Using plant data to estimate biodegradable COD fractions – case study kwaMashu WWTP

AD

Anaerobic digester

ADM1

Anaerobic Digester Model No. 1

AS

Activated sludge

BOD

Biological oxygen demand

COD

Chemical oxygen demand

DAF

Dissolved air flotation

FSA

Free and saline ammonia

IS

Inorganic solids

ISS

Inorganic suspended solids

MLE

Modified Ludzack-Ettinger

MLSS

Mixed liquor suspended solids

OHO

Ordinary heterotrophic organism

Ortho P

Orthophosphate

PST

Primary settling tank

PWM_SA

Plant-wide model – South Africa

RAS

Return activated sludge

SNL

Supernatant liquor from the secondary digester

SRT

Sludge retention time

SSE

Sum of squared errors

SST

Secondary settling tank

TKN

Total Kjeldahl nitrogen

TP

Total phosphorus

TS

Total solids

TSS

Total suspended solids

UCTADM1

University of Cape Town Anaerobic Digester Model No. 1

VFA

Volatile fatty acid

VSS

Volatile suspended solids

WAS

Waste activated sludge

WWTP

Wastewater treatment plant

COD_f

Filtered COD

COD_up

Unbiodegradable particulate COD

COD_us

Unbiodegradable soluble COD

C_xH_yO_zN_aP_b

Stoichiometric formula for a generic organic compound

f_codf

Soluble (filtered) fraction of influent COD

f_codup

Unbiodegradable particulate fraction of influent COD

f_codus

Unbiodegradable soluble fraction of influent COD

f_cv,i

COD to VSS ratio of component i

f_fsa

FSA fraction of TKN

f_iss

Influent ISS to TSS ratio

f_n

Nitrogen content per gram organic

f_ns,codup

Fraction of influent unbiodegradable particulates in the settled sewage

f_ns,iss

Fraction of influent ISS in the settled sewage

f_nsPST

Fraction of influent TSS in the settled sewage

f_OHO

OHO fraction of influent COD

f_p

Phosphorus content per gram organic

f_tknf

Soluble fraction of TKN

f_tpf

Soluble fraction of TP

f_tss

Influent TSS to COD ratio

f_vfa

VFA fraction of influent COD

H

Hessian matrix

N_p

No. of fitting parameters

N_d

Number of measured data points

P

Vector of fitting parameters

R_AD

Hydraulic retention time of the anaerobic digesters

S_F

Soluble fermentable organics

S_NH

Ammonium

S_PO4

Phosphate

S_U

Soluble unbiodegradable organics

S_VFA

Volatile fatty acids

TKN_f

Filtered TKN

TP_f

Filtered TP

V_AD

Anaerobic digester volume

V_AS

Activated sludge reactor volume

X_BInf

Biodegradable particulate influent organics

X_ISS

Inorganic suspended solids

X_OHO

Ordinary heterotrophic organisms

X_UInf

Unbiodegradable particulate influent organics

σ

Covariance matrix

When the configuration of an existing plant is to be changed, the most critical part of the model calibration is the influent wastewater fractionation. Furthermore, due to the complexity of the systems involved, model calibration protocols, for example, Reiger et al. (2012), typically involve the sequential calibration of the various subsystems starting with the influent characterization. As a result, errors in the influent characterization are propagated through the other calibration steps (Grau et al. 2007).

The constantly varying characteristics of wastewater make experimental determination of an adequately representative set of components, using protocols such as those recommended by the IWA Guidelines (Reiger et al. 2012), difficult, time-consuming and expensive, which constitutes significant barriers to the adoption of modelling by many municipalities. Biodegradable organic fractions in raw and settled sewage are typically determined via biological oxygen demand (BOD) measurements (Hulsbeek et al. 2002) or respirometric methods (Vanrolleghem 2002). Both of these methods take days to get results for a single sample, and many municipalities simply do not have the equipment or experienced personnel to undertake these types of characterization studies. Furthermore, translating laboratory results to full scale WWTP plants can be quite challenging due to important differences between the two types of systems (Sin et al. 2005).

With the exception of the soluble unbiodegradable fraction of the influent (f_codus), which is typically assumed to equal the soluble COD in the secondary effluent (Ekama & Wentzel 2008), the probabilistic fractionator has no way of distinguishing biodegradable and unbiodegradable fractions. Therefore, the influent COD fractionation parameters (f_codf, f_codp, f_vfa) are usually best guesses based on the available literature.

In the PWM_SA model (Ekama 2009; Ikumi et al. 2015), each of the organic components is assigned an elemental composition C_xH_yO_zN_aP_b, from which other important component properties can be calculated, including its COD to VSS ratio (f_cv), and nitrogen and phosphorus content per gram (f_n and f_p). The biodegradable influent particulate component X_BInf is assumed to represent a mixture of proteins, carbohydrates and lipids. These are separate components in some other models, e.g. Anaerobic Digestion Model No1 (ADM1) (Batstone et al. 2002); however, since this type of characterization data is very seldom available, they are lumped together in the University of Cape Town Anaerobic Digestion Model No 1 (UCTADM1) (Sötemann et al. 2005), which is used to represent anaerobic digestion in PWM_SA. However, the ratio of proteins, carbohydrates and lipids in raw sewage depends on the diet of the population, how much food waste is entering the sewer, and the impact of industrial and commercial wastewater contributions. Therefore, the average composition of the biodegradable particulate fraction may vary significantly between different treatment works. Table 1 shows the elemental compositions for the X_BInf model component for various implementations of the plant-wide model compared with ADM1 components for proteins, carbohydrates and lipids (Batstone et al. 2002). The listed compositions were all experimentally determined at some point, but it is often not clear which plants the samples they are based on came from and how representative they are of conditions in other plants.

Component compositions are fixed in the probabilistic fractionator, and errors in the composition of X_BInf will affect the fitting of the model components to the measured data. In the plant-wide model, the assumed composition of X_BInf also affects the predicted release of nutrients in the anaerobic digester, which has implications for nutrient removal.

In this study, we explore the possibility of extending the probabilistic fractionator by including routine process measurements with the routine influent measurements, and coupling the fractionator with a simplified plant-wide steady state model, so as to obtain a more accurate fractionation of the influent. This technique could make wastewater treatment modelling accessible to a wider range of municipalities.

The detailed plant layout is shown in Figure 3. After screening and de-gritting, raw sewage is pumped to two primary settling tanks (PSTs). The raw sludge from the PSTs is pumped to two gravity thickeners and the thickened primary sludge is then pumped to the two primary anaerobic digesters. The five-cell secondary digester separates the digested solids from the supernatant liquor (SNL) under gravity. The digested sludge is pumped to the dewatering plant while the SNL is pumped back to the head of works. The primary settler effluent (settled sewage) flows to a Modified Ludzack-Ettinger (MLE) activated sludge process consisting of a four-lane aeration unit followed by eight secondary settlers (SSTs). The secondary effluent is discharged to a series of three maturation ponds and the pond effluent is chlorinated prior to discharge to the Umhlangane River.

The waste activated sludge (WAS) is drawn from the return activated sludge (RAS line). The WAS is pumped to the DAF units for thickening. The DAF sludge is sent to the dewatering plant while the sub-natant flows back to the head of works.

The dewatering plant consists of eight screw presses with polymer dosing. The filtrate from dewatering is returned to the head of works while the dewatered sludge is supposed to be trucked to offsite disposal. However, during the study period, sludge was not being removed from the site on a regular basis due to problems with the disposal contracts and consequently, the operating staff were forced to reduce the sludge wasting rate. As a result, the plant was often being operated at sludge ages of greater than the 12 to 20 days the plant was designed for. Furthermore, since the wasting rate data had been lost when the data storage system was damaged, the actual SRTs during the study period were not known.

Routine monitoring data for 2018 were provided by eThekwini Water and Sanitation (EWS) from their laboratory information system and plant operating records. June and July were a period of little rain and relatively stable plant operation, so these data were selected for modelling. Table 2 shows the plant sampling schedule for the sampling points shown in Figure 3. Influent flowrate was the only measurement made every day. Additional plant operating data recorded by the SCADA system had been lost when the data storage system was damaged by power disruptions.

In addition to the data listed in Table 2, sludge cake production data was available as monthly totals.

A simplified version of the plant-wide steady state model (Ekama 2009) was set up to track the organic and particulate components through the plant, from the raw sewage to the secondary effluent and sludge filter cake. The purpose of the model was threefold:

• (1)

  To estimate the missing internal flows; specifically: the flow to the anaerobic digesters (ADs) and the waste activated sludge (WAS) flow, which determine the hydraulic retention (R_AD) time of the ADs and the sludge retention time (SRT) of the activated sludge (AS) plant, respectively.

• (2)

  To estimate the biodegradable organic fractions of the influent based on the overall reduction in solids and COD.

• (3)

  To estimate f_cv and f_n of the biodegradable influent particulate component X_BInf.

The influent sub-model for the simplified steady state model includes only COD and particulate components, i.e. the components shown in Figure 1.

In the simplified steady state mode, the PST sub-model lumps together the primary settlers and primary sludge thickeners such that its outputs are the settled sewage and the feed to the anaerobic digesters. The PST sub-model parameters are f_setsew, f_nsPST, f_ns,codup and f_ns,iss, where (1 − f_setsew) is the flow to the anaerobic digesters as a fraction of the plant influent flow and determines the anaerobic digester hydraulic retention time (R_AD), while f_nsPST, f_ns,codup and f_ns,iss are the settled sewage TSS, X_UInf and ISS fluxes as a fraction of their corresponding fluxes in the influent.

= digester volume, m³

Raw flow = raw sewage flow to the primary settlers, m³/d

where

SRT = sludge age, d

= activated sludge reactor volume, m³

MLSS = mixed liquor suspended solids, kg/m³

Wasting rate = DAF sludge solids flux fed to sludge dewatering, kg/d

Effluent solids flux = secondary effluent flux, kg/d

Note that is assumed that the impact of recycling the secondary digester supernatant, DAF underflow and dewatering plant filtrate on the flow and solids fluxes have a negligible effect on the overall flow and solids balances. Since the simplified steady state model does not include nutrient removal, the impacts of these internal recycles on the nitrogen and phosphorus fluxes are not considered.

The simplified model and optimization were coded in R (R Project 2021). The standard R non-linear optimization routine optim was used for the parameter regression.

The plant measurements used in the fitting procedure and initial set of fitting parameters are listed in Table 3. Outliers in the raw measurements were removed using the interquartile range and fences method.

Some of the parameters listed in Table 3 can be estimated directly from the available raw plant data, using simple mass balances on individual treatment units where necessary, specifically, the fitted raw COD, secondary effluent solids flux, f_setsew, f_tss, f_codus and f_ns,PST. However, the results obtained this way may be subject to errors and systemic biases, e.g. due to data sparsity, how and when samples are collected and how averages are calculated. Furthermore, uncertainties increase when parameter values are interdependent. For the above mentioned parameters, the raw data was used to calculate the initial estimates and expected ranges. The initial estimates for the fitted raw COD and effluent solids flux were calculated as the average of the available measurements. The fraction of the raw flow going to the activated sludge plant (f_setsew) was estimated from a solids balance on the PST using the average raw and settled TSS and thickened primary sludge %TS. The initial estimate of f_tss was the ratio of the average measured raw COD and TSS. The initial estimate of the soluble unbiodegradable COD fraction (f_codus) was calculated from the average secondary effluent COD and TSS as described by Ekama & Wentzel (2008). The non-settlable influent solids fraction (f_ns,PST) was estimated as the ratio of the average measured settled sewage and raw TSS.

The upper and lower fitting bounds of the wasting rate corresponded to SRT values of 12–30 days. Initial estimates and ranges for the settled sewage fractionation parameters f_ns,codup and f_ns,iss were specified based on typical values of 0.2–0.3 and 0.1–0.2, respectively, reported by Wentzel et al. (2006). Initial estimates and bounds for the other fractionator parameters were selected based on values in Henze & Comeau (2008) and typical values for various South African wastewater treatments included in a plant-wide steady state design (PWSSD) tool developed by Wu (2015).

wi= weight of measurement i

x_i,j = jth observation of measurement i

X_i = predicted steady state value of measurement

To estimate the significance of the regressed parameters P, the parameter covariance matrix was calculated from the inverse of the Hessian matrix H at the optimum point where (Vanrolleghem et al. 1995). H is a measure of the local model sensitivity to the parameters at the optimum point, and is an output of the optim routine, and ⁠, where N_d is the number of measured data, and N_p is the number of regressed parameters. The significance of parameter i was expressed as its standard error ⁠: the ratio of its estimated standard deviation to its value. This approximate procedure implies linearizing the model in the vicinity of the optimum point.

The results of the parameter regression are summarized in Tables 4 and 5. Where a parameter was found to be identifiable in the analysis, the best-fit value is listed. Table 4 also shows the % change from the initial values calculated from the raw plant data.

It was possible to obtain best fit values for the wasting rate, flow and solids flux split between the anaerobic digesters and AS plant ((1 − f_setsew) and f_ns,PST in Table 4) and the X_BInf composition parameters (f_cv,XBinf and f_n,XBinf in Table 5). The best fit values for f_cv,XBinf and f_n,XBinf corresponded most closely to biodegradable particulate compositions reported by Gaszynski (2021) and these were used to estimate the X_BInf stoichiometry used in the probabilistic fractionator and WEST model. The SRT corresponding to the best fit wasting rate was 20 days, which was at the limit of the design range of 12–20 days.

For the raw sewage fractionation, f_codus and f_codup (Table 4) were identifiable, but f_codf (Table 5) was not. This means that the multi-parameter regression could not estimate a unique ratio of soluble and particulate biodegradable COD that explained the observed data. For the settled sewage fractionation, f_ns,iss and f_ns,codup (Table 5) were not identifiable. Inorganic solids (%IS) were measured only in the dewatering filter cake at kwaMashu, and the regression could not determine in which ratio the organic and inorganic inert particulates follow the two flow paths available in the simplified model.

Parameters that were excluded from the regression were fixed at their initial estimates, and these fixed values will affect the best-fit values of the identifiable parameters, and the subsequent calibration of the detailed model (Brun et al. 2002). However, the results of the analysis also point to a simple and practical solution to the problem: of the three parameters that were not observable from the available data, only f_ns,codup requires specialized analytical techniques for its measurement. Any wastewater laboratory should be able to conduct measurements for filtered COD to reduce the uncertainty in the estimation of the parameter f_codf. Similarly, the estimation of f_ns,iss can be improved by measuring the %IS of the primary, digested or waste sludge or the ISS of the settled sewage. In fact, it is quite common to measure the inorganic or ash content of sludge fed to anaerobic digesters.

= mass of X_Uinf in the reactor

= flux of X_Uinf fed to the reactor

While the wasting rate and effluent solids flux, which are used in calculating the SRT, are identifiable in the regression, the corresponding flux of X_Uinf calculated by the model will be very sensitive to variability in the regressed values of these parameters. In fact, when the value of the wasting rate was manually fixed in the regression, the parameter f_ns,codup became identifiable. Since the SRT is a critical operating parameter, the sludge wasting rate ought to be part of routine plant records, but was not available in this case. The results of the identifiability analysis suggest that with reliable sludge wasting rate data and some additional inorganic solids measurements, even the fractionation of the settled sewage at kwaMashu WWTP can be estimated using the simplified steady state model.

Note that, in principle, the multivariate parameter estimation could be carried out in WEST using the full plant-wide model; however, it is computationally intensive, and will take many hours to converge to a solution. The simplified steady state model provides a convenient and much faster way (a few minutes at most) to find the best fit parameters for the COD and solids balances. This then reduces the number of parameters that require tuning in the detailed model, for example, relating to the nitrogen removal model.

As noted above, inorganic solids (IS) were only measured in the belt press filter cake at kwaMashu WWTP. While ISS is usually only a small fraction of the TSS of raw wastewater, it makes up a substantial part of the overall filter cake production (∼40% in Figure 8), which is why including ISS in the raw wastewater characterization is important.

In Figure 9, the influent COD (S1 in Table 2 and Figure 3), effluent (S20) and sludge cake (S22–29) plus the oxygen uptake and methane production represent the overall plant COD balance, while the primary (S3), digested (S6–7) and DAF thickened waste activated sludge (S18–19) are internal streams. The potential for solids and COD reduction is determined by the biodegradable content of the plant influent, while the extents of the reductions are functions of the sludge age of the AS plant and the hydraulic retention time of the anaerobic digesters, which is why the R tool is able to estimate their values from the observed solids and COD reduction.

As shown in Figure 9, all of the soluble biodegradable COD is consumed while the unbiodegradable soluble fraction mostly exits in the effluent. The bulk of the biodegradable influent particulate is fed to the anaerobic digester in the primary sludge and consumed while the unbiodegradable influent particulate plus some unbiodegradable residue of biomass endogenous respiration ends up in the sludge cake. The biodegradable particulate fraction of the DAF sludge consists of biomass (OHOs) rather than influent organics but makes up only 43% of the COD compared to 68% in the influent.

Particulate COD in the effluent is a small fraction of the total COD balance in Figure 9 but makes up 53% of the secondary effluent COD. This was due to the accumulation of solids in the reactor as a result of the long solids retention time leading to the overloading of the secondary clarifiers.

We have developed a modelling tool which estimates COD fractionation and other mass balance parameters from plant data using a simplified steady state model of a WWTP. This can greatly facilitate setting up models of treatment plants using routine monitoring data even when some critical operating data, such as the sludge wasting rate, is missing or unreliable. While not all the fractionation parameters investigated were identifiable from the available data in this case study, the results of the significance analysis pointed to practical ways in which data collection could be improved to provide data that is useful for modelling, process optimization and design. Thus, for kwaMashu, it was concluded that all the parameters investigated would likely have been identifiable if the sludge wasting rate records had been available and the routine plant measurements had been augmented by some measurements of filtered raw COD, settled wastewater ISS and/or the IS of the primary, digested or waste activated sludge, as detailed in the discussion. The influent fractionation and calibrated plant model of kwaMashu WWTP will be used by the eThekwini municipality as a benchmark against which the current plant performance is evaluated, and in the design of the proposed upgrade.

This research was funded by the DANIDA fellowship center through the project ‘Evaluation of Resource Recovery Alternatives in South African Water Treatment Systems (ERASE)’ (Contract-No: 18-M09-DTU) and by eThekwini Water and Sanitation.

The probabilistic influent fractionator code and the R implementation of the simplified steady state WWTP model is available on request. To express interest please contact Chris ([email protected]) and Barbara Brouckaert ([email protected]) at the WASH R&D Centre at the University of KwaZulu Natal (South Africa).

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

The authors declare there is no conflict.