Modelling and set-point definition for the development of a joint control system of two interconnected wastewater treatment plants and its application in practice

AE

aeration energy

ASM

activated sludge model

b_AUT

autotrophic decay rate

CAS

conventional activated sludge

COD

chemical oxygen demand

CSO

combined sewer overflow

DAF

dissolved air flotation

DO

dissolved oxygen

E_i

overall model efficiency

EQN

effluent quality index (nitrogen)

f_A

XI/(particulate COD)

f_B

ISS/TSS

f_S

SI/(total COD)

f_COD

fraction of SS from bio-degradable COD

GMP

good modelling practice

IQN

influent quality index (nitrogen)

ISS

non-volatile (inert) suspended solids

iWW

industrial wastewater

iWWTP

industrial wastewater treatment plant

JCS

joint control system

saturation constant for autotrophic microorganism

L

likelihood

LSA

local sensitivity analysis

ME

mixing energy

MLR

multiple linear regression

mWW

municipal wastewater

mWWTP

municipal wastewater treatment plant

N/E

nitrogen removed/(energy consumed in the whole plant)

NSE

Nash–Sutcliffe efficiency

PE

pumping energy

PID

proportional integral derivative

PLC

programmable logic controller

pWW

power plant wastewater

RAS

return activated sludge

RMSE

root mean square error

RTU

remote terminal unit

R²

coefficient of determination

SAE

standard aeration efficiency

SCADA

Supervisory Control and Data Acquisition

SI

soluble non-biodegradable (inert) COD

S_ij

sensitivity coefficient

SS

soluble biodegradable COD

SRT

sludge retention time

SOTR

standard oxygen transfer rate

SVI

sludge volume index

TN

total nitrogen

TP

total phosphorus

TSS

total suspended solid

VSS

volatile suspended solid

WAS

waste activated sludge

WSIS

water smart industrial symbiosis

XI

particulate non-biodegradable (inert) COD

XS

particulate biodegradable COD

μ_AUT

maximum growth rate for autotrophic microorganisms

Due to climate change and the limitation of valuable resources, the demand to save water, materials and energy is strongly increasing and the promotion of circular economy becomes crucial. New forms of cooperation between the industry and the water sector, also known as ‘Water Smart Industrial Symbioses’ (WSISs), are becoming an important basis to implement circular economy solutions. WSIS consist of companies or entities, which enter a partnership to cooperate for their mutual advantages. For example, the industry produces wastewater as a valuable resource and the water sector treats the wastewater to recover water, materials and energy, which can be reused by the industry again.

Large industrial parks have usually a high potential to establish a WSIS. Typically, an upstream industrial wastewater treatment plant (iWWTP) does not exchange real-time water quality and quantity data with a downstream municipal wastewater treatment plant (mWWTP). As a result, the mWWTP often operates with a conservative aeration regime that consumes more energy than is actually required for treatment. Still full-scale demonstrations of joint control strategies are rarely described in literature. The most common and widely used controllers in practice are proportional integral derivative (PID) controllers, while only a few advanced dissolved oxygen (DO) control strategies have achieved successful implementation due to the lack of high-performance hardware and the risk of equipment failure (Li et al. 2022). Bertanza et al. (2021) used a fuzzy control system to optimize the aeration of a nitrification reactor and saved more than 25% of its former energy consumption for aeration. Revollar et al. (2020) suggested a new cascaded control system to improve the global performance of a WWTP and compared its potential to existing control strategies such as a DO control, cascade control of effluent ammonium concentration and combined nitrate and DO control. For the new control strategy, the nitrogen removal (kg N) per energy consumed (kWh) of the whole plant index (N/E index) served as a global performance indicator to adapt the DO set-point in the activated sludge process. Modelling showed promising results, but to the author's knowledge it was not applied in practice until now. Both strategies of Bertanza et al. (2021) and Revollar et al. (2020) did not use any TN predictions based on real-time sensor data for their influents.

To test new control strategies and to define set-points for process control, wastewater treatment processes are usually modelled. The activated sludge model (ASM) is commonly used in modelling programmes such as SIMBA# and was stepwise developed and extended. ASM1 considers carbon oxidation, nitrification and denitrification, while ASM2, ASM2d and ASM3 consider additionally fermentation and chemical as well as biological phosphorus removal (Henze et al. 2000; Siegrist et al. 2002). Typical default values for the fractionation of chemical oxygen demand (COD) for municipal wastewater (mWW) are well known (Henze et al. 2000). In the case of a high fraction of industrial wastewater (iWW) however, the fractionation of the influent to the mWWTP has to be adapted and the calculation of different parameters may also need adjustment. Therefore, each plant needs individual calibration, what is considered as a bottleneck of ASM (Koch et al. 2000; Mannina et al. 2011). In sensitivity analyses, the effectiveness of the model to replicate and/or predict reality can be evaluated by conducting correlation analyses to reveal the model efficiency. Important correlation coefficients are the coefficient of determination (R²), the Nash–Sutcliffe efficiency (NSE), the likelihood measure (L), the root mean square error (RMSE) as described in detail, e.g. by Ahnert et al. (2007), Mannina et al. (2011), Guo et al. (2015) and Carreres-Prieto et al. (2023).

In the European Horizon 2020 project ULTIMATE, the WSIS Kalundborg in Denmark is one of nine case studies, in which an iWWTP pretreats pharmaceutical and biotechnological wastewater, before it is discharged to the mWWTP. The iWWTP is owned and operated by the industry, while the mWWTP is owned and operated by a public utility. The treated iWW is discharged to the mWWTP and corresponds to approximately 50% of the total influent to the mWWTP. Both plants have their own control systems and the exact loads coming from the iWWTP are unknown to the utility. This can be a big challenge, especially when sudden nitrogen shock loads occur. Due to the obligation to comply with legal concentration thresholds in the effluent of the mWWTP, the aeration system of the secondary treatment of the mWWTP is often operated at higher DO set-points than actually needed. This can result in an over-aeration especially during low load periods. Therefore, it can be very effective for an energy efficient operation of the mWWTP to know, e.g. the nitrogen load resulting from the iWWTP in advance in order to operate the aeration system for the nitrification on demand and to increase its DO set-points only, when it is needed, to avoid any over-aeration.

Therefore, the aim of this study was to develop a joint control system (JCS) to optimize the aeration strategy of the mWWTP. The goal was to allow a demand-driven aeration based on the TN load coming from the iWWTP predicted by an empirical equation and to increase the N/E index of the mWWTP. In addition to the TN prediction, the aeration strategy was optimized using the ASM model in SIMBA#. The research questions were:

• How reliable can an empirical equation based on existing sensors serve as an early warning for increasing TN loads coming from the iWWTP?

• How accurately can SIMBA# model the mWWTP, which receives a challenging and quite unique wastewater mix from the biotech industry and the municipality?

• What DO controller equations can be determined using sensitivity analysis within the SIMBA# model for a demand-driven aeration strategy for the JCS?

• How much can the JCS increase the energy efficiency of the nitrification (determined as N/E index) in the model and in the real environment?

Optional, phosphorus can be removed chemically, if needed. The activated sludge is separated in secondary clarifiers and a subsequent dissolved air flotation (DAF) unit. The effluent of the iWWTP enters the mWWTP after the mechanical treatment of the mWW. The united streams are treated in a conventional activated sludge (CAS) process for carbon oxidation, nitrification and denitrification. In the mWWTP, phosphorus is removed chemically and biologically supported by the side stream hydrolysis. In the case the flow rate of the joint wastewater streams exceeds 2,160 m³/h, the combined sewer overflow (CSO) is activated and discharges the pretreated mWW into the Great Belt.

For the mWWTP, the JCS is designed to operate the aeration on demand and to avoid over-aeration. It is therefore crucial to anticipate the nitrogen load from the industrial wastewater treatment plant (iWWTP) as early as possible, as it can vary considerably and typically represents between 50 and 60% of the total nitrogen (TN) load at the mWWTP under normal conditions. In order to have sufficient time to adjust the aeration system at the WWTP for a high nitrogen load, it is advantageous to have an early warning before the iWWTP effluent enters the WWTP. Therefore, an investigation was carried out to determine if an empirical equation could be developed using the concentrations of ammonium, nitrite and nitrate measured in the effluent from each bioreactor, referred to as the ‘activated sludge plant’ in Figure 1, together with the total suspended solids (TSSs) content measured in the DAF. The ammonium, nitrate and nitrite concentrations and the TSS are measured 37 h and 70 min before the iWW enters the mWWTP and reaches the BioTector (B7000 Hach), respectively. The BioTector analyses the TN concentration and is located in the iWW inlet to the mWWTP.

Online measured data from the iWWTP were used for the development of the TN prediction equation. They comprised the flow rates and the concentrations of NH₄-N, NO₂-N and NO₃-N in the seven activated sludge tanks and two online TSS measurements in the DAF effluents (Figure 1). Outliers within the online data sets were selected and excluded based on the operator's experience. To develop the empirical equation for the prediction of the TN concentration in the treated iWW, the TN online measurement (BioTector, Figure 1) was used and correlated with the calculated TN concentration for the iWW stream. The ammonium, nitrate and nitrite sensors in the iWWTP were maintained once a week and calibrated, when a drift occurred, while the ammonium sensors in the mWWTP were automatically calibrated every day and they were cleaned every 4 weeks or earlier, if a drift occurred. In addition, grab samples were taken and analyzed in the laboratory to confirm the online measurements.

For the implementation of the ASM, input data sets were generated based on measured (online and in laboratory) data. Influent (iWW and mWW) and effluent (mWWTP) data, measured in the laboratory, such as the concentrations of COD, TSS, TN, NH₄-N, NO₃-N, TP and PO₄-P were collected daily, using 24 h flow proportional automatic samplers. Process parameters such as the content of mixed liquor suspended solids (MLSS), the sludge volume index (SVI) and the content of TSSs within the return activated sludge (RAS) were also determined using daily collected grab samples. Online flow rate data were recorded for the influent flows (iWW, mWW and pWW), the RAS and wasted activated sludge (WAS) flows. Concentrations of DO, NH₄-N and NO₃-N were also measured online in the aerated tanks of the mWWTP. All data were evaluated and outliers were excluded based on the operator's experience. Table 1 shows the median values of the measured concentrations in the 24 h flow proportional mixed samples of the standard parameters for the iWW, mWW and WWTP effluent.

For the TN concentration, the TN online measurement (BioTector, Figure 1) was used instead of laboratory data. After a fist testing with a broader range from 0 to 10 for the factors x, y, and z, the values between 0 and 3 revealed to be in a useful range for the correlation analyses. The factors x, y, and z were systematically varied between 0 and 3. The running averages of the TN concentrations, determined with the BioTector, were correlated with the running averages of Equation (1). For the correlation analysis R², NSE, L and RMSE were used (refer Section 2.6).

Depending on the TN concentration measured in the influent stream from the iWWTP to the mWWTP, different DO set-point functions were applied to control the DO concentration in the aerated tanks of the mWWTP.

The aeration control is implemented as a cascade scheme where the inner loop controls the airflow rate by manipulating the valve position and the outer loop controls the DO concentration. The supervising ammonium controller controls the ammonia concentration in the reactor by calculating the DO set-point for the DO controller.

The development of the activated sludge model presented in this study followed the ‘Good Modelling Practice (GMP)’ guidelines published by Rieger et al. (2012). The guidelines provide recommendations for data collection and evaluation, model setup, calibration and validation, and interpretation of the selected model outputs. The developed model can describe the nitrogen and phosphorus removal processes and was implemented into the simulation environment of SIMBA# version 5.0 (SIMBA#, ifak, Magdeburg Germany).

As biokinetic model the EAWAG ASM3 + BioP (Siegrist et al. 2002) with the HSG parameter set according to Alex et al. (2015) was used to simulate the processes of the investigated full-scale WWTP. The effluent parameters COD, TSS, TN, NH_x-N, NO_x-N, TP and PO₄-P were selected as model outputs.

To account these fractions, the stoichiometric parameters for soluble inert nitrogen (iNSI = 0.055) and particulate inert nitrogen (iNXI = 0.054) were adjusted within the influent converter model. Also, the particulate part of the TN (0.054*TSS) was incorporated into the influent converter model.

The calibration of the model follows a four-step approach. In the first step, the calculated fractionation parameters f_A, f_B, f_S, f_COD and the ratio of VSS to COD were calibrated. In addition, the process parameters for sludge production such as MLSS, WAS flow (Q_WAS) and SVI were adjusted. The aim was to achieve a good correlation between measured and modelled COD and TSS loads in the first step. Table 3 compares the actual and modelled process parameters after calibration.

In the second step, the most influential kinetic parameters were determined on the basis of published data. For the definition of the upper and lower bounds of the kinetic parameters, a relative change of ±10% for the default values was used. The aim was to reduce the set of possible parameters as much as possible in order to avoid over-parameterization of the model. The main focus was on the kinetic parameters for nitrification and denitrification, as nitrogen removal was identified as a bottleneck. Table 4 summarizes three influential kinetic parameters for nitrification and denitrification based on the literature (Mannina et al. (2011), Zaborowska et al. (2016), Szeląg et al. (2022)) and their range of variation.

Table 4

Selected kinetic parameters for sensitivity analysis based on the literature (Mannina et al. (2011), Zaborowska et al. (2016), Szeląg et al. (2022))

Variable .	Unit .	Range .	Default .	Calibrated .
μAUT	d−1	0.56–1.68	1.12	1.18
bAUT	d−1	0.09–0.27	0.18	0.18
	mg O2 L−1	0.25–0.75	0.5	0.5

Variable .	Unit .	Range .	Default .	Calibrated .
μAUT	d−1	0.56–1.68	1.12	1.18
bAUT	d−1	0.09–0.27	0.18	0.18
	mg O2 L−1	0.25–0.75	0.5	0.5

View Large

In the third step, the selected kinetic parameters were systematically varied individually within their range of variation. After each iteration step, the influence of TN, NHx-N and NOx-N on the correlation between measured and simulated model outputs was investigated. Instead of calculating the sensitivity coefficients (S_ij) according to Petersen et al. (2002), the correlation factors described in Section 2.6 were used for the evaluation. The aim of these local sensitivity analyses (LSA) was to obtain the best correlation result for the model output of TN. It was found that the maximum growth rate for autotrophic microorganisms (μ_AUT), followed by the saturation constant for autotrophic microorganisms (⁠⁠) and the autotrophic decay rate (b_AUT), had the strongest influence on the nitrogen parameters (TN, NHx-N and NOx-N). Zaborowska et al. (2016) and Szeląg et al. (2022) found similar sensitivities for the nitrogen parameters NHx-N and NOx-N. In the fourth step, the calibration was evaluated at a global level by considering the overall model efficiency. In order to modify the original ASM + BioP as little as possible, only the maximum growth rate for autotrophic microorganisms (μ_AUT) was calibrated to model the low NOx-N load of the mWWTP effluent, in addition to the stoichiometric parameters iNSI and iNXI to account for local nitrogen fractions. Instead of using the default value of μ_AUT = 1.12 d⁻¹, μ_AUT = 1.18 d⁻¹ was used. Another way to model the low NOx-N load would be to reduce the saturation constant for autotrophic microorganisms (K_HNO3), which had a similar effect compared to μ_AUT.

In Figure S1 (Supplementary Material), the process flow diagram implemented in SIMBA# is shown.

To determine the correlation between the measured and modelled values in a reliable way, a combination of different correlation factors was used. Ahnert et al. (2007) recommended the use of the R², modified Nash–Sutcliffe Efficiency (NSE₁), and the RMSE (Equation S1, Supplementary Material) as measures for the goodness of fit. In this study, the NSE ( Equation S2, Supplementary Material) was used instead of the modified NSE (NSE₁). To give an indication of the overall model efficiency (E_i, Equation S3, Supplementary Material), the likelihood measure (L, Equation S4, Supplementary Material) was also calculated according to Mannina et al. (2011).

The R² indicates how close the estimated or modelled values are to the measured values. R² typically ranges from 0 to 1, with 1 representing perfect fit. The NSE is considered to be a statistically reliable way of accessing the goodness of fit of hydrological models by taking the variation of the specific pattern into account. The NSE ranges from −∞ to 1, where a value of 1 indicates a perfect fit and a value of 0 or negative indicates that the mean values would have produced the same level of accuracy. Along with this, another aspect of negative values of NSE with relatively high values for R² (>0.5) indicates that the model has some relevance to reality in terms of variation, but fails to reproduce the mean. To calculate an overall model efficiency (E_i), the approach according to Mannina et al. (2011) was used. Therefore, the likelihood measures (L) were calculated for each model output. To give an indication of how well the model fits in terms of numerical deviation, the RMSE was used.

Within this study the energy demand for the removal of pollutants can be divided into three categories: the aeration energy (AE), which quantifies the amount of energy used for aeration as a function of KLa, the pumping energy (PE) the sum of the PE of each unit, and the mixing energy (ME). The methodology for calculating AE, PE and ME is based on the approach described by Gerneay et al. (2014). For the calculation of AE, the standard oxygen transfer rate (SOTR) provided as a model output and the standard aeration efficiency (SAE = 1.2 kg O₂ kWh⁻¹) were employed.

The performance of the implemented JCS in the real environment was evaluated using average data from the first 3 months extrapolated to 1 year of operation and comparing them to average data resulting from a full year of operation without JCS. Therefore, the influent and effluent concentrations of COD and TN were used to calculate the eliminated loads of COD and TN for each year. To determine the COD/E and N/E-indices, the eliminated loads were divided by the consumed energy related or extrapolated to 1 year using the average specific energy consumptions. Because the flow rates differed by a factor of 1.3, COD/E and N/E indices were corrected by normalizing the ratios based on the flow rate, still using the actual eliminated loads of COD and N for each period.

The best fit considering high and low concentrations was reached for x; y; z = 0.6; 1.8; 2.3 with correlation coefficients between 0.68 and 0.77 and a RMSE of 0.3 mg L⁻¹. However, especially for smaller concentrations below 25 mg L⁻¹, the correlation with R² < 0.5 and NSE < 0 was not as strong as for peak concentrations. For concentrations of 25 mg L⁻¹ and lower, the optimal combination of x; y; z was 2.0; 0.5; 2.0, for which the correlation coefficients were between 0.54 and 0.59 with an RMSE of 1.8 mg L⁻¹. Because of the lower concentrations, measurement inaccuracies had a greater effect on the correlation, which was less strong. For the operator, however, predicting concentrations of c_TN ≤ 25 mg L⁻¹ was more important than accurately predicting peak concentrations in order to detect increasing TN concentrations as early as possible. Although the NSE of these equations was negative for higher concentrations, but combined with an R² > 0.5, the equations ‘only’ overestimated the peak concentrations that still indicated a warning, which was the actual purpose of the early warning system. The prediction of lower concentrations (c_TN ≤ 25 mg L⁻¹) considered more effective for an early warning. Also, the probability of the occurrence of events with concentrations higher than 25 mg L⁻¹ was expected to be very low. Therefore, the main focus was put on the best fits for concentrations below 25 mg L⁻¹.

Aiming at simplifying Equation (1) by neglecting either the organic nitrogen concentration or nitrite and nitrate concentrations, the combinations 2.5; 0.6; 2.0 as well as 3.0; 0.0; 0.0 for concentrations c_TN ≤ 25 mg L⁻¹ resulted in a still useful replication of the TN concentrations (R² > 0.5 and NSE > 0) even though the correlation coefficients were less good than for the other combinations neglecting the organic nitrogen concentration resulted in the loss of the TN peak at day 40. Nevertheless, it had the advantage that the TN concentration could be predicted 36 h earlier than using an equation involving the TSS content for the organic nitrogen concentration. Neglecting the nitrite and nitrate concentrations had the advantage, that only 16 sensors were needed instead of 23 sensors. Considering the time benefit of the combination that neglected the TSS content for the organic nitrogen concentration, it made sense to use it as an early warning signal. However, it is seen as necessary in a second step to confirm and/or revise the warning by using the combinations 2.0; 0.5; 2.0 and/or 3.0; 0.0; 0.0. Finally, in the third step, the BioTector could confirm/revise both signals. For testing the JCS in the real environment the operator decided to use the equation in the first 3 months (see Section 3.3) together with the signal of the BioTector to confirm the prediction. In the near future, the equation will be integrated additionally in the JCS to use the advantage of the early warning.

Even though the correlation coefficients for the best fits for c_TN ≤ 25 mg L⁻¹ were far from 1.0, indicating an identical fit, the accuracy was good enough to at least provide a first warning using data from 22 sensors. The combination of the R² coefficient with >0.5 and the NSE with > 0 showed that the predicted values were close to the actual measured values and that the variation of the measured values was reflected in the empirical equation. The combination of R² < 0.5 and with the NSE <0 were considered as not useful and therefore, the fit for 0.6, 1.8, 2.3 was not selected for the empirical equation. Furthermore, for the second warning based on 23 sensors, the RMSE of 1.8 mg N L⁻¹ was still in an acceptable range, as the increase in the TN concentration to 20 mg N L⁻¹ and above was considered as critical. Considering the high number of sensors being 16, 22 and 23, on which the TN prediction equations depended on, the correlation coefficients for R² and NSE ranging between 0.51 and 0.77 as well as 0.08 and 0.54, respectively were a very good result. Guo et al. (2015) determined the R² and NSE for two machine learning models to predict the TN concentrations based on daily water quality and meteorological data. Although the method of predicting the TN concentrations was different from our approach, the way of interpreting the results of the correlation coefficients can be compared. Guo et al. (2015) obtained values between 0.46 and 0.56 for their correlation coefficients, which they interpreted as a good fit. These values were in a similar range as our best fits, confirming our interpretation of a good fit for values between 0.51 and 0.77.

Extreme values that were higher or lower than the 1.5-fold of the interquartile range (IQR), are not shown. The measured and simulated COD loads were very similar in terms of their variation, mean and median values. The measured and simulated loads of TN, NO_x-N, NH_x-N, TP and PO₄-P had their medians on similar levels, even though the simulated loads varied less.

This is related to the controller set up for aeration and the phosphorus precipitation within the simulation environment, which were more effective than the manual operation in reality. Another reason may be the solver used (iterator, damped), which tends to smooth numerical extremes within the simulation. In contrast to the other model outputs, the TSS load was less well predicted with a deviation of 40% by the model with respect to the measured values, because the tree layer model (Takács et al. 1991) for the secondary clarifier was more effective than in reality. This uncertainty is related to the model structure (e.g. separation processes) and the simplification of its process (Belia et al. 2021). Nevertheless, the model is suitable to simulate the same median loads, but it tends to underestimate the variation for almost all model outputs. To evaluate its suitability for the scenario investigation, the goodness of fit and the overall model efficiency were determined with the following correlation coefficients.

Table 6 presents the correlation coefficients (R², NSE and L) for the selected model outputs. R² is greater than 0.7 for almost all parameters except for the NO_x-N load (0.29). The NSE coefficient is also greater than 0.7 for almost all parameters expect for the TSS and NO_x-N loads. This indicates that the model fits well to the actual measured loads of COD, TN, NH_x-N, TP and PO₄-P. For TSS, R² is higher than 0.5 and the NSE is still positive but lower than 0.5 which implies that the model has some relevance with the reality in terms of variation, but it fails to reproduce the mean. One reason for this issue is the three-layer model of the secondary clarifier used and the different settling abilities between the real conditions and the simulated model results. Calibration trials to improve the settling ability did not lead to satisfactory results. Only the SVI had a major impact on the goodness of fit. For the NO_x-N load, the NSE was negative and R² was lower than 0.5 (0.29). This implies that this parameter is not adequately predicted by the dynamic model. However, the small concentrations of the nitrite and nitrate should be noted in this context ranging from 1.0 to 1.2 mg L⁻¹ with an average concentration of 1.1 mg L⁻¹. Already small deviations in their concentrations can lead to quite poor correlation results. Similar observations were made by Mannina et al. (2011) and Solís et al. (2022), who also simulated a full-scale WWTP.

Table 6

Correlation analysis (R², NSE, L) and αj for the model outputs of COD, TSS, TN, NH_X-N, NO_X-N, TP, PO₄-P under steady state conditions, simulated for 60 days

View Large

In order to calculate the overall model efficiency, the likelihood measure (L) was also determined for each parameter. As with NSE, L is used for parameter calibration of hydrological and hydrogeological models (Molin et al. 2020). The results for L are on a similar level as those for the NSE and vary for this model between 0.35 for NO_x-N and 0.91 for COD. Besides for NO_x-N, the results indicate a good fit between the simulation outcomes and the observed data. For comparison, Mannina et al. (2011) achieved results for L between 0.12 for TP and 0.7 for TSS for the same selection of parameters and categorize their results as satisfactory. Hence, our results were as good as and for some parameters even better as the results of Mannina et al. (2011). The overall model efficiency was 0.73, indicating that the model predicted 73% accurately the real-environment conditions of the mWWTP.

To give an indication of how well the model fits in terms of numerical deviation, the RMSE was also calculated for each individual model output. In Table 7, the average concentrations for the measured and simulated concentrations along with the RMSE are highlighted.

The calibrated model was used for scenario analysis in order to define the set points for the JCS. For this investigation, the simulation was run until a steady state was achieved for at least 60 days, corresponding to approximately two sludge ages (SRT ≈ 27 d), before the results were generated.

Within the scenario analysis, the behaviour of the simulated mWWTP was investigated with respect to the significance of high TN concentrations originating from the iWWTP. The JCS exerts a direct influence on the aeration controller. The intermittent aeration of the mWWTP is controlled by the measured ammonium concentration in the aeration tanks. The relationship between the ammonium concentration and the calculated DO set-point within the controller is linear. In order to manipulate the slope of the linear set-point equation, it is necessary to define the minimum and maximum DO set-points. This was achieved by modelling different controller settings for different TN concentrations. However, it is necessary to control the aeration in order to ensure that the permitted discharge limits are still met. These limits are as follows: COD = 75 mg L⁻¹, TSS = 30 mg L⁻¹, TN = 8 mg L⁻¹ and TP = 1.5 mg L⁻¹. As the mWWTP has a low fraction of biologically available COD for denitrification due to the high fraction of iWW, it is important to avoid over-aeration. This was achieved by subsequently limiting the maximum DO concentration within the control concept to a specific threshold. When the TN concentration in the iWW stream exceeded normal levels, the DO concentration in the aeration tanks of the mWWTP was increased more rapidly to prevent nitrification limitations. Therefore, the JCS triggered a new slope for the aeration controller, when the TN concentration in the industrial influent of the mWWTP exceeded TN > 20 mg L⁻¹.

Figure 8 on the left compares the amount of energy used for aeration. The combination of an active JCS with optimized DO settings showed a decrease of 4% for the AE compared to the actual DO-setting. Figure 8 on the right compares the removal efficiency for COD (COD/E) and for nitrogen (N/E). For the COD/E index, an increase of 3% (0.8 kg kWh⁻¹) and for the N/E index, an increase of 4% (1.94 kg kWh⁻¹) were observed. The JCS, with its optimized settings, increased slightly the energy efficiency of the mWWTP in the model. However, the modelled increase in efficiency is very modest and within the range of possible simulation errors. Nevertheless, the result is interpreted as correct as further observations are included in the assessment. One reason for this discrepancy might be the already better efficiency in terms of pollutant removal (Table 7) of the simulated mWWTP compared to the actual mWWTP. For instance, the NO_x-N removal was 21% more efficient than in reality, while the NH_x-N removal remained unchanged. The air consumption of the real mWWTP is also 18% higher than that of the simulated mWWTP and the already low effluent concentrations of NH_x-N (0.4 mg L⁻¹) and NO_x-N (1.1 mg L⁻¹) also contributed to a decrease in the energy efficiency potential. Another reason might lie in the method used to determine the N/E index. By optimizing the DO controller, lower DO concentrations in the aeration tank were achieved in order to prevent the biologically available COD from being completely oxidized and thus supported the denitrification process. However, this led to a slight increase in the NH_x-N concentration, which was weighted 30 times in the calculation of the EQN, in contrast to the NO_x-N concentration, which was weighted 10 times only in the calculation. This led to a very low increase in the N/E index.

The performance of the JCS was evaluated after the first 3 months of operation in the real environment (Figure 9). Before the implementation of the JCS, the COD and TN loads in the influent were higher, while in the effluent, the COD and TN loads were lower and higher, respectively. While the COD concentrations in the effluent were nearly the same after the JCS was in operation, the TN concentration in the effluent decreased by a factor 1.5. After the start-up of the JCS, the COD/E and N/E ratios seemed to decrease from 0.56 kg COD (kWh)⁻¹ to 0.44 kg COD (kWh)⁻¹ and from 1.52 kg N (kWh)⁻¹ to 1.38 kg N (kWh)⁻¹, indicating the consumption of the same energy amount to eliminate a lower COD and N load with the JCS in operation. However, simultaneously, the average flow rate increased by a factor 1.3 after the implementation of the JCS. Hence, the higher flow rate led to a dilution of the wastewater and required more energy for pumping and mixing. Nevertheless, after the JCS was in operation, the specific energy consumption of the mWWTP decreased from 0.4 to 0.32 kWh/m³. Therefore, to exclude the effect of the higher flow rate during the operation of the JCS, the COD/E and N/E ratios were normalized using the flow rate. Hence, the normalized ratios indicate the treatment of the actual COD and TN load, but in a smaller water volume normalized to the flow rate before the JCS's implementation. The comparison of those ratios showed that the normalized COD/E ratio was on a similar level as the COD/E ratio of the operation without JCS, while the normalized N/E ratio was slightly higher with a factor of 1.18. Hence, the plant operation with the JCS in operation seemed to have a positive effect on the TN concentration in the effluent as well as on the energy efficiency for N elimination.

Those were first results and need to be confirmed with data of a full year of operation. Also, the normalization of the COD/E and N/E ratios was a rough estimation which did not consider any changes in environmental conditions. Shock loads of COD or TN were not observed in the testing time, also indicated by the lower COD and TN loads in the influent compared to the time period without JCS. The JCS was mainly designed to properly treat those events. The evidence therefore is still missing, but in a real environment the priority is to maintain ‘normal’ operating conditions. Therefore, it was only possible to test the system under normal operating conditions until now. Nevertheless, the increase in the mWWTP's energy efficiency under normal operation conditions by a factor of 1.18 is in its expected range. Bertanza et al. 2021 observed also an increase in the energy efficiency of their plant after the optimization of the aeration system. They saved 25% of the energy for aeration. Assuming that 60% of the total energy consumption of a plant refers to the operation of the aeration system, this would correspond to 15%, which is roughly in the range of our observation.

For two interconnected WWTPs, an industrial and a municipal WWTP, a JCS was developed. Using SIMBA# for modelling the mWWTP, the set-points for the JCS were determined and based on correlation analyses an empirical equation was derived to be used as an early warning system for high TN loads coming from the iWWTP and entering the mWWTP.

The JCS aimed to enhance the nitrogen elimination process in the mWWTP and to increasing its energy efficiency via a demand-driven aeration to avoid over-aeration. Therefore, an early warning system was developed using an empirical equation to predict the TN concentration coming from the iWWTP as early as possible. A stepwise warning was found to be useful 37 h, 70 and 0 min before the iWW enters the mWWTP. The TN prediction was highly dependent on the quality of the sensor measurements implemented in both WWTPs. The maintenance strategies were found to be sufficient, as good results were obtained for the correlation coefficients of higher than 0.5, even though 16 and 23 sensors were used to predict the nitrogen concentration for the stepwise warnings. In general, however, a minimum of four sensors is needed to measure flow, ammonium, nitrite/nitrate and TSS. As the BioTector is a quite expensive sensor, the TN prediction could replace the sensor. However, as the correlation was good, but not perfect and it depended on a large number of sensors, the confirmation by the BioTector provided additional reassurance.

The selected ASM3 + BioP biokinetic model in SIMBA# was adapted to the local conditions through a series of sensitivity analyses and calibration procedures. Especially the fractionation of the iWW turned out to be the bottleneck of the modelling part. In order to assess the goodness of model calibration and the model's predictive capability, by different statistical tests (R², NSE and L) were conducted. The overall model efficiency was calculated to be 0.73 for the selected model outputs (COD, TSS, TN, NH_x-N, NO_x-N, TP and PO₄-P), indicating that the model accurately predicted 73% of the real-world conditions on site. The calibrated model was used for scenario analysis in order to define the set points for the novel JCS. By varying the TN concentrations within the iWW entering the mWWTP the plant behaviour was investigated and the DO-setpoint for the aeration controller and JCS were improved. The results showed that the optimized settings for the aeration controller (DO_{min,optimised} = 1 mg O₂ L⁻¹, DO_{max,optimised} = 3 mg O₂ L⁻¹) with a subsequent limitation of the DO concentration to 1.5 mg L⁻¹ was recommended. To compare the actual and optimized settings, the plant-wide performance indices N/E and COD/E were computed. The optimized settings involving the JCS were related to an improved efficiency. The AE decreased by 4%, while the energy efficiency for COD and nitrogen removal increased by 3 and 4%, respectively.

To test the JCS in the real environment, the early warning for the TN load coming from the iWWTP (70 min before the iWW enters the mWWTP) together with the TN signal of the BioTector in the influent of the mWWTP was tested in the first 3 months. The evaluation showed an almost 2-fold lower TN concentration in the effluent and a 1.18 higher energy efficiency to eliminate nitrogen in the mWWTP under normal process conditions. However, these results still need to be confirmed considering a longer time period and under different environmental conditions during a full year of operation. In particular, the treatment of nitrogen shock loads coming from the iWWTP will provide deeper insights into the effectiveness of the TN prediction and the JCS. Therefore, the empirical equation to generate an early warning already 37 h before the iWW enters the mWWTP will be integrated into the JCS in the next step.

This system can be replicated at any other interconnected WWTPs which are also equipped with sensors for flow rates, ammonium, nitrate/nitrite and TSS. As long as the sensors are well maintained and provide reliable data, the BioTector sensor could even be replaced by the TN prediction. In addition, however, the modelling in SIMBA# is considered as necessary to provide valuable insights into the optimization potential of the aeration system and to define the DO setpoints of the JCS.

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

The authors declare there is no conflict.

The joint control system was implemented and the results of the presented study were obtained within the project “ULTIMATE” funded by the Horizon 2020 Research and Innovation Programme of the European Union under grant agreement number 869318.