Tertiary treatment using ultrafiltration in an existing sewage treatment plant for industrial reuse: a modelling approach using artificial neural network with uncertainty estimation

The United Nations (UN) – General Assembly formulated 17 goals known as Sustainable Development Goals (SDGs), where 193 UN member countries pledged to implement these SDGs to achieve prosperity and peace for the people and world before 2030. The 17 goals of the SDG framework cover diverse domains, including the economy, society, and environment. Although the goals have varying purposes and different strategies for implementation, it is believed that goals mutually complement one another. For instance, Obaideen et al. (2022) conducted a comprehensive analysis of all SDG objectives, revealing that better and more efficient wastewater treatment (SDG 6: Clean Water and Sanitation) could potentially pave the way to achieve 11 out of 17 SDGs; this significant impact is due to the augmented availability and access to water.

When it comes to water availability, different countries possess varied accessibility. Let's consider India as an illustration, having a populace of 18% of the global population, comprising just 4% of the world's freshwater resources, and their availability is scattered across the region. Another cause of concern is that out of 345 river stretches in India, 45 rivers are severely polluted by domestic and industrial wastewater. As a result, India is ranked as the 13th most water-stressed country globally. Concerning water resources management, the Indian constitution declares that water and sanitation fall within states' subjects, which means intra-state water and wastewater management is the responsibility of the concerned states. With such challenges, like massive population, scarce water availability, and excessive water pollution, it is highly challenging for government utilities in developing countries like India to effectively meet the demand for various needs and safeguard water resources.

Effectively managing water resources involves not just augmenting water but also reducing water pollution that arises due to the emission of wastewater into freshwater resources. According to the recent report by CPCB (2021), 72,368 × 10³ m³/day of domestic sewage is generated in urban centres of India, where the available treatment capacity of sewage treatment plants (STPs) for treating up to secondary treatment level is 31,841 × 10³ m³/day. There seems to be a significant gap between available treatment capacity and functional capacity (sewage collected and treated), which is about 63%. The matter of concern is whether the mentioned operating capacity can adhere to produce the effluent quality to the prescribed stringent effluent discharge limit. The prescribed effluent standards in India are set and revised by the respective State, Central Pollution Control Board, and Honourable National Green Tribunal (NGT); revision of standards in India over time is well-reported by Schellenberg et al. (2020).

Despite the multitude of challenges, the Government of India introduced several schemes over time to improve water and sanitation infrastructure, such as Jawaharlal Nehru National Urban Renewal Mission (JNNURM) (launched in December 2005), Atul Mission for Rejuvenation and Urban Transformation (AMRUT) (launched in June 2015), and the recently launched AMRUT 2.0 (October 2021). Primarily, AMRUT 2.0 targets revitalizing water bodies, expanding clean water access from 500 cities to 4,800 statutory towns (earlier AMRUT's mission was to provide water supply for 500 cities). In addition, managing sewerage and septage (collection and treatment) in 500 cities and, more importantly, prioritizing the safe reuse of treated wastewater (SRTW). It is essential to remember that SDG 6.3 and AMRUT 2.0 apprise not only wastewater treatment but also a significant increase in recycling and SRTW to promote a circular economy (Zhang et al. 2016).

In order to achieve a circular economy, STPs act as a resource recovery factory to extract resources from sewage, such as energy, nutrients, and reusable water (Ali et al. 2022). A report by the Ministry of Jal Shakti (MoJS), Government of India, states that treated wastewater as ‘Apna Jal’ means ‘our water’ to detach the perception of treated sewage as a liability instead as a resource. The Government of India formulated several acts and policies pertaining to water reuse, but it all started with the Water Prevention and Control of Pollution Act of 1974 to restore aquatic resources by not discharging sewage or pollutants. In 1986, the Ganga Action Plan and the Environment Protection Act were passed to protect river Ganga, and the latter is an act relating to protecting and improving the environment of human beings along with plants and other living creatures. Concerning water reuse, emphasis has been given in the National Environment Policy of 2006, which spotlights the need to prepare action plans by addressing water pollution and regulatory frameworks for major cities in India. The National Urban Sanitation Policy drafted in 2008 emphasizes the safe disposal and handling of human waste and endorses recycling and reuse.

The National Water Policy 2012 (revised draft – 2020) necessitates recycling and reusing wastewater. The growing water demand and limited freshwater availability obliged the formulation of a comprehensive nation-level framework for SRTW; thus, the MoJS collaborated with the India-European Union (EU) Water partnership and GIZ, a German development agency, to craft a strategic policy (MoJS 2022). At this point, it is essential to mention that the developed framework is intended for non-potable reuse, such as aquifer recharge, on-site usage within treatment plants (landscaping), municipal uses (parks, fire fighting, toilet flushing), agriculture, aquaculture, and most importantly, industrial use.

Agriculture requires significant water consumption, where the national demand is almost 70%, but providing treated wastewater for industries with higher effluent quality is even more challenging. India's water demand for various industrial purposes varies between 8 and 10%, whereas state-level demands deduce that Maharashtra state is the second largest consumer, with an annual requirement of 6.7 billion cubic metres (Joseph et al. 2019). Along with the higher water demands for industries, CPCB (2021) reports that Maharashtra stands first in generating a significant volume of municipal sewage of 9,107 × 10³ m³/day with a proposed and operational capacity of STP as 9,295 × 10³ m³/day, creating an excellent opportunity for SRTW in the industrial sector. This endeavour aims to mitigate the depletion of freshwater resources and eliminate the harmful effects caused by wastewater on the natural environment. In this context, Maharashtra has several noteworthy initiatives in SRTW; for example, in 2008, Maharashtra Generation Company Limited (MahaGenCo) sought to expand its power capacity, requiring additional water supply. Instead of relying on freshwater sources, MahaGenCo approached Nagpur Municipal Corporation to provide raw wastewater, where the transportation and treatment are under the scope of MahaGenCO, where the treated wastewater is used for power generation. The project was mutually beneficial for both parties involved with significant spillover benefits; interested readers can avail more information regarding this case study from the World Bank report (World Bank 2019).

Another important step taken by the Maharashtra government to improve water reuse was a policy devised on 1 November 2017 to use only recycled wastewater for industrial purposes, with no more freshwater utilization in industries; a deadline of 3 years was given to 71 municipal bodies to devise a plan and start using treated wastewater. In this context, municipalities investigated neighbouring industrial areas’ water requirements along with specific types of industries present (known as demand profiling) and the operational efficiency of STPs to devise a plan for SRTW for the non-potable purpose.

In this context, Navi Mumbai Municipal Corporation (NMMC), one of the progressive and exemplary urban local bodies of Maharashtra, proposed a compelling plan to integrate the tertiary treatment plant (TTP) into the existing couple of STPs to meet the industrial water demand of Maharashtra Industrial Development Corporation (MIDC) areas. The effectiveness of TTP is dependent on the STP's capability to treat sewage up to the required input quality levels for TTP. To illustrate the importance, a report by NGT on compliance of treatment facilities in India states that out of 15,403 STPs (non-municipal and municipal), 608 are not following the adhered-to prescribed standards, and legal cases are filed on 15 STPs (NGT 2019). Within this frame of reference, there is an obligatory requirement to model the performance of existing STPs and to check the reliability of input parameters to TTP, such as flow rate (Q_T), biochemical oxygen demand (BOD), chemical oxygen demand (COD), and total suspended solids (TSS).

Over the years, research studies have reported that modelling an STP is an intricate task because the time series dynamics of quantity and quality parameters of wastewater comprise seasonality, trend, and periodic components, creating high non-linearity and non-stationarity in the data. Modelling Q_T of STP is complex because of the stochastic variation in time series due to some of the uncertainty involved, such as the non-linear rainfall-runoff relationship of the combined sewer system, wearing infrastructure, groundwater infiltration, and many more. Modelling Q_T would be beneficial to understanding the quantum of secondary treated effluent (STE) available diurnally to TTP (to fix the capacity of TTP), planning operational schedules, and optimizing pump plans. We reviewed the various academic literature on modelling Q_T of STP, where different researchers tried varied artificial intelligence (AI) and time series modelling techniques for forecasting daily Q_T data: Fernandez et al. (2009) utilized the neuro-fuzzy technique using two input features, Boyd et al. (2019) applied autoregressive integrated moving average (ARIMA), Zhu & Anderson (2019) developed and employed iterated stepwise multi-linear regression, and Heo et al. (2021) integrated multiple deep learning models to predict both long- and short-term forecast using ensembled deep learning architecture.

Some researchers worked with low temporal resolution hourly Q_T data using non-linear autoregressive exogenous input neural networks (Ghazouli et al. 2021) and seasonal ARIMA (Do et al. 2022). Wang et al. (2022) critically reviewed the modelling of non-linear data of STPs and acclaimed that an artificial neural network (ANN) would be a propitious option for modelling. ANN is a forerunner while modelling municipal STPs (Hamed et al. 2004; Wadkar et al. 2021); many such studies investigated and compared different modelling techniques with ANN and inferred the resilience of the approach (Asami et al. 2021; Hejabi et al. 2021). However, it is crucial to remember that ANN can mimic closer to the actual values of the interested parameter of STP if sufficient training data establishes a more robust model architecture.

Establishing a robust model architecture would be a challenging problem because of the number of hyperparameters involved in ANN modelling; detecting optimal hyperparameters would be an arduous and time-consuming task because of the vast search space. Different ways to figure out optimal hyperparameters include grid search, random search, and applying heuristic algorithms. Most of the research works attempted grid search (an exhaustive search technique through a predefined combination of parameters), which requires ample time and high computational requirements. Some studies used heuristic and evolutionary algorithms such as differential evolution, gene expression programming, and genetic algorithm (GA), to mention a few. For instance, Liu & Chen (2023) applied evolutionary and greedy search optimization algorithms, such as GA and sequential feature selection methods, to model BOD and found that support vector regression integrated with GA provided better performance than other modelling techniques. In wastewater optimization modelling, GA performed better, along with different AI algorithms reported in several studies (Bagheri et al. 2015; Mu et al. 2016; Picos & Peralta-Hernández 2018); to the best of our knowledge, there has been no research work involving ANN-GA for modelling Q_T. In addition, most studies try optimization to mimic the real-world system perfectly by improving the performance metrics and not including uncertainty in the modelling (Hejabi et al. 2021). This research employed the prediction interval (PI) method to capture the uncertainty associated with the outcome of ANN-GA modelling. By incorporating this modelling approach, we can assess the reliability and resilience of the TTP input parameters.

This research focuses on two primary objectives; as prior art, an ANN-GA modelling technique is leveraged to model the volume of the STE from existing STP to conclude the reliability of the input flow available to TTP along with its quality parameters such as BOD, COD, and TSS. This objective intends that TTP relies upon a range in which the influent parameters must be available; a higher magnitude of parameters from STE (or failure of existing STP to produce intended quality) may lead to the failure of TTP. This soft sensor model will inform a decision on whether to proceed with the treatment of STE using TTP or divert STE to creeks as a preventive measure against the potential failure of TTP. Another reason for this objective is that Maharashtra's policy for industrial reuse strictly mentions using treated wastewater, so it is necessary to check the uninterrupted availability and the intended quality of STE. The second objective is to assess the performance of TTP by examining the tertiary treated wastewater (TTW) and statistically inferring the adherence to standards of NMMC for industrial reuse of effluent parameters: BOD, COD, and TSS, from the inception of TTP operation from September 2022.

This research article is outlined as follows in the subsequent section. The Study area section briefs about the existing Koparkhairane STP and the plan of action by NMMC on the construction and commissioning of TTP. The Materials and methods section provides a systematic way of achieving the research objective of modelling and performance evaluation. The Results section delineates modelling for forecasting and performance evaluation of TTP. A section on Discussion has been included to provide more insights about results with some additional analysis. Finally, the conclusion with the future scope and practical implications of the study are provided in the Conclusion section.

NMMC, located in Maharashtra state, is one of the largest planned cities globally and has a successful streak of being one of the cleanest mega cities in India; additionally, it holds a five-star rating for its maintenance of a city free of waste (NMMC 2021). The NMMC region is situated on the eastern shores of Thane Creek, in the Thane district, comprising seven fully developed nodes spanning from Airoli in the north to CBD-Belapur in the south with an area of 108.6 km². Out of this area, 60% of the region is dedicated to urban zones, including residential, commercial, and industrial, which includes transport infrastructure such as rails, roads, and many more. Concerning municipal sewage treatment infrastructure, seven STPs are placed strategically in seven nodes comprising different capacities with sequential batch reactor (SBR) technology. NMMC STPs, better performance and spillover effect provided an excellent opportunity for advancing towards SRTW (Jothiprakash et al. 2020).

In this context, NMMC conducted demand profiling within the corporation and determined that the significant demand for this resource would be in the MIDC industrial area. MIDC is situated on the eastern side of Thane-Belapur road; the industrial nodes of MIDC include Vashi, Turbhe, Ghansoli, Airoli, and Koparkhairane, requiring almost 40 × 10³ m³/day of recycled water. In response, NMMC decided to unify ultrafiltration (UF) technology with ultraviolet treatment to existing STPs. The existing STPs, namely Koparkhairane and Airoli, have been selected to unify the TTP of 20 × 10³ m³/day each due to their proximity to the MIDC area.

Since the study involves both the modelling of STE and performance evaluation of TTP, it requires enormous data to conclude the effectiveness of both treatment modules. Data have been collected to perform ANN-GA modelling from Koparkhairane STP for STE parameters such as Q_T, BOD, and COD having diurnal data of 3 years in length (January 2020 to December 2022), where TSS has been modelled using available minimal data (January 2020 to November 2020). For the performance evaluation of TTP, we performed the Wilcoxon signed-rank test on a data length of almost 2 months from the inception of the functioning of TTP (September 2022 to October 2022). A step-by-step methodology to attain the desired objective is outlined in the subsequent section.

Given that the research encompasses two primary objectives, the methodology to accomplish the intended objective is elucidated in two parts. Thus, the Materials and Methods section is further divided and explained step by step as modelling STE of Koparkhairane STP and performance of Koparkhairane TTP.

In the pre-processing phase, data are examined to identify missing data and outliers from the operational records of STP. We employed spline interpolation to impute the missing values in the time series dataset. Concerning the elimination of outliers, we calculated the interquartile range (IQR) of the time series and then excluded the values below (Q1–1.5 IQR) and above (Q3 + 1.5 IQR) (Alsulaili & Refaie 2021).

ANN is a mathematical modelling tool mimicking the human brain and nervous system for prediction and forecasting. This mathematical framework comprises the ability to portray any complex non-linear phenomena by capturing the relationship between the input and output of any system. In a simple context, there is no requirement for an explicit explanation about the physical elucidation of the system required for modelling. One of the popular types of ANN framework is the feed-forward neural network, which means the data transmission happens in the forward direction.

The MLP presented in this research comprises a deep feed-forward and back-propagation network architecture encompassing five distinct layers: one input layer, three hidden layers, and one output layer. Two dropout layers have been added, (i) between the first and second hidden layers, and (ii) from the second to the third hidden layer. The compelling reason behind applying the dropout layer is to control overfitting because of the deep architecture of ANN; a dropout regularization randomly disables some nodes during training, resulting in numerous different network configurations. Yaqub et al. (2020) suggest avoiding more dropout layers, which may result in loss of trend and accuracy in wastewater modelling; that is the reason behind restricting to two dropout layers. We tried different activation functions between all the network layers, such as Sigmoid, Hyperbolic tangent (Tanh), Rectified Linear Unit (ReLU), and Softmax with the optimizer as Adam algorithm. The learning rate parameter of the Adam optimizer is 0.0001; this is a critical hyperparameter governing the overriding of old data by new data, which is crucial in the modification of weight based on loss gradient. Thus, the parameters mentioned above for formulating the network architecture, such as input vector lag (past historical data), number of hidden neurons in each layer, dropout rate, and activation functions, are tuned using GA to get minimum loss function (mean squared error).

The primary goal of optimization is to figure out the optimal solution out of all the available solutions. Among several optimization algorithms, GA is a metaheuristic algorithm that draws inspiration from Darwin's theory. GA employs the concept of survival of the fittest by direct search method mimicking natural selection and genetics, which effectively handled optimization problems in different scientific areas, such as reservoir operating policy (Jothiprakash & Shanthi 2006) and desalination efficiency of treatment plants (Sadi et al. 2019), to name a few.

GA follows a set of protocols, such as the creation of population, selection, crossover, and mutation. GA starts with generating a set of initial solutions within a feasible range using randomization called population; a solution retrieved from the current population is used for developing a new population. GA employs genetic operators and a selection process to create a new population. By employing a genetic operator, new solutions (offspring) can be made based on the current solution (parents). Whereas the selection process involves choosing some individuals from the population to create a new generation, the selection is based on the fitness function, which relies on a higher fitness score. Some ways to carry out the selection process include remainder, shift linear, stochastic uniform, roulette wheel, and tournament (Kılıç et al. 2014). Another crucial genetic operator is the crossover, which intends to facilitate the exchange of genetic information (genes) between individuals and is essential for the convergence of GA. In the crossover phase, the next generation is formed by merging two randomly chosen current-generation individuals by different methods like single point, double point, or uniform crossover. In essence, from the crossover, two individuals, known as parents, create offspring called sons; the son inherits traits of both parents. Furthermore, the parents are individuals selected from the original population, where the number of chosen parents for crossover depends on the crossover rate. Another crucial strategy in the GA algorithm is a mutation, meant to avoid local solutions and incorporate variability in the new population with some unprecedented examples.

PI is prevalent while forecasting with a regression model, where the PI provides the range in which the actual outcome may fall (Meeker et al. 2017). We employed Gaussian PI, which involves the following steps; the retrieved optimal architecture from the GA is used to create PI. Several models are fitted in the training dataset with the same network architecture retrieved from GA; it is to be remembered that each model should make different predictions. Secondly, imputing non-identical random seeds generates random initial weights on the network architecture, making different predictions due to the stochastic gradient optimization algorithm. In this research, we assigned 20 random seeds to generate 20 models that produced varying predictions; 95% Gaussian intervals' upper and lower limit can be computed using the following Equations (6) and (7):

The performance assessment or compliance concerning its effluent standards at the TTW end (NMMC fixed effluent standards of TTP based on the requirements of industries) are checked for parameters such as BOD, COD, and TSS. To evaluate the performance of TTP, a non-parametric statistical approach known as the Wilcoxon signed-rank test was employed; detailed information regarding the various statistical tests to infer the compliance of STPs is provided by Ramkumar et al. (2022). The test procedure to conduct the Wilcoxon signed-rank test is as follows:

• (a)

  The TTP effluent parameter of a single sample of n values is denoted as ⁠. For each for i = 1, 2,…, n, the signed differences were found by ⁠, where is the test value or, in this case, the prescribed effluent standard for the parameter.

• (b)

  The next step is to compute the absolute difference, and ranking with ⁠.

• (c)

  The number of non-zero values is found.

• (d)

  Each ranked item is affixed with its corresponding sign of ⁠, mathematically, it can be written as, ⁠.

• (e)

  The summation of positive-signed ranks was computed using ⁠, which can be rewritten as ⁠.

• (f)
  With the above set of procedures, one can compute the W and (number of non-zero s), which was further utilized to calculate the normal test statistic Z using the following Equation (8):

  

  (8)
• (g)

  Based on the Z, the value of p can be computed to conduct a significance test on the prescribed discharge standard of TTP against the alternate hypothesis; this research utilized a significance level (⁠⁠) of 0.05.

The reliability of achieving expected quality in the TTW relies on the intended quality of STE to receive at the influent of TTP. For instance, UF technology at Koparkhairane STP requires influent quality to be less than 10 mg/L for BOD, 50 mg/L for COD, and 20 mg/L for TSS. In this context, the retrospective data from STE, such as Q_T, BOD, COD, and TSS, are modelled for forecasting.

Data pre-processing is carried out by imputing the missing value, and outliers are ascertained and removed by computing the quartiles of univariate data and imputing them with the spline interpolated data. The raw STE data comprised missing datapoint entries of 37 (Q_T), 4 (BOD), and nil for COD and TSS; concerning the number of outliers, 5 in the case of Q_T, 18 for BOD, 19 for COD, and 6 for TSS. The corresponding pre-processed data statistics are provided in Table 1. Other than TSS, the rest of the parameters are trained and tested with 3 years of data. One conspicuous value in Table 1 is the skewness (Sk) value of BOD, which is highly right-skewed with a value of 0.82; it needs to be explored in detail to understand the effect in modelling. The rest of the parameters range between – 0.1 and 0.34 (inferring nearly normal distribution). Further data is normalized before carrying out modelling.

One of the necessary steps for capturing the temporal pattern of STE using ANN-GA is to select the range in which ANN and GA hyperparameters must be varied; this involves some random search. The range of selection for regulating the hyperparameters of ANN and corresponding parameters in GA for the search process for fine-tuning implemented in this research work is provided in Supplementary Table S1. This study used previous time lags (number of past observations) as input for forecasting the next time step; different researchers tried varying lags with the lags retrieved from the autocorrelation function or by the trial-and-error method. For a case in point, Surakhi et al. (2021) tried GA for tuning the input time lag for forecasting univariate time series, which provided better results; this research persuaded us to adjust the input vector using GA. Along with time lag, the rest of the hyperparameters, such as activation function, hidden neurons, and dropout rate, are tuned using GA.

The corresponding optimal network architecture retrieved using the GA algorithm is depicted in Table 2. The input vector's lag length varies between 2 and 10 for all four parameters; the maximum is recorded for BOD, whereas the minimum is recorded for COD. For each STE parameter, the optimal architecture of the ANN exhibits varying numbers of hidden neurons in each layer, as illustrated in Table 2. The dropout rate within the three hidden layers comprises a range of values between 0.13 and 0.77. Concerning activation function, other than softmax, all three functions are found to be in the ANN model architecture. With this optimal architecture, testing has been carried out; their performance is provided in Table 3.

The batch size and epochs, two crucial parameters of ANN, are kept as 32 and 1,000, respectively. The model architecture with optimal hyperparameters provided better results in almost all STE parameters. Concerning the training stage, parameters such as Q_T, BOD, and COD, which are trained using nearly 2 years in data length (877 data points), have an R² above 0.90 in all cases, as provided in Table 3. Whereas the TSS, where the training was carried out with 244 data points, provided a lower R² of 0.78. Meanwhile, the R² in the testing set is almost comparable to the training set, ranging from 0.94 to 0.72, deducing there is no overfit occurred in training; the reason may be due to the data pre-processing, application of dropout rate parameter, and optimal architecture retrieved using GA.

As shown in Table 3, the RMSE value of the testing set is more than the training set. Especially BOD, the RMSE with a value of 0.484 mg/L, and the corresponding R² is about 0.88 due to the extreme values in the testing set. For instance, the statistical analysis of the training and test set is provided in Supplementary Table S2. This table infers that the training BOD's mean and SD vary significantly from the testing set, which provided slightly inferior results compared to the training set.

The RSR is an error index statistic that reports the model's robustness. The optimal value of RSR ranges from zero to any positive value; a lower RSR or near zero means an ideal model or vice versa. The RSR of the training set differs between 0.2 and 0.42, where the maximum RSR is observed for TSS; it is essential to mention that if TSS has been trained with higher data points, a better performance with a lower RSR may be possible to achieve. One of the critical aspects to be noticed in the testing set RSR is 0.43 observed for BOD, whereas the training set has an RSR of only 0.23. This is because of the higher standard deviation (SD) of BOD in the testing set. TSS is observed with a higher RSR of 0.42 in the training and 0.48 in the testing set, respectively.

Another noteworthy concern is that the Q_T is decreasing in the trend at Koparkhairane STP over time, observed from statistical analysis. The average yearly Q_T at the influent of STP is 47.7 × 10³ m³/day (2020), 35.67 × 10³ m³/day (2021), and 32.81 × 10³ m³/day (2022), respectively. The decrease in average Q_T may be due to a reduction in stormwater flow into STPs due to STP being designed for combined sewer overflow. We further analysed the data of non-monsoon flow into STP, which provided an average Q_T of 39.13 × 10³ m³/day (2020), 35.18 × 10³ m³/day (2021), and 32.8 × 10³ m³/day (2022), inferring a substantial reduction in influent Q_T even during dry seasons. This may be due to the uncertainties, such as infrastructure ageing and loss during sewage transmission. This crucial factor needs to be considered, and further scrutinized analysis in this area is the future scope of work. Regarding the current situation, STP can provide 20 × 10³ m³/day of better-quality water of STE for further processing into TTP, and the calibrated model can be used as a soft sensor for forecasting, which can be integrated into the influent end of TTP.

The chosen treatment technology for TTP involves using UF technology alongside a UV disinfection unit. The adoption of this particular technology is driven by the higher effluent quality standard mandated by industries located within the MIDC area (for example, TSS ≤ 1 mg/L and turbidity ≤ 0.5 NTU). The inclusion of UV treatment into the UF system stems from its remarkable ability to eliminate microbial contamination effectively. Given that a significant volume of TTW is supplied for non-potable purposes in residential areas of NMMC, the adoption of UV treatment was made obligatory.

Moreover, gaining a comprehensive understanding of the treatment modules of TTP is imperative as it significantly brings the STE to higher quality. The TTP modules consist of several distinct units, beginning with the pre-treatment system, membrane filtration units (UF treatment), disinfection unit (UV), and maintenance cleaning units and ending with a sump to store TTW. The corresponding TTP layout of Koparkhairane with treatment modules is depicted as a schematic diagram in Figure 4(b). The pre-treatment unit comprises a feed tank and a cleaning strainer. The STE from STP must have residual chlorine of 0.3–0.5 mg/L and is stored in a reinforced concrete feed tank, where two pumps lift the raw STE (additionally, one pump acts as standby) to feed it into a cleaning strainer consisting of varying sizes of stainless steel meshes to filter out coarse material to safeguard UF trains.

The membrane unit (UF train) is specifically designed to treat 833 cumecs per hour of STE; these versatile units work in different stages, including production, backwash, cleaning, and standby or shutdown. In the production stage, water is drawn from the strainer using pumps and delivered to the UF membrane columns under pressure, where the particles get trapped. Each membrane train is equipped with a flow meter and pressure transmitter that measures the transmembrane pressure to evaluate any decrease in permeability due to membrane fouls. The backwash mode maintains the transmembrane pressure in the membrane column by reversing the flow direction. During the backwash stage, the membrane is subjected to a gentle air stream to remove accumulated particles and flushed to remove the solid particles where the backwash water is not added with any chemical solutions. Prior to commencing the cleaning stage, it is necessary to conduct a backwash in conjunction with a gravity drain of backwash water. During the cleaning stage, chemical solutions like low-concentration sodium hypochlorite or citric acid are used for regular maintenance or intensive cleaning. This process involves circulating these solutions into the UF train using clean-in-pump tanks. Based on the incoming flow, some UF trains get into standby mode if the inflow volume is less than the design flow, and shutdown mode is for major maintenance. The filtered water from UF trains, known as permeate, flows to the subsequent treatment units. After the UF treatment, the UV unit disinfects the treated water to remove microbiological contamination. Then, TTW is stored in a sump and pumped to OHT for supply to the NMMC residential and MIDC industrial area through a network of pipes.

Concerning adherence to prescribed standards, as per NMMC, the effluent quality expected for industrial reuse should be BOD ≤ 5 mg/L, COD ≤ 20 mg/L, TSS ≤ 1 mg/L, and Turbidity ≤ 0.5 NTU. We further statistically analysed the observed 38 samples collected between September 2022 and October 2022 of BOD, COD, and TSS using the Wilcoxon signed-rank test, where the formulated Null hypotheses (H_o) and Alternate hypotheses (H_a) are provided below:

• (a)

  H_o: median of the parameter = Prescribed effluent standard of the parameter, H_a: median of the parameter <> Prescribed effluent standard of the parameter.

• (b)

  H_o: median of the parameter ≤ Prescribed effluent standard of the parameter, H_a: median of the parameter > Prescribed effluent standard of the parameter.

• (c)

  H_o: median of the parameter ≥ Prescribed effluent standard of the parameter, H_a: median of the parameter < Prescribed effluent standard of the parameter.

Since the inception of treatment using TTP in September 2022, at an average of 1 × 10³ m³/day, the flow has been maintained for the trial run and supplied for the NMMC residential area for non-potable purposes. The median for BOD, COD, and TSS is 4.8, 19.6, and 1 mg/L, respectively. The performed hypothetical analysis infers that BOD and COD adhere to the prescribed discharge standards for industrial reuse. But in the case of TSS, there seems to be an infringement deduced by a hypothetical test concerning the standards. The reason may be due to the relatively minimal flow at the inception of TTP, leading to a mean magnitude of TSS around 1.3 mg/L. From the middle of October, the flow was further increased to 20% of its design capacity. From this period onwards, TSS was well below the limit of 1 mg/L. It is essential to remember that proper mitigation measures should be taken if the TTW quality does not reach the prescribed standards. There were provisions given to distribute TTW based on the effluent quality. The hardware sensors will meticulously monitor the effluent at the TTP sump and compare it against the prescribed standard; then, the real-time decision will be taken whether the TTW will be distributed to the NMMC residential area or MIDC industrial area or bypassed into creeks. Thus, the present study with hypothetical analysis infers that TTP produced the intended quality required for industrial reuse over time.

STP and TTP's sustainable and reliable operation relies on several key elements: obtaining timely data about the quality and process parameters, utilizing acquired data to forecast essential parameters, and making informed decisions based on this data to manage the system. In this context, the objective was formulated to develop a data-driven soft computing model that can effectively forecast influent parameters (STE) for TTP, informing decisions on whether to proceed with the treatment of STE using TTP or divert STE to creeks as a preventive measure against the potential failure of TTP (if STE encompasses prescribed standards of NMMC). Another objective of this research is to evaluate the performance of TTP in acquiring prescribed discharge standards for industrial water reuse.

Concerning the first objective of preparing a soft computing model based on hybrid ANN-GA along with uncertainty estimation provides excellent forecasting accuracy results. The R² of Q_T, BOD, COD, and TSS in the testing stage is 0.94, 0.88, 0.93, and 0.72, respectively. One of the advantages of this study is that it focuses solely on utilizing past data of the parameter to model the time series rather than depending on a variety of water quality parameters. Another crucial aspect is that when finding optimal input features to model time series, several studies rely on a trial-and-error approach or autocorrelation analysis to find optimum time lags. This study utilized GA to tune the input features and model hyperparameters (hidden layer, dropout layer, and activation function) to achieve minimum mean squared error and improve the forecasting performance, considerably saving time in model development compared to grid search and random search technique. Uncertainty estimation in the developed hybrid model will provide the upper and lower threshold within which the outcome may fall, which is also a novelty of the study. It is imperative to mention here that this study utilized diurnal data for developing the model, which is one of the limitations because the model developed with a higher frequency of data may provide more real-time information, leading to better control over the system. In addition, we employed ANN to model time series; several other independent AI techniques were also available, such as long short-term memory networks, random forest, support vector regression, and many more. However, after thoroughly reviewing the existing literature and considering the success achieved with ANN modelling, we opted for this method due to its ability and simplicity in modelling with higher forecasting accuracy. Our future work will be modelling hard-to-measure parameters with high-frequency data by employing diverse state-of-the-art AI techniques hybridized with optimization algorithms to integrate with SCADA or PLC to ameliorate STPs energy and chemical consumption.

Another objective is assessing the TTP's adherence to established guidelines for industrial water reuse using the Wilcoxon signed-rank test. This assessment implies improved performance in BOD and COD levels while TSS gradually meets prescribed standards over time. The influent characteristics of TTP (STE) should follow a range of values to meet the effluent standard at the TTP stage. These criteria encompass pH within the range of 6.5–9, temperature ≥22 °C, BOD ≤ 10 mg/L, COD ≤ 50 mg/L, TSS ≤ 20 mg/L, and residual chlorine within the range of 0.3–0.5 ppm. Within this context, we have expanded our analysis by thoroughly examining the magnitude of the acquired STE data comprising pH levels, temperature readings, BOD and COD concentrations between September and October 2022. We aimed to identify any observed values that deviate from the specified range of acceptable values, which may affect TTP's performance. Analysis deduces that these characteristics fall within acceptable limits, with average pH, temperature, BOD, and COD values measuring 6.6, 28.5 °C, 4.8 mg/L, and 34.45 mg/L, respectively. The treatment implemented in TTP substantially reduced the level of BOD and COD in STE, as indicated by removal efficiency analysis. The average removal efficiency for BOD was recorded as 22%, while an even higher rate of 43.4% was achieved for COD. These findings were derived from a trial run of TTP with limited capacity utilization and analysed with the collected diurnal data. As discussed for STE modelling, an increased frequency of data attained with a more reliable sensor presents an opportunity to develop a digital twin for these treatment systems. This technological advancement allows for the continuous monitoring of current conditions and enables predictive analysis to anticipate real-world scenarios. Our forthcoming research on TTP will concentrate on improving digital twins, explicitly focusing on TTP's forecasting and optimization aspects.

The primary objective of this research is to deduce the quantum of Q_T along with the anticipated quality of STE water available for TTP to assess the reliability of its input parameters' dynamic range. This objective is inferred using ANN-GA modelling along with an uncertainty estimation; the developed soft sensor models provided excellent results for Q_T, BOD, and COD with an R² value of 0.94, 0.88, and 0.93 in the testing stage, whereas TSS shows a value of 0.72. The lower forecasting accuracy of TSS is due to the minimal data length used for model development and testing. The model dynamics and the upper and lower limit of uncertainty estimate provide strong evidence that 20 × 10³ m³/day of good-quality STE water is always available for further processing into TTP. The developed model can be used as a soft sensor and requires only past parameter data, which needs to be forecasted rather than diverse water quality or quantity parameters. Another advantage of this soft sensor is that it helps in decision-making on whether to proceed with the treatment of STE using TTP or divert STE to creeks as a preventive measure against the potential failure of TTP. The decreasing trend of Q_T was highlighted, which may be due to the age of infrastructure, loss during sewage transmission and emphasized the need for further study. Another objective of this study includes assessing the adherence to prescribed standards by TTP for industrial water reuse. Analysis of TTW data for September and October 2022 using the Wilcoxon signed-rank test deduces that BOD and COD are well within the prescribed standards, where TSS shows infringement during the initial period and started following standards over time; consistency of parameters on achieving the prescribed standards infers the TTW can be used for industrial reuse. The developed ANN-GA model and performance evaluation will help to control the treatment system and decision-making. One of the limitations of this research is that the diurnal data was utilized to develop the model. STP comprising high-frequency data with diverse water quality, quantity, and process parameters can be even more efficient in modelling with the advantage of real-time control. Our future scope is to develop a more robust model with state-of-the-art AI techniques tuned with various optimization algorithms to integrate with digital twin systems for anomaly prediction and ameliorate STP energy and chemical consumption. Thus, the research work covered different aspects of water reuse, such as policy initiatives in India on water reuse, modelled the STE using ANN-GA along with uncertainty estimation of Koparkhairane STP, and TTP performance evaluation using the Wilcoxon signed-rank test.

The authors thank NMMC for providing the necessary data to carry forward this research successfully. We also thank the Maharashtra Pollution Control Board (MPCB) and the Government of Maharashtra for funding this project.

This project is carried out with funds distributed for studying the compliance of sewage treatment plants of Maharashtra by the Maharashtra Pollution Control Board (MPCB). Project funding code: RD/119 – MPCB-009-001.

Data cannot be made publicly available; readers should contact the corresponding author for details.

The authors declare there is no conflict.