from literature, the adsorption of six heavy metals (lead, cadmium, nickel, arsenic, copper, and zinc) on 44 biochars was modeled using ANN and random forest (RF). According to biochar properties, metal sources, environmental factors (such as temperature and pH), and the initial concentration ratio of metals to biochar, regression models were generated and optimized to estimate the adsorption capacity. In terms of adsorption efficiency, the RF model outperformed the ANN model (R2 = 0.948) in terms of accuracy and prediction performance (R2 = 0.973) (Zhu et al. 2019).
For the detection and identification of contaminants, especially micropollutants in wastewater, image-based AI models combined with spectroscopy techniques, such as scanning electron microscopy (SEM), Fourier transform infrared spectroscopy (FTIR), and Raman spectroscopy, have been found to be helpful (Enders et al. 2021). The effectiveness of salinity may be precisely assessed, and pollutant removal procedures can be predicted using a new wavelet neural network (WNN) model (Cai et al. 2019).
Quick propagation (QP), batch backpropagation (BP), genetic algorithm (GA), and Levenberg–Marquardt (LM) (algorithms for ANN model training to minimize the RMSE value for Cu and Pb ion elimination prediction) were shown to have lower incremental BP efficiency than incremental BP, according to Khandanlou and group (Khandanlou et al. 2016).
Using a Levenberg–Marquardt backpropagation (LM-BP)-trained algorithm of an ANN model, the prediction of biosorption of Ni(II), Cd(II), and Pb(II), was clarified with good accuracy (Varshney et al. 2016). The adsorption technique was used in the studies for the removal of various groupings of heavy metals (i.e., Cu–Zn and Cd–Ni–Cu–Zn), and the parameters obtained in the output included equilibrium concentrations, breakthrough curve, and the adsorption capacity exhibited by bone char (Hernández-Hernández et al. 2017; Mendoza-Castillo et al. 2018).
According to a study by Shanmugaprakash et al. (2018), to generate the best prediction of the removal efficiency of Zn(II), the LM-BP learning technique can be used for training MLP topology of an ANN. The study also suggested the use of numerous optimized variables gathered by the center composite design (CCD) of RSM.
The effectiveness of Pb, Cd, and Zn removal from wastewater using complexation-microfiltration process was predicted using ANN models. To enable their simultaneous prediction, the models were created using two distinct architectures, namely backpropagation neural networks (BPNNs) and general regression neural networks (GRNNs). As input, operating process parameters and heavy metals physical parameters were employed. BPNN operates passing information through the hidden layers to the output layer after which the output is then calculated and compared with the target value. If an error is found, the information propagates backwards toward the input layer, adjusting the connection weights layer by layer to minimize the error. However, a common issue with training ANN models is overfitting, which can diminish the model's generalization performance. In contrast, the GRNN learns quickly and converges to the optimal regression surface thereby having less chance of model overfitting. ANN trained using GRNN had R-squared values from 0.717 to 0.852 and from 0.897 to 0.955 for BPNN (Sekulić et al. 2017).
Bhagat and group presented a detailed work comparing RF, SVM, and ANN models developed to find the optimal prediction of removal of copper (Cu) in an adsorption process using raw attapulgite clay in an aqueous solution. They used different sets of inputs for the comparison which included parameters like ionic strength, pH, adsorbent concentration, contact time, and initial concentration of Cu. The results showed that RF and ANN models were more accurate than SVM. The correlation coefficient for the RF and ANN was greater than 0.99 while the correlation coefficient for SVM was 0.93 (Bhagat et al. 2021).
Nutrients
Nutrient removal from wastewater is a critical aspect of environmental engineering and water quality management. It involves the process of eliminating excess nutrients, such as nitrogen and phosphorus, from wastewater before it is discharged into natural water bodies like rivers, lakes, or oceans (Rout et al. 2021b). Excessive nutrient discharges can lead to water pollution, eutrophication, and harm to aquatic ecosystems (Wurtsbaugh et al. 2019). To address these concerns, various models have been used in wastewater treatment, including an ANN model for predicting nitrogen content with 90% accuracy (Chen et al. 2003) agent based model (ABM) for improved removal of biological phosphorus, which outperforms traditional models with a 38% improvement in acetate uptake rate (Zhao et al. 2018). In addition, strategies have been developed to optimize nutrient removal in wastewater treatment processes. An improved Q-learning algorithm-based optimization method for controlling aerobic and anaerobic processes in the removal of biological phosphorus and aiming for excellent and stable effluent quality was developed (Pang et al. 2019). Antwi and group developed two novel feedforward BP-ANN models for simulating the TN and NH4+ removal during the anammox and partial nitritation process, achieving high R2 values (0.989–0.997) and IA (0.993–0.998) (Antwi et al. 2019).
Manu and Thalla conducted a study in Mangalore, India, using AI models such as SVM and ANFIS to examine the real-time efficiency of the removal of nitrogen from a WWTP. They used input variables like total solids, pH, influent nitrogen, NH3, and COD. The study found that SVM outperformed ANFIS in forecasting efficiency (Manu & Thalla 2017). In a comparative study conducted in Korea, researchers aimed to estimate the one-day interval nitrogen concentration in wastewater effluent at the Ulsan Wastewater Treatment Plant. They employed ANN and SVM models, and both models were found to be effective in predicting effluent concentration and water quality. However, the ANN model demonstrated superior results and accuracy compared with the SVM model, showcasing its effectiveness in this context. Further, Latin hypercube one-factor-at-a-time (LH-OAT) sensitivity analysis was employed for input parameters that might potentially influence the TN concentration prediction in the effluent. In LH-OAT sensitivity analysis, the sampling of all parameters is done with OAT method precision, ensuring that any modification in the output value is credited to the input which is modified. The sensitivity ranking for the input parameters performance on the concentration of TN in the effluent highlighted the significance of spatial and temporal variables in the predictions of the model (Table 3). For the ANN model, temperature was the most crucial parameter, followed by the TN of inflow water and pH. In contrast, for the SVM model, the three most important parameters were the month, COD, and SS (Guo et al. 2015).
Rank . | ANN . | SVM . | ||
---|---|---|---|---|
VARIABLE . | Final effect . | Variable . | Final effect . | |
1 | Temperature | 38.59 | Month | 1.45 |
2 | Total nitrogen of inflow | 33.37 | COD | 1.34 |
3 | pH | 32.60 | Suspended solid | 1.33 |
4 | Volumetric flow rate of inflow | 30.58 | pH | 1.29 |
5 | Suspended solid | 26.89 | Temperature | 1.28 |
6 | Total nitrogen of food waste leachate | 22.31 | Total nitrogen of inflow | 1.24 |
7 | Month | 23.58 | Volumetric flow rate of inflow | 1.22 |
8 | COD | 17.64 | Total nitrogen of food waste leachate | 1.17 |
Rank . | ANN . | SVM . | ||
---|---|---|---|---|
VARIABLE . | Final effect . | Variable . | Final effect . | |
1 | Temperature | 38.59 | Month | 1.45 |
2 | Total nitrogen of inflow | 33.37 | COD | 1.34 |
3 | pH | 32.60 | Suspended solid | 1.33 |
4 | Volumetric flow rate of inflow | 30.58 | pH | 1.29 |
5 | Suspended solid | 26.89 | Temperature | 1.28 |
6 | Total nitrogen of food waste leachate | 22.31 | Total nitrogen of inflow | 1.24 |
7 | Month | 23.58 | Volumetric flow rate of inflow | 1.22 |
8 | COD | 17.64 | Total nitrogen of food waste leachate | 1.17 |
Source: Reproduced with permission from Guo et al. (2015).
Mesoporous nanohybrids have proven to be a valuable, cost-effective, and rapid decontaminant material for treating organic contaminants and other pollutants (such as nitrogen, phosphorus, polyphosphate, ammonium, and nitrate) in wastewater. In a study conducted by Martín de la Vega and Jaramillo-Morán, they utilized self-organizing maps (SOMs) to identify three parameters, namely aeration–oxidation, oxidation–reduction transition, and non-aeration–reduction, to provide the same information as 16 variables in a municipal WWTP. This identification was achieved by monitoring dissolved oxygen (DO) and oxidation–reduction potential (ORP) across a significant number of aeration–non-aeration cycles, totaling 3,200 cycles. The application of this method has the potential to enhance the efficiency of nutrient removal in WWTP, contributing to improved wastewater treatment processes (Martín de la Vega & Jaramillo-Morán 2018).
Bustillo-Lecompte & Mehrvar (2017) addressed slaughterhouse wastewater treatment through combined aerobic and anaerobic processes for the generation of biogas and organics and nutrients removal. RSM was used to optimize process parameters, including flow rate, influent TOC concentration, and pH. The study found that at the optimum operating conditions (pH 6.84, feed flow rate of 63 mL/min, and TOC concentration of 343 mg/L), maximum TN and TOC removals of 72.10 and 85.03%, maximum biogas yield of 116.56 mL/min, minimum TSS residual of 19.54 mg/L were achieved.
Khalid and group conducted a study where Chlorella sorokiniana was employed to remove and NH4+ from palm oil mill effluent, showcasing the potential of microalgae for the removal of nutrients in wastewater treatment. Under the identified optimum conditions (200 μmol photon m2 s−1 light intensity, 12-h photoperiod, and 28% inoculum size), an impressive 94.50% of and 93.36% of NH4+ were effectively treated (Khalid et al. 2019). Mojiri and group investigated the treatment of synthetic wastewater using a combination of anammox bacteria and biochar in a fixed-bed column for performance improvement. The study employed the ANN model for process optimization. Two reactors were compared: Reactor 1, which contained both anammox bacteria and biochar, and Reactor 2, which served as a control with only biochar. Reactor 1 consistently outperformed Reactor 2 in nitrogen removal from wastewater. The optimum effectiveness for removal of nitrogen and rate (g/L/day) for Reactor 1 varied across different phases of the study, ranging from 69.5 to 90.9% and from 6.9 to 13.0, respectively (Mojiri et al. 2020).
The process for removal of biological nutrients in a WWTP situated in Beijing, China, using the anaerobic–anoxic–oxic (A/A/O) process was simulated and subsequently optimized using an integrated system model. This model combines an extended version of activated sludge model No. 3 with an accelerating genetic algorithm (AGA) and SVM. The model produced input data for the SVM simulation which established the relation between the quality of effluent water and operating conditions. AGA was then used to further optimize the operating conditions according to the quality of the effluent. With the help of the model, it was possible to reduce the anoxic tank volume by 11% and the internal recycle ratio by 250–300% while still meeting the effluent quality requirements (Fang et al. 2011).
Real-time monitoring of water quality parameters for process optimization
Monitoring the environmental water quality involves a sequence of steps, including acquiring data, transmitting data, preserving data, and making decisions by analyzing that data. This process requires comprehensive integration of software and hardware which makes an excellent architecture the foremost priority for the water quality monitoring system as a composite system (Zhang et al. 2011).
The growing availability of measurable data, AI models, and advanced multivariate statistical techniques has led to real-time prediction and data-driven modeling becoming increasingly appealing (Zhang et al. 2023). A summarization of the AI applications in real-time wastewater monitoring is given in Table 4.
Parameter . | Model . | Key findings . | Determination coefficient (R2) . | Reference . |
---|---|---|---|---|
BOD, COD | FFNN, ANFIS, SVM, MLR | ANFIS performed better than all models in the calibration and verification phase for BOD, while ANFIS and FFNN recommended for COD | 0.7640 for BOD (using ANFIS) 0.9363 for COD (using FFNN) | Nourani et al. (2018) |
BOD, COD, TDS, TSS | ARIMA-ORELM | The hybrid ARIMA-ORELM performs better than linear ARIMA and nonlinear ORELM models | 0.99 for BOD (ARIMA-ORELM) | Lotfi et al. (2019) |
COD | ARMA + VAR | ARMA + VAR outperforms BPNN and GA-BPNN models | 0.94 | Man et al. (2019) |
bCOD | ANN | An inter-relationship between COD and trace metals in water was predicted | 0.98–0.99 | Matheri et al. (2021) |
TDS | ANN | A direct correlation between TDS and EC was observed | 0.9829 | Tarke et al. (2016) |
Parameter . | Model . | Key findings . | Determination coefficient (R2) . | Reference . |
---|---|---|---|---|
BOD, COD | FFNN, ANFIS, SVM, MLR | ANFIS performed better than all models in the calibration and verification phase for BOD, while ANFIS and FFNN recommended for COD | 0.7640 for BOD (using ANFIS) 0.9363 for COD (using FFNN) | Nourani et al. (2018) |
BOD, COD, TDS, TSS | ARIMA-ORELM | The hybrid ARIMA-ORELM performs better than linear ARIMA and nonlinear ORELM models | 0.99 for BOD (ARIMA-ORELM) | Lotfi et al. (2019) |
COD | ARMA + VAR | ARMA + VAR outperforms BPNN and GA-BPNN models | 0.94 | Man et al. (2019) |
bCOD | ANN | An inter-relationship between COD and trace metals in water was predicted | 0.98–0.99 | Matheri et al. (2021) |
TDS | ANN | A direct correlation between TDS and EC was observed | 0.9829 | Tarke et al. (2016) |
Biological oxygen demand (BOD)
AI is currently making an appearance in the wastewater treatment industry because of its effectiveness, speed, and independence from human operations. The term ‘biological oxygen demand’ (BOD) refers to one of the many characteristics that are affected by water pollution. MATLAB (Matrix laboratory) is the platform that is widely utilized to create prediction models for BOD using AI and ML approaches (Sunori et al. 2022). This method has been employed to track the water treatment facilities in the wastewater industry. The primary three AI models employed in the wastewater industry are ANN, FL, and GA. According to studies, employing ANN and hybrid intelligent systems, R2 values of 0.99 can be achieved for BOD (Malviya & Jaspal 2021).
The outlier robust extreme learning machine (ORELM) and ARIMA (autoregressive integrated moving average)-ORELM hybrid model may be employed to calculate the BOD of wastewater, and their respective R2 determination coefficients were determined to be 0.96 and 0.99, respectively (Lotfi et al. 2019). The method of link networking analysis and multilayer network introduced by Esquerre and group was used to determine BOD for units for biological treatment of wastewater (lagoon) (Esquerre et al. 2004). When predicting the BOD effluent of the Tabriz Wastewater Treatment Plant, it was discovered that the supervised committee fuzzy logic (SCFL) model was superior to individual FL due to its lower MAPE of 4% and higher R2 value of 0.960 (Nadiri et al. 2018).
To forecast the effectiveness of BOD removal in the process of the treatment of wastewater, three distinct AI-based nonlinear models – ANFIS, feedforward neural network (FFNN), and SVM were used in addition to a traditional MLR. Using weighted averaging ensemble, simple averaging ensemble, and neural network engine (NNE), the efficiency of AI modeling in predicting BOD was improved by 14, 20, and 24%, respectively, after the verification phase (Nourani et al. 2018). To estimate BOD, model trees (MTs), gene expression programming (GEP), and evolutionary polynomial regression (EPR) have been used. Nine input factors, including Ca2+, Na+, Mg2+, , , , EC, pH, and turbidity, were chosen as effective variables to create the proposed models (Emamgholizadeh et al. 2014). Results of these approaches, training and testing phases have been looked into. Performance results showed that the EPR technique was generally superior to the GEP and MT models. Gamma testing was used to identify crucial variables for BOD prediction. According to research by Najafzadeh and group, Ca2+, , and pH have the biggest effects on BOD (Najafzadeh & Ghaemi 2019).
Varkeshi et al. applied ANN, ANN-GA, and a co-active neuro-fuzzy logic inference system (CANFIS) to predict the performance of a WWTP in Gorgan, Iran based on input variables like pH, TSS, charge (Q), BOD, and COD. BOD, TSS, and COD were the output parameters. Results showed that the ANN-GA model outperformed the FL technique, with NRMSE values of 0.15 for COD, 0.19 for TSS, and 0.15 for BOD, and correlation coefficients up to 0.930. The study concluded that ANN-GA provides an effective tool for understanding and simulating the nonlinear behavior of WWTPs, offering high accuracy and requiring fewer input parameters, making it cost-effective for plant operators and decision-makers (Varkeshi et al. 2019).
Chemical oxygen demand (COD)
COD signifies the presence of chemical compounds that do not easily break down and are harmful to the microorganisms in wastewater (De Canete et al. 2016). Predicting the quality indicator of wastewater, COD is important as considerable expenses and time are involved in its measurement. Further, the prediction offers decision-makers the chance to fine-tune operational parameters for optimal energy usage and detect any unusual changes in the influent of WWTPs resulting from potential upstream discharges (Ching et al. 2022).
Numerous models are available for predicting, optimizing, and simulating, the COD removal in both physicochemical and biochemical treatment processes within WWTP. A hybrid AI model, utilizing ARMA and vector auto-regression (VAR), was employed for COD forecasting in urban sewage treatment plants. The predictive accuracy of this model surpassed that of BPNN and GA-BPNN, reaching nearly 99% (Man et al. 2019). Nourani and group compared three distinct models for assessing the effectiveness of a wastewater treatment facility concerning variables such as BOD, TSS, and COD. They found that the ensemble neural network model exhibited superior reliability in predicting BOD (with a 24% improvement) and COD, along with total nitrogen in the effluent (approximately 5% improvement for both) (Nourani et al. 2018). An ANN model with ML was developed by Matheri and group to predict and understand the relationship between trace metals and COD in WWTPs in South Africa. The model predicted the parameters with R2 of 0.98–0.99 and a sum of square error (0.000129–0.1598), thus making it a potential tool to predict the WWTP performances (Matheri et al. 2021). In a study used to optimize an anaerobic upflow sludge blanket reactor for the treatment of saline wastewater, the ANN model in combination with GA was used to predict the COD parameter. Good optimization of the process was observed with a COD removal efficiency greater than 70%, without adding any energy to the process and under the conditions of high organic loading rate and high conductivity (Picoz-Benitez et al. 2017). In a recent study, AI models including FFNN, SVR, and ANFIS were used to predict the COD of the effluent from the Tabriz Wastewater Treatment Plant, Iran using the data from 2016 to 2018. Among the studied models, SVR provided the most reliable results. Further, the jittering data preprocessing method and the data post-processing ensemble method improved the accuracy of the prediction by up to 20% (Nourani et al. 2021). In a study, the COD of Karaun River, Iran was predicted using least-square support vector machine (LS-SVM) and multivariate adaptive regression spine (MARS) ML models. The results were compared with those obtained using MLR, ANFIS, and multinomial logistic regression (MNLR) models. It was observed that MARS and LS-SVM performed better for COD prediction in terms of F test and external validation criteria (Najafzadeh & Ghaemi 2019). Wan and group developed an ANFIS model for predicting COD and suspended solids (SS) for a paper mill wastewater treatment process effluent in Dongguan (China). The ANFIS model was able to achieve minimum MAPEs of 1.003 and 0.5161% for COD and SS, respectively, and R values of 0.9912 and 0.9882 for COD and SS, respectively (Wan et al. 2011).
In a study by Aghdam and group, GEP, MLP, MLR, gradient boosting, k-nearest neighbors, and regression trees-based models were used to predict the COD values of the effluent from seven WWTPs in Hong Kong during a 3-year period. Of all the models, GEP provided the most accurate COD prediction with an R2 value of 0.861 (Aghdam et al. 2023). Aghdam et al. used Monte Carlo simulation (MCS) to carry out the sensitivity analysis of the model, which provided best predictions according to the statistical metrics of the models. COD and BOD were the target parameters as they can be used as alternative indicators for estimating both the consumption of energy and production of sludge in a WWTP. The predictors used for BOD and COD included TSS, organic nitrogenous compounds, NH3, inorganic phosphorous compounds, and organic phosphorous compounds. After collecting 1,000 data points for both target parameters and predictors, the sensitivity of COD and BOD5 to each predictor is determined by varying the predictor value by ±10%, while holding the other predictor values constant. It was seen that TSS had the maximum influence on BOD and COD estimation and NH3 is the second most influential factor in the estimation of COD after TSS, while InorgP, OrgP, and OrgN have a negligible effect on the estimated value of COD (Table 5; Aghdam et al. 2023).
Parameter . | Average value . | Variation in the average value . | BOD (mg/L as O2) change (%) in comparison with the average value . | COD (mg/L as O2) change (%) in comparison with the average value . |
---|---|---|---|---|
TSS | 309.11 mg/L | +10% | 7.94 | 7.918 |
−10% | −7.88 | −8.269 | ||
NH3 | 23.42 mg/L as N | +10% | 0 | 5.176 |
−10% | 0 | −5.176 | ||
Organic nitrogenous compounds (OrgN) | 16.28 mg/L as N | +10% | −0.448 | 0.415 |
−10% | 0.585 | −0.415 | ||
Organic phosphorous compounds (OrgP) | 2.49 mg/L as P | +10% | 0.358 | −0.723 |
−10% | 0.077 | 1.580 | ||
Inorganic phosphorous compounds (InorgP) | 2.92 mg/L as P | +10% | 0 | 0.142 |
−10% | 0 | −0.142 |
Parameter . | Average value . | Variation in the average value . | BOD (mg/L as O2) change (%) in comparison with the average value . | COD (mg/L as O2) change (%) in comparison with the average value . |
---|---|---|---|---|
TSS | 309.11 mg/L | +10% | 7.94 | 7.918 |
−10% | −7.88 | −8.269 | ||
NH3 | 23.42 mg/L as N | +10% | 0 | 5.176 |
−10% | 0 | −5.176 | ||
Organic nitrogenous compounds (OrgN) | 16.28 mg/L as N | +10% | −0.448 | 0.415 |
−10% | 0.585 | −0.415 | ||
Organic phosphorous compounds (OrgP) | 2.49 mg/L as P | +10% | 0.358 | −0.723 |
−10% | 0.077 | 1.580 | ||
Inorganic phosphorous compounds (InorgP) | 2.92 mg/L as P | +10% | 0 | 0.142 |
−10% | 0 | −0.142 |
Source: Reproduced from Aghdam et al. (2023). This article is published under a Creative Commons Attribution 4.0 International License.
Other water quality parameters (pH, turbidity, hardness, temperature, TDS, and conductivity)
Post and group integrated a CNN model with laser-induced Raman and fluorescence spectroscopy (LIRFS) to enable real-time monitoring of water quality parameters. Their approach yielded a strong correlation with an R2 = 0.74 across all samples. These findings indicate that this approach can provide highly accurate measurements, surpass detection thresholds, and identify micropollutants that traditional monitoring methods cannot detect (Post et al. 2022). The water quality of the Godawari River was modeled using ANNs. Over an 11-year period, input parameters related to total dissolved solids were carefully selected. The MATLAB software was used to implement and evaluate the ANN model's architecture and training effectiveness. The study revealed a strong correlation coefficient of 0.946 between total dissolved solids and EC, indicating a direct relationship. The results provided by the model were highly precise and dependable (Tarke et al. 2016). The various water quality parameters like DO, pH, temperature, and training and transfer functions in subsurface and surface water were predicted (Sarda & Sadgir 2015).
Zhang et al. implemented ANN-GA trained on data from around 45 drinking water treatment plants in China to forecast drinking water production using input variables like COD, temperature, electricity consumption, etc., and the results showed that the ANN-GA performance was much better than only ANN. The performance of ANN-GA, in terms of coefficient of determination (R2), was around 0.93 (with a MAPE of 20.3%) while the R2 value of ANN alone was around 0.70. The model was trained on around 11 input parameters. Out of these 11, it was seen that changes in turbidity, NH4, COD, and pH of raw water, and the amount of residual chlorine in treated water greatly impacted the model simulation accuracy, whereas the influence of temperature on the model performance was very limited. In terms of operational parameters, variations in coagulant, electricity consumption, and active chlorine dosage significantly influenced the accuracy of the model. Notably, active chlorine dosage had the most pronounced influence during this test, while the dosage of lime hydrate and the tertiary process had minimal effects among all variables (Zhang et al. 2019).
Fault detection, diagnosis, and prognosis in wastewater treatment
In fault diagnosis, classification algorithms like SVM, NN, relevant vector machine (RVM), and extreme learning machine (ELM) are frequently employed to assign labels to different classes, identifying normal and abnormal samples. Single-class classifiers are employed for the detection of single faults, while multi-class classifiers allow the simultaneous identification of multiple faults. Clustering and classification can also help determine the main cause of faults detected (Zhang et al. 2018).
Process faults
Process faults encompass a wide range of issues such as equipment malfunctions, process inefficiencies, or other anomalies that disrupt the effective treatment of wastewater. The process faults in an urban wastewater system mainly consist of sludge bulking, sewer corrosion, and performance failures of technologies for wastewater treatment (Liu et al. 2023).
Sludge bulking
Sludge bulking is a major issue in over 50% of WWTPs that use the activated sludge process (ASP). It primarily occurs due to the overgrowth of filamentous bacteria present in the secondary clarifier. Sludge bulking leads to ineffective separation of activated sludge and causes solid material to be lost in the final effluent. This not only results in increased operational costs but also leads to poorer overall plant performance (Liu et al. 2016).
Han et al. (2019) introduced a smart identification method and self-organizing fuzzy neural network to detect and sort various sludge bulking types based on the sludge volume index (SVI). However, this method relies primarily on single-factor fault detection, limiting its ability to capture the complex relationships involved in sludge bulking events. In a recent study, an advanced soft sensor using hybrid deep learning was devised which incorporated a sparse constraint stacked autoencoder (SAE) model. This sensor aimed to evaluate the SVI and classify bulking states in WWTP effluent, enhancing the overall quality of effluent. The proposed approach surpasses existing methods in SVI modeling by a notable margin, ranging from 38 to 78%. Benitez et al. (2023) proposed a novel graph-based monitoring framework utilizing advanced techniques for detecting and diagnosing sludge bulking events. Historical datasets were utilized in the framework during normal operations to understand relationships between process variables and pertinent features. The dynamic Bayesian network (DBN)-based method for diagnosis and accurate identification of the major sludge bulking root causes, particularly those associated with sudden drops in COD concentration, achieving a 98% accuracy – a significant 11% improvement over state-of-the-art techniques.
Han et al. (2021) proposed two models for identifying sludge bulking based on data knowledge-driven (DKD) methods. Firstly, a data-driven detection model employing a recursive kernel principal component analysis (RKPCA) algorithm and a DKD model based on the Bayesian network (BN). The developed DKD method was tested on real WWTP data, and the experimental results demonstrated that the proposed DKD method can be used for diagnosis and accurate identification of the major sludge bulking root causes.
Liu and group introduced an innovative maintenance framework that can facilitate cooperative actions, including causality analysis, fault detection, assessing the remaining useful life (RUL), prediction, and conducting maintenance. This framework was designed to provide support for the management of sludge bulking in wastewater treatment processes (Liu et al. 2020). Zhao and group employed a deep U-Net model, utilizing data augmentation, for automated floc and filament segmentation identification through phase contrast microscopy (PCM) image analysis. This approach enables the derivation of an SVI sensor by analyzing the segmentation of flocs and filaments, thus facilitating the early detection of filament bulking with the help of image detection (Zhao et al. 2019). Further, a kernel extreme learning machine (KELM) model was proposed which was optimized by an improved mutation bald eagle search (IMBES) optimizer. The experimental findings revealed that, in comparison to other methods like CNN, long short-term memory (LSTM), and integrated bagging ensemble system with least squares support vector machines (IBES-LSSVM), this approach exhibits a notable enhancement in accuracy prediction. Additionally, at an equivalent confidence level, it maintains a high rate of fault detection while producing smaller confidence intervals (Zhou et al. 2023).
Sewer corrosion
H2S emission is a major reason of corrosion and odor issues in sewer networks. Proper decision-making in sewer corrosion management can be achieved through accurate prediction of H2S or sulfide emissions in sewer networks (Pikaar et al. 2014). Research has shown that advancements in sensing systems designed to detect crucial indicators related to corrosion of pipe, such as pH, H2S, and temperature, have played a significant role in minimizing corrosion caused industrial equipment damage in sewage pipelines (Foorginezhad et al. 2021).
Sensors like gas sensor, pH sensor, and temperature and humidity sensors can be used for corrosion monitoring in sewerage. Montazeri et al. (2018) introduced an innovative sensor consisting of a selective microfluidic gas channel paired with a sensitive metal–oxide–semiconductor (MOS) sensor to identify nuisance gases in sewer environments. The results obtained were consistent, demonstrating a robust detection capability. Temperature plays a role in the generation of H2S in sewers. When the temperature falls within the range of 10–30 °C, a fast corrosion process is anticipated (Ams et al. 2017). Alwis et al. (2016) developed fiber Bragg grating (FBG)-based sensor for temperature-compensated relative humidity (RH) measurement in harsh environments. Rente et al. (2019) further used the FBG-based sensors by applying it in a working sewer system in Sydney, Australia, which was a successful follow-up project. This could assess a large range of changes in temperature which was not feasible in the previous study by Alwis et al. (2016). Giovanangeli et al. (2019) introduced an innovative prototype sensor based on drill resistance to assess the depth to which the concrete has corroded microbiologically. Laboratory investigations demonstrated good accuracy in millimeters, highlighting the capability of the device to detect the location of the aggregate, providing valuable information for structural evaluations of concrete.
Li et al. (2019) conducted a comparison among ANN, MLR, and ANFIS for forecasting corrosion rate and corrosion initiation time in concrete sewers. The study found that both ANN and ANFIS outperformed MLR. Additionally, the research concluded that H2S concentration is the most crucial factor in the prediction of the service life of the sewer system. Zounemat-Kermani et al. (2021) conducted a comparison of the online and standard versions of NN, Kernel ELM, and MLR models. Their findings indicated that the online ELM model performed better than other models in the prediction of loss of concrete mass for various concrete types.
Equipment faults
Instrumentation defects are mostly caused by sensors, controllers, and actuators. If the sensors, actuators, or controllers are part of a closed loop, the impact of such errors may be more severe (Liu et al. 2023).
Sensors are generally used in gathering data and information in urban wastewater treatment systems. There are two categories of sensor faults: sudden failure (abrupt, noisy, random) and deterioration failure (bias, drift, gain) (Li et al. 2020). Principal component analysis (PCA) was utilized by Luca and group to identify many common DO sensor failures (Luca et al. 2021). For sensor failure detection in WWTPs, deep learning techniques like variational residual, autoencoders, and deep dropout neural networks have been explored and used in other studies (Mali & Laskar 2020; Ba-Alawi et al. 2022).
The key components of the urban wastewater treatment systems that allow it to respond appropriately to a control system are actuators (Liu et al. 2023). Zhou and group developed an upgraded multi-way principal component analysis (MPCA) to address the multi-period fault diagnosis of actuators in a sequencing batch reactor (SBR). This approach minimizes interference and interactions between faults in different times by dividing the sub-periods and applying similarity measurement methodologies. This ensures reliable sequential fault diagnosis and recognition. Numerous variables, including the blower valve opening, blower current, DO, and wastewater level, were monitored in the paper mill using the suggested methodology. The outcomes demonstrate the viability and dependability of the suggested MPCA technique (Zhou et al. 2021). A data-driven subspace identification (SID) approach was presented by Purbowaskito and group to identify and characterize aberrant occurrences in an induction motor (Purbowaskito et al. 2021).
In a study, tests were conducted on a simulated WWTP that had six distinct fault kinds (actuators and sensors). Test findings showed that the FFNN performed well in identifying errors in 97.2% of the cases that were examined. The following seven classes are taken into consideration: normal operation state, biomass concentration sensor fault, DO concentration sensor fault, supply pump fault, recirculation pump fault, excess sludge pump fault, and partial supply pump fault (functioning at 25% of capacity) (Miron et al. 2018). In an approach by Mamandipoor and group, LSTM in deep neural networks was specifically built to capture the temporal behavior of sensor input. Over 5.1 million sensor, data points from a real-world dataset were used to assess the suggested approach. The approach outperformed conventional techniques and allowed for the prompt detection of collective problems, achieving a fault detection rate (recall) of over 92% (Mamandipoor et al. 2020).
COST–BENEFIT ANALYSIS
AI has the potential to enhance efficiency in WWTP by optimizing processes and operations. This optimization can lead to decreased energy consumption, more effective utilization of resources, and an overall improvement in plant performance. Through process optimization, AI has the capacity to minimize energy consumption, resulting in long-term cost savings. Additionally, AI can be used for the prediction of equipment failures and suggest preventive maintenance measures, thereby reducing downtime and extending the lifespan of machinery (Dwivedi et al. 2021).
It has been observed that hybrid systems using various artificial technologies demonstrated a prediction accuracy within the range of 0.64–1.00. Furthermore, the implementation of these hybrid systems resulted in a notable 30% reduction in operational costs (Zhao et al. 2020). An adaptive kernel function model based on an enhanced multi-objective optimal control strategy was proposed by Han and group and this model decreased energy consumption by 2.2, 1.2 and 1.6%, under stormy, rainy, and dry weather conditions, respectively, when compared with the adaptive multi-objective differential evolution algorithm and PI controller strategy (Han et al. 2018).
In a report by Asadi and group, a model was developed by data mining for optimizing the aeration process at a WWTP in Detroit, Michigan. Through aeration-induced oxygen reduction, they achieved a 31.4% reduction in energy consumption, when compared with the previous data for years 2012–2014. The effluent quality was maintained at a level higher than the standard (Asadi et al. 2017). In an innovative approach, Filipe et al. (2019) combined statistical learning and deep reinforcement learning (RL) for predictive control to achieve a significant reduction in electricity consumption. The implemented strategy led to a remarkable 16.7% decrease in electricity consumption in comparison to normal operating conditions.
Data-driven neural networks were employed by Zhang and coworkers to enhance the sewage pumping system performance, demonstrating the capability to maintain pumping efficiency and conserve energy. On an average, the implemented neural networks achieved energy savings of approximately 10%. In the most favorable scenario, there was a notable 25% decrease in energy consumption, highlighting the potential for significant energy efficiency improvements in sewage pumping systems (Zhang et al. 2016). A hierarchical control strategy with two levels, combining model predictive control (MPC) and feedforward (FF), was developed to optimize operating costs, considering both energy consumption and quality of wastewater. This approach resulted in a reduction of overall costs by 0.8% and a decrease in aeration energy of around 6% (Santín et al. 2015). A mathematical programming framework, guided by expert knowledge, was established to recommend a multi-criteria retrofitting measure (Bozkurt et al. 2016). The framework employed various models such as DM, GA, ANN, NF, FL, and ES to optimize construction and operational costs, including those related to energy and reagents. The application of these models may result in cost reductions of up to 30%.
There are various real-time practical examples which demonstrate the cost-effectiveness of the implementation of AI techniques in wastewater treatment. Streamwise D.I., an Australian start-up, introduced an AI-powered intelligence platform tailored for digitization at industrial grade in the treatment of wastewater. The platform utilized real-time data from IoT-based sensors, cameras, and other instruments fitted on machinery present in customer facilities, allowing users to monitor their entire setup and its functions in real-time. By leveraging this data, operators can make real-time adjustments to mechanisms and treatment processes, ensuring compliance and achieving improved efficiencies. The plant successfully reduced trade waste charges by 26%, and streamlined operator oversight by 80%, resulting in a total annual savings of $250,000 (Forbes 2021). In Cuxhaven, Germany, AI found practical application in a WWTP through the implementation of the xylem treatment system optimization. This system accurately predicted optimal set points for operating aerators, resulting in a remarkable 30% reduction in aeration energy consumption. This reduction translates to an annual saving of 1.1 million kilowatt-hours, equivalent to powering 64 homes for an entire year, as per calculations by the US EPA (Xylem 2021).
It is quite challenging to give an exact estimate about the cost-effectiveness of an AI system implementation in a WWTP because the actual cost-effectiveness of AI will vary depending on a number of factors, such as the size and complexity of the plant, the type of wastewater being treated, and the specific AI system being used. AI implementation in WWTPs involves significant upfront costs as well. These include the purchase of AI software and hardware, sensors, and other necessary equipment. Staff members may need training to understand and operate the AI systems effectively. This training incurs additional costs and time. Integration of AI systems with the existing infrastructure of the WWTP may require modifications and upgrades, leading to additional costs. Regular maintenance and updates to AI systems are essential to ensure their optimal performance. This incurs ongoing costs for software updates, hardware maintenance, and troubleshooting. However, the examples provided suggest that AI has the potential to be a very cost-effective way to improve the efficiency of WWTPs.
POLICIES AND PROJECTS FOR THE INTRODUCTION OF AI IN WASTEWATER TREATMENT IN VARIOUS COUNTRIES
Various countries have been exploring the integration of AI in wastewater treatment to enhance efficiency, sustainability, and overall water management. Here are some examples of countries that have shown interest in incorporating AI into wastewater treatment:
In 2019, the Japanese government launched the AI Strategy, and between 2019 and 2020, it increased the number of projects that received funding under this plan. Making the most of this chance to develop current products and services with cutting-edge AI technology would be advantageous for the commercial sector (Takeda et al. 2021).
The new Wastewater Treatment Plant of Nicosia (NWWTP) was planned to take care of 270,000 populations with the project horizon year 2025. Membrane bioreactor (MBR) technology has been considered during the planning of the new facility. With MBR technology, it is presently the second-largest WWTP in Europe, meeting the demands of both Greek and Turkish Cypriots (Nourani et al. 2018).
SENTRY, a US company, offers wastewater treatment systems with real-time water quality monitoring capabilities. The company makes use of unique sensor technology to assess wastewater's optical fingerprint and identify variations in water quality indicators including phosphate, ammonia, nitrate, organic matter, and pathogens. An Israeli company called Kando offers wastewater intelligence solutions for public health and water quality. Utilizing AI and ML techniques on wastewater data, the company's cloud-based software platform helps organizations adopt a more data-driven approach to manage environmental and public health issues (Omdena 2023).
DARROW researchers from Spain, Belgium, the Netherlands, and Germany will train AI models using data from a variety of sensors that monitor the water quality and treatment process at the RWZI Tilburg wastewater treatment facility, which is located in the Netherlands (Aquatech 2023). InovaYa is a France company that develops and offers sustainable membrane filtration technology with the goal of protecting water resources and enhancing access to drinking water for remote people worldwide. To maximize system performance and upkeep, the company additionally employs AI and data analytics. Puraffinity is a UK-based company that develops intelligent materials for use in the environment, such as eliminating dangerous contaminants from water and wastewater by employing AI and data science to enhance system upkeep and performance (UK) (Omdena 2023).
In order to estimate the quality of water of the Yamuna River in Delhi, India, the potential of four distinct neuro-fuzzy embedded metaheuristic algorithms was examined: particle swarm optimization, harmony search, GA, and teaching-learning-based optimization algorithm. In its sewage treatment plants, the Delhi Jal Board (DJB) has created a method for water treatment. The goal of this technology, known as ISASMA-CD (Intelligence Self-Administered Self Monitored Automatic Chemical Dosing), is to lower TSS and BOD to less than 10 ppm. It has been implemented at four DJB plants in Okhla and Yamuna Vihar (Kisi et al. 2023).
ETHICAL ISSUES RELATED TO THE IMPLEMENTATION OF AI IN WWTP
The integration of AI in WWTPs presents several ethical challenges that should be addressed to ensure the responsible and equitable use of these advanced technologies.
Ensuring transparency and accountability in AI systems used in WWTPs is critical for trust building among stakeholders and regulatory bodies. The processes for decision-making by AI must be well-documented and accessible to allow stakeholders to understand how outcomes are derived. This transparency helps in holding operators and developers accountable for the decisions made by AI systems. For instance, when AI optimizes aeration processes to balance energy use and treatment efficiency, clear documentation of these processes is necessary to ensure that the outcomes align with regulatory standards and community expectations (Monday et al. 2024). The deployment and development of AI technologies in WWTPs can contribute to increased energy consumption and electronic waste generation. While AI-driven optimizations, such as those in aeration control, aim to reduce energy usage, the infrastructure supporting AI (e.g., servers and sensors) can have a significant environmental footprint. It is essential to assess and mitigate these detrimental environmental impacts to promote sustainable AI applications (Bolón-Canedo et al. 2024).
AI systems can receive biases present in their training data, which can lead to unfair decisions. In the context of wastewater treatment, such biases can disproportionately affect marginalized communities, who may already face environmental injustices. For example, if AI models are trained on data from affluent areas with better infrastructure, their recommendations might not be as effective for under-resourced communities. This disparity highlights the need for careful consideration of data sources and the inclusion of diverse datasets to ensure fair and equitable AI applications (Alprol et al. 2024). AI applications in wastewater treatment often involve the collection and processing of large amounts of data, raising privacy and security concerns. Clear policies on data usage and ownership should be established to protect the privacy of individuals and communities affected by wastewater treatment operations (Richards et al. 2023).
FUTURE RESEARCH PROSPECTS
The application of AI in WWTPs is evolving rapidly, driven by technological advancements and the increasing need for efficient, sustainable, and resilient water management solutions. However, there are still certain limitations which exist and more research needs to be done to overcome these challenges.
An AI model like an ANN can capture and define complex nonlinear relationships between various inputs and outputs. However, while the AI model provides a mapping relationship between inputs and outputs, it lacks the ability to offer mechanistic insights into the problem being analyzed. This limitation arises because the underlying mechanisms of many wastewater treatment issues remain unclear in current research. By the combination of an AI model (data-driven approach) with a traditional mathematical model (knowledge-driven approach), data requirements can be significantly reduced, and obtaining meaningful results becomes more straightforward (Wang et al. 2023). The hybrid model, which integrates black box and white-box approaches, shows great potential as a tool for investigating the underlying mechanisms of wastewater treatment systems.
Further, the performance of processes forecasted by AI tools can sometimes diverge from actual outcomes in certain situations. For instance, abrupt changes in water quality and operating parameters may lead to inaccurate predictions by AI tools (Alam et al. 2022). It is essential to enhance the prediction capabilities of AI tools to ensure they can be effectively used under diverse conditions and adapt to sudden fluctuations in input variables.
AI models often suffer from low interpretability because the parameters like hidden layers, weights, neurons, and biases – lack physical significance. This results in challenges in understanding how these models make decisions (Fan et al. 2018). There is currently no standardized method for determining the optimal network architecture, which tends to be problem-specific. Trial and error are commonly used to find a suitable architecture, but this approach can easily lead to overfitting or underfitting (Jabbar & Khan 2015). Further research is needed to establish reasonable methods for constructing AI models. Additionally, more theoretical studies on AI techniques are necessary to address the challenges of neural network training, poor reproducibility, optimization of parameters, and low interpretability, thereby advancing AI applications in wastewater treatment.
For enhancing the AI applications in real wastewater treatment processes, operational data from real water treatment plants can serve as inputs for AI models. This will enable more accurate predictions of pollutant removal. AI technology has the potential to play a crucial role in sustainable wastewater treatment by significantly reducing operating costs and protecting the environment. Beyond predicting the water treatment process efficiency, AI tools can be utilized to integrate the entire water treatment process, including water discharge, transportation, sludge management, environmental impacts, economic considerations, and policymaking. However, when reporting data, it is important to include details such as data sources, locations, process environments, and dataset ontology (Lowe et al. 2022).
CONCLUSION
The increasing application of AI in wastewater treatment marks a pivotal advancement aimed at boosting efficiency, refining processes, and optimizing resource utilization. Using AI algorithms to analyze extensive datasets in real-time has emerged as a transformative strategy for enhancing the overall operation of wastewater treatment processes. This review focuses on three major applications: the prediction of removal efficiency for nutrients, organic and inorganic pollutants, real-time monitoring of essential water quality parameters (such as BOD, pH, COD, turbidity, TDS, and conductivity), and fault detection in the processes and equipment integral to wastewater treatment. The key findings highlight the efficacy of AI models in predictive analytics, enabling accurate forecasts of pollutant removal efficiency. Real-time monitoring, facilitated by AI, ensures timely responses to dynamic changes in water quality parameters, contributing to proactive and efficient treatment. Additionally, the incorporation of AI in fault detection prevents equipment failures, reducing downtime and maintenance costs. Further, it has been observed that through process optimization, AI has the capacity to minimize energy consumption, resulting in long-term cost savings.
While AI has shown significant promise in improving wastewater treatment processes, there are challenges and potential problems associated with its implementation which need to be overcome. To overcome these challenges, there are several areas where improvements can be made. Hybrid AI models can be proposed to improve the predictive accuracy of the pollutant removal efficiency and water quality parameters. AI models should be designed to dynamically adapt to changing conditions. The ability to continuously learn and adjust in response to fluctuations in wastewater characteristics ensures that the models remain effective over time. Further research in advanced ML algorithms, real-time optimization techniques, and integrated sensor networks could inform future research directions aimed at unlocking the full potential of AI in optimizing wastewater treatment processes. In addition, beyond predicting the water treatment process efficiency, AI tools can be utilized to integrate the entire water treatment process, including water discharge, transportation, sludge management, environmental impacts, economic considerations, and policymaking.
ACKNOWLEDGEMENTS
The authors are thankful to Vivekananda Institute of Professional Studies-Technical Campus, Delhi and Shyam Lal College, University of Delhi for the infrastructural support.
DATA AVAILABILITY STATEMENT
All relevant data are included in the paper or its Supplementary Information.
CONFLICT OF INTEREST
The authors declare there is no conflict.