15
9 th European Waste Water Management Conference 12-13 October 2015, Manchester, UK www.ewwmconference.com Organised by Aqua Enviro Limited THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE ACCURACY OF WASTEWATER TREATMENT WORKS MODELS Ashagre, B. B. 1* Fu, G. 1 Davidson, K. 2 Butler, D. 1 1 University of Exeter, UK, 2 Scottish Water, UK * Corresponding author [email protected] Abstract In order for wastewater treatment works (WwTW) models to be used with confidence for active control it is important to perform model calibration to accurately represent its performance. Data is usually a limitation in achieving high levels of calibration since it can be costly and time consuming. Thus it is important to carefully assess the minimum data requirement and determine the need for further data collection. This study assesses how far model performance can be improved by considering more frequent and perhaps costly monitoring .This is achieved by comparing simulated quality indicators with a measured dataset after performing sensitivity analysis to identify parameters to which the model is most sensitive. WwTW models are tested using three different datasets of increasing number of quality variables. Calibration accuracy, as measured using R 2 and RMSE goodness-of-fit tests, increases for TSS and NH3-N concentrations as compared to the baseline for scenarios two and three, albeit still with low absolute values. In this case study, the results indicate the importance of characterising influent wastewater organic matter and nitrogen concentrations to reduce prediction uncertainty and help build confidence in the use of models for active control. Keywords Active control, calibration, data reconciliation, wastewater treatment Introduction Future demand for model based control of wastewater treatment works is expected to increase. In Europe, including the UK, models are mostly a research subject whereas in other parts of the world like North America Wastewater Treatment Works (WwTWs) are predominantly used as an engineering tool in practice (Hauduc et al. 2009). This is now changing in the UK, and water utilities have started to incorporate WwTW models in decision making, process control and optimisation. According to UKWIR (2013) WwTW models are now being used for advanced process control and this practice is expected to increase significantly in the future due to tighter regulations and the potential of this approach to save energy, chemical usage and greenhouse gas emissions. One of the challenges is the availability of data and their quality. High quality data is crucial for the effective use of WwTW models. The reliability of model results is strongly linked to the amount of the data used to set up and calibrate the model (Rieger et al. 2010). A carefully designed and collected dataset can reduce time for the subsequent modelling study and also can increase the confidence in using the model for practical application. In addition, data scarcity and low quality data can distort the simulation results and increase the chance of faulty conclusions, which might lead to very expensive decisions and/or could cause breaching of licenses. Historical data can be used to understand the long-term behaviour of the treatment works. Dynamic modelling for control purposes requires high resolution spatial and temporal data, which includes sub- daily monitoring of various parameters. Stoichiometric/kinetic data can also be monitored to accurately estimate the model parameters. However, this can be costly and demand experience to achieve all these datasets. Thus, it is important to determine what level is sufficient for the model- based study.

THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

  • Upload
    others

  • View
    3

  • Download
    0

Embed Size (px)

Citation preview

Page 1: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE ACCURACY OF

WASTEWATER TREATMENT WORKS MODELS

Ashagre, B. B.1* Fu, G.1 Davidson, K.2 Butler, D.1

1University of Exeter, UK, 2Scottish Water, UK *Corresponding author [email protected]

Abstract

In order for wastewater treatment works (WwTW) models to be used with confidence for active control

it is important to perform model calibration to accurately represent its performance. Data is usually a

limitation in achieving high levels of calibration since it can be costly and time consuming. Thus it is

important to carefully assess the minimum data requirement and determine the need for further data

collection. This study assesses how far model performance can be improved by considering more

frequent and perhaps costly monitoring .This is achieved by comparing simulated quality indicators

with a measured dataset after performing sensitivity analysis to identify parameters to which the

model is most sensitive. WwTW models are tested using three different datasets of increasing number

of quality variables. Calibration accuracy, as measured using R2 and RMSE goodness-of-fit tests,

increases for TSS and NH3-N concentrations as compared to the baseline for scenarios two and

three, albeit still with low absolute values. In this case study, the results indicate the importance of

characterising influent wastewater organic matter and nitrogen concentrations to reduce prediction

uncertainty and help build confidence in the use of models for active control.

Keywords

Active control, calibration, data reconciliation, wastewater treatment

Introduction

Future demand for model based control of wastewater treatment works is expected to increase. In

Europe, including the UK, models are mostly a research subject whereas in other parts of the world

like North America Wastewater Treatment Works (WwTWs) are predominantly used as an

engineering tool in practice (Hauduc et al. 2009). This is now changing in the UK, and water utilities

have started to incorporate WwTW models in decision making, process control and optimisation.

According to UKWIR (2013) WwTW models are now being used for advanced process control and

this practice is expected to increase significantly in the future due to tighter regulations and the

potential of this approach to save energy, chemical usage and greenhouse gas emissions. One of the

challenges is the availability of data and their quality.

High quality data is crucial for the effective use of WwTW models. The reliability of model results is

strongly linked to the amount of the data used to set up and calibrate the model (Rieger et al. 2010).

A carefully designed and collected dataset can reduce time for the subsequent modelling study and

also can increase the confidence in using the model for practical application. In addition, data scarcity

and low quality data can distort the simulation results and increase the chance of faulty conclusions,

which might lead to very expensive decisions and/or could cause breaching of licenses.

Historical data can be used to understand the long-term behaviour of the treatment works. Dynamic

modelling for control purposes requires high resolution spatial and temporal data, which includes sub-

daily monitoring of various parameters. Stoichiometric/kinetic data can also be monitored to

accurately estimate the model parameters. However, this can be costly and demand experience to

achieve all these datasets. Thus, it is important to determine what level is sufficient for the model-

based study.

Page 2: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

The importance of monitoring wastewater within the WwTW has been suggested to be crucial both for

design and modelling purposes (Gernaey et al. 2006, Melcer 2003, Metcalf & Eddy 2004, Rosén et al.

2003), but the question is what level of data is sufficient to have reasonable confidence in the model

to be used for active control purpose. The aim of this paper is therefore to investigate the model

performance using different levels of dataset. Hence, two scenarios were used in this study, in

addition to a baseline scenario, each with a different level of data availability in order to set up and

calibrate the model. The differences in model performance among the scenarios are used to assess

the benefit of using the specific dataset considered in the corresponding scenario.

Methodology

The case study

The case study wastewater treatment works has a design capacity of 15,000 P.E. (Population

Equivalent) with current load of 16,000 PE or a throughput of 89L/s. The incoming flow enters the works

through an inlet chamber and passes to a band screen. Screened sewage passes under a weir and

flows below 3 x DWF (90L/s) flow to a grit trap and any flow above 3 x DWF bypasses to the storm

tanks. Following screening and de-gritting, the sewage gravitates into a wet well from where it is pumped

by 3 duty/standby/assist pumps to the oxidation ditch for secondary treatment. Based on the water level

in the inlet wet well at the pumping station, water will return from the storm tanks and the liquor buffer

tank back to the wet well through gravity. The return flows combined with online sewage flow are then

pumped to the oxidation ditch which has a capacity of 2980m3 and a maximum depth of 4m. The

schema of the whole treatment work is presented in Figure 1.

In this study, the IWA Benchmark Simulation Model 2 (BSM2) was used to model the WwTWs under

consideration. BSM2 uses the MATLAB® platform and was built using Simulink®. By modifying the

existing typical wastewater scheme into BSM2, the scheme and processes of the existing wastewater

treatment plant were represented.

Biochemical processes in oxidation ditches have previously been modelled using ASM1 (e.g. (Abusam

and Keesman 1999, Abusam et al. 2003, 2001) and it has been suggested that 10 continuously stirred

reactors are sufficient for adequate representation of the aeration configuration (Abusam and Keesman

1999). In this instance, twelve reactors were found to be the required amount to represent the

configuration of blowers and the influent and effluent flows from the ditch.

From the oxidation ditch, the biologically treated sewage (the 'mixed liquor’) passes into two final

settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel

from where it gravitates to a sampling chamber and is discharged into the nearby river.

The two final settlement tanks are modelled as a single tank with a surface area of 928 m2 and volume

of 1628 m3, using the a ten-layer model, based on Takács et al. (1991). This assumes no biological

activity occurs in the final settlement tanks.

Dataset

Data monitoring locations are shown in Figure 1 and the data availability corresponding to these

points is discussed in this section. The model was run to simulate 200 days; the first 100 days were

used to warm up the model, the next 100 days were used for calibration purposes.

The flow data were used to cross check control philosophies collected from operational manuals and

interview with operators, to complete the hydraulic balance and assess the hydraulic performance of

Page 3: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

the work (the hydraulic performance of the WwTW is not the focus of this paper and won’t be

discussed further) flow to the WwTW, to inlet well, to oxidation ditch from inlet well, RAS and SAS

flow, and final effluent.

Water level data were used to estimate the return flow rate from storm tanks to inlet wet well. The inlet

wet well level and the flow from the inlet well to the oxidation ditch were used to create a correlation

between flow from pumping station and water level in the inlet wet well. This correlation was used to

model the flow from inlet wet well to oxidation ditch based on water level in the wet well. The 15

minute time step influent flow was used in the dynamic influent model input dataset.

Quality data were available on a daily basis, which were measured using a daily composite sample.

These measured quality data include; biochemical oxygen demand (BOD), total suspended

solids(TSS), ammonia (NH3-N or SNH), pH, temperature, soluble reactive phosphorus (SRP), and total

phosphorus (P). Most of these measured quality data from the influent wastewater, except SRP and

P, were used in creating dynamic influent flow concentrations which are discussed in the baseline

model calibration section.

Figure 1 shows location of points where data was available. The available information at each location

varies both in content and temporal resolution. Points 1, 4, and 7 have flow data at 15 minutes time

step and daily average data on wastewater quality. Data points 8, 11, and 18 have only daily average

quality data. Some only have flow data at 15 minutes time step, this includes; data points 3,5,12, and

13. Imported wastewater and sludge from different works are deployed at point 2 and 16 respectively.

Volume and time at which the wastewater/sludge imported was available. Data on volume and time of

surplus activated sludge (SAS) removal is available, point 14. Data on wells’ and tanks’ level at point

9, 10, 15, 17, and 19 were available at 15 minutes time step. Mixed liquor (MLSS) level and dissolved

oxygen (DO) within the oxidation were measured at 15 minutes time step, point 6.

Figure 1: Scheme of the wastewater treatment work

Screens

Grit trap

Inlet wet well

Oxidation

ditch

Storm tank 1 Tanks

Final settlement tank 1

Final settlement tank

2

Final effluent

chamber

Screen to skip

Combined sewer

network

Tankered imports

Sludge dewaterin

g tank

Super-natant pump well

Decanted liquor

chamber

Sludge holding

tank Centrifuge

Liquor buffer tank

Sludge cake

Final

effluent

SAS impor

1 3 4

5 6

10

9

18

8

17

14 13

12

15

7

16

19

11

2 Backwash

water

Wastewater

RAS/SAS/Sludge

Liquor return

Storm water

Screenings/grit

Potable water

Grit to skip

RAS

SA

S

Page 4: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

Model calibration

In order to assess what data set plays a crucial role in improving model calibration accuracy in

simulating TSS and NH3-N, three different scenarios were assessed. The first scenario, which is the

baseline, was set up using available data but it was assumed that MLSS and DO data were not

available and were not used in model calibration. In the second scenario, which is the MLSS and DO

scenario, the 15 minutes time step data was known and were used for model calibration. In the last

scenario, a different approach in characterising influent wastewater was used. The details on each of

the scenario are discussed below.

Baseline

The WwTW model contains information on operational procedures, existing set points for control

purposes, and infrastructure details such as dimensions, maximum capacity and working capacity. In

addition, commonly monitored parameters such as flowrates which were monitored at a 15 minute

time step at the influent and effluent points, and daily average influent and final effluent measured

quality indicators (TSS and NH3-N) were used. TSS and NH3-N were measured using a daily

composite sample for each day of the whole calibration period. Converting the daily data into sub

daily data and fractionation of COD was carried out by imitating the sub daily pattern of pollutants

from the BSM2 model and using fractions used in Gernaey et al. (2005),.

The model inputs that are required to characterise influent wastewater are given in Table 1.

Table 1: BSM2 model inputs: Influent wastewater characteristics

Soluble COD Particulate COD Nitrogen Others

Soluble inert organic matter (SI) Particulate inert organic matter (XI)

Nitrate and nitrite nitrogen (SNO)

Oxygen (SO)

Readily biodegradable substrate (SS)

Slowly biodegradable substrate (XS)

NH3 nitrogen (SNH) Alkalinity (SALK)

Active heterotrophic biomass (XB,H)

Soluble biodegradable organic nitrogen (SND)

Total Suspended solid (TSS)

Active autotrophic biomass (XB,A)

Particulate biodegradable organic nitrogen (XND)

Flow rate

Particulate products arising from biomass decay XP

Temperature

The influent wastewater concentration of XB,A, particulate products arising from biomass decay (XP),

oxygen (SO), and nitrate and nitrite nitrogen (SNO) were assumed to be zero (Jeppsson et al. 2007).

The particulates concentration; particulate inert organic matter XI, slowly biodegradable substrate XS,

and active heterotrophic biomass XB,H were estimated from the measured daily average TSS and

fractions used in Gernaey et al. (2005).

Readily biodegradable substrate SS is the difference between the soluble inert organic matter (SI) and

the total soluble COD (bCOD). bCOD was estimated from measured biochemical oxygen demand

(BOD) and the ratio between ultimate biological oxygen demand (UBOD) and BOD; for typical

domestic wastewater UBOD⁄BOD is equal to 1.5 (Metcalf & Eddy 2004). Using synthesis yield

coefficient of 0.67(Jeppsson et al. 2007), bCOD was estimated to be equivalent to 75% of BOD. Since

the measured BOD has considerable gaps, correlation between BOD and TSS was used to generate

a complete BOD dataset. The correlation in Figure 2 showed a significant correlation between BOD

and TSS with a regression coefficient R2 equal to 0.615.

Page 5: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

Figure 2: Correlation of measured BOD and TSS

The soluble inert organic matter, SI, was estimated to be 21mg/L which was based on the ratio

between total average of measured TSS and total average of TSS in Jeppsson et al. (2007).

For this case study, the raw influent diurnal ammonia concentration (SNH) was estimated based on the

ratio between daily measured SNH and the daily average of SNH from BSM2 dynamic influent data.

Due to data paucity on total Kejldhal nitrogen, the particulate organic nitrogen XND were estimated by

using BSM2 XND and the ratio of daily measured average SNH of the WwTW and the daily average of

SNH of BSM2. SND was calculated using the fraction; fraction of organic nitrogen that is soluble and

degradable (𝑆𝑁𝐷_𝑓𝑟 = 0.247) of the total organic nitrogen, and fraction of organic nitrogen that is

particulate and degradable (𝑋𝑁𝐷_𝑓𝑟 = 0.753) (Gernaey et al. 2005).

Using MLSS and DO data

In this case the modeller has data on Mixed Liquor Suspended Solid (MLSS) and Dissolved Oxygen

(DO) in the oxidation ditch at 15 minutes time steps. The MLSS data were used to further calibrate the

SAS flow control operation and the DO data were used to calibrate the DO control feedback loop

used in the model. MLSS and DO at 15 minutes time steps were used to understand the DO control

and SAS removal better. The SAS control was revised and its removal operation was triggered based

on the MLSS level in the oxidation ditch. The SAS removal in the baseline was set up to be instigated

once every other day to discharge 200m3 of SAS into the sludge holding tank. In this scenario, the

removal operation was continuous but the SAS removal would not be instigated if MLSS is lower than

3000 mg/L. The DO control in the baseline was set up using a PID (Proportional Integral Derivative)

controller to a set point of 0.75mg/L. In this scenario DO control is set up the same as the baseline but

the blower will switch off if the DO level in the oxidation ditch is higher than 2.25mg/L. In addition, the

sensor in the baseline line was ‘type A’ which is close to an ideal sensor but in this scenario, by

observing the measured DO level in the oxidation ditch, sensor type B was selected which is a

intermittently measuring sensor with a time delay of 10 minutes and a measuring interval of 5 minutes.

Using a phenomenological influent generator to estimate pollutant concentrations

In this case, modelling WwTWs has become standard practice in order to implement automated

controls and optimisation of waste water treatment works’ operations. Water service providers accept

that a detailed characterisation of the wastewater organic matter is crucial in modelling wastewater

treatment works. This can be achieved by fractionating the total COD into fractions with different

microbiological properties (Henze et al. 2000).

y = 0.3945x + 120.33R² = 0.615

0

100

200

300

400

500

600

700

800

900

1000

0 500 1000 1500 2000

BO

D m

g/L

TSS mg/L

Page 6: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

Due to data limitations the benefit of characterising wastewater influent to the WwTW was assessed

in this study by generating pollutant patterns using a phenomenological wastewater generator model

(Gernaey et al. 2011). This produced wastewater influent data for the BSM2 taking into account

process control and optimisation. Since the model is ‘open source’, users can modify the original

structure in order to customise the model for a specific study area in hand. Considering that this

approach is flexible, and easily linked to BSM2, it was used in this scenario to characterise

wastewater influent to the WwTW.

Sensitivity analysis, calibration and model performance

It is essential to calibrate developed models due to the uncertainty in model structure, model

parameters or underlying uncertainties in the (incomplete) measured data. In complex process-based

models, calibration can be a complex and time-consuming process due to high numbers of model

parameters and the poor identifiability associated with large parameter sets. Sensitivity analysis helps

to identify the parameters that most influence the model output and hence, reduce the number of

parameters to consider for calibration purposes.

The sensitivity is carried out using one-factor-at-a-time (OAT) approach in order to identify the most

significant parameters for model calibration purposes. Since the focus was to identify the most

sensitive parameters for the purpose of model calibration, evaluation criteria that indicate the WwTWs

model’s performance in simulating TSS and NH3-N were used. The evaluation criteria assess how

close or far simulated TSS and NH3 are from measured TSS and NH3-N respectively both in terms of

pattern and residual error. This was represented by the use of statistical tests, the Root-Means-

Square-Error (RMSE),𝑹𝑴𝑺𝑬= √∑(𝒚𝒊−𝒙𝒊)𝟐

𝒏 1, and

coefficient of regression R2, 𝑹𝟐=∑[(𝒙𝒊−�̅�)×(𝒚𝒊−�̅�)]𝟐

[∑(𝒙𝒊−�̅�)𝟐]×[∑(𝒚𝒊−�̅�)𝟐] 2.

𝑹𝑴𝑺𝑬 = √∑(𝒚𝒊−𝒙𝒊)𝟐

𝒏 1

Where; 𝑦𝑖 is simulated daily average final effluent quality indicator (TSS or NH3-N), 𝑥𝑖 is measured

daily average (TSS or NH3-N), 𝑛 is number of data point or days used for analysis.

𝑹𝟐 =∑[(𝒙𝒊−�̅�)×(𝒚𝒊−�̅�)]𝟐

[∑(𝒙𝒊−�̅�)𝟐]×[∑(𝒚𝒊−�̅�)𝟐] 2

Where; �̅� is the average of the measured daily average values, �̅� is the average of the simulated daily

average values.

The R2 and RMSE of the uncalibrated model were calculated first. The percentage change in the

value of these evaluation parameters due to change in value of sensitivity parameter was assessed in

order to identify the most sensitive parameter, for example, percentage change in R2 in simulating

TSS at literature average and upper bound of b_H.

The wastewater treatment parameters listed (kinetic and stoichiometric parameters) for the ASM

(Activated Sludge model) were adopted from Jeppsson et al. (2007), see Table 2. Details on

parameters related to settlement tanks can be found in the studies Jeppsson et al. (2007) and Takács

et al. (1991). Additional parameters which are reported in other studies as being sensitive but

dependent on the operational philosophy of the plant for example oxygen transfer rate are not

Page 7: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

assessed in the sensitivity analysis. Instead, they were manually adjusted and calibrated based on

available information.

Table 2: Parameters used for sensitivity analysis

Para-

meters Description

Literature

average

Lower

bound

Upper

bound Remark/ References

mu_H Maximum specific growth rate for

heterotrophic biomass [d-1]

8.000 3.000 13 a, b, e

K_S Half-saturation coefficient for heterotrophic

biomass

[g COD.m-3]

95.000 10.000 180 Ks (mg/L) value varies from 50 - 120 for domestic waste

(g, b, a)

K_OH Oxygen Half Saturation Coefficient for

heterotrophic biomass

[g (-COD).m-3]

1.150 0.300 2 K_OH value was reported

1mg/L (g, h, a)

K_NO Nitrate NO3-N half saturation coefficient for

heterotrophic biomass [g NO3-N.m-3]

0.150 0.100 0.2 a

b_H Decay coefficient for heterotrophic biomass

[d-1]

0.825 0.050 1.6 b_H without recycling was found to vary from 0.05 day-1 to 1.6 day-1 for some food-

processing wastes (a)

mu_A Maximum specific growth rate for autotrophic

biomass [d-1]

0.495 0.340 0.65 This parameter is strongly associated with the removal

of ammonia nitrogen (a)

K_NH Ammonia half-saturation coefficient for

autotrophic biomass

[g NH3-N.m-3]

1.000 0.500 1.5

c

K_OA Oxygen Half Saturation Coefficient for

autotrophic biomass

[g (-COD).m-3]

1.350 1.200 1.5

a, f

b_A Decay coefficient for autotrophic biomass [d-

1]

0.100 0.050 0.15 a

ny_g Correction factor for mu_H under anoxic

condition [dimensionless]

0.800 0.600 1 b, a

k_a* ammonification rate [m3.(gCOD.d)-1] 0.050 0.025 0.075 b, c

k_h* Maximum specific hydrolysis rate [g SBCOD.

(g cell COD. d)-1]

3.000 1.500 4.5 b, c, d

K_X* Half-saturation coefficient for hydrolysis of

slowly biodegradable substrate [g SBCOD.

(g cell COD)-1]

0.100 0.050 0.15

b, d

ny_h* Correction factor for hydrolysis under anoxic

condition [dimensionless]

0.800 0.400 1.2 b, d

Y_H Yield for heterotrophic biomass

[g cell COD formed. (g COD oxidized)-1]

0.575 0.460 0.69 a

Y_A Yield for autotrophic biomass

[g cell COD formed. (g COD oxidized)-1]

0.150 0.020 0.28 a

f_P* Fraction of biomass leading to particulate

products [dimensionless]

0.080 0.040 0.12 b, c

i_XB* Mass of nitrogen per mass of COD in biomass

[g N.(g COD)-1 in biomass]

0.080 0.040 0.12 b, c

i_XP* Mass of nitrogen per mass of COD in

products from biomass

[g N.(g COD)-1 in biomass]

0.06 0.030 0.09

b, c

v0* Secondary settlement maximum settling

velocity [m.d-1]

250 125 375 b, c

v0_max

*

Secondary clarifier maximum visilind velocity

[m.d-1]

474 237 711 b, c

r_h* Secondary settlement hindered zone settling

parameter [m3.(gSS)-1]

0.0006 0.0003 0.0009 b, c

r_p* secondary settlement flocculants zone settling

parameter [m3.(gSS)-1]

0.00286 0.00143 0.00429 b, c

f_ns* Secondary clarifier non-settleable fraction

[dimensionless]

0.00228 0.00114 0.00342 b, c

Henze et al (2002) argues that most kinetic parameters value varies widely and very dependent on the nature of the wastewater being treated.

Page 8: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

*Parameters whose upper bound (UB) and lower bound (LB) were taken ±50% of the average literature value.

a - Henze (2002), b - Benedetti et al. (2008), c - Jeppsson et al. (2007), d - Sweetapple et al. (2013), e - Dold (1986), f - Picioreanu et al. (1997)

g - Horan (1990), h - Henze (2008)

The model was run twice for each parameter with its lower and upper bounds values while keeping

the rest of the parameters’ value to the literature average. The six parameters to which the model was

sensitive, in terms of model performance measures (RMSE and R2), were selected for model

calibration purposes. In this study the term model performance is used to refer to how well the model

output is close to the observed dataset.

Four different values for the top six most sensitive parameters were used for model calibration; the

average value from literature, the lower/upper bound of the parameter and two more values in

between. The choice between lower and upper bounds was based on which limit (upper or lower) of

the parameter gave the highest percentage change in model calibration accuracy indicators, R2 and

RMSE. 4096 different combinations were created and the WwTW model ran 4096 times.

In order to assess the impact of having different levels of dataset on model performance, it is

necessary to follow the same calibration procedure in all the scenarios. This can avoid the

introduction of model performance increase/decrease due to inconsistent calibration levels. For this

reason, the same calibration procedure was carried out for each scenario.

Results and Discussion

The OAT sensitivity analysis results are presented in 4 and 5, showing the percentage change in

model performance indicator with respect to the base case (literature average value) when each

parameter was set to its respective upper and lower bounds. The variation of parameter values within

the feasible range can have a significant effect in model performance. The OAT sensitivity result

showed that the model performance indicators R2 and RMSE are highly sensitive to; b_H, r_h, Y_H,

Y_A, V0, f_ns, f_P, i_XB, K_S and K_OH which were also identified as sensitive parameters by

Benedetti et al. (2008) and Sweetapple et al. (2013). However, the model was not sensitive to

changes in b_A and to changes to V0_max at its upper bound.

The parameters to which the model is highly sensitive were selected in such a way that the first four

parameters that impact TSS R2 and RMSE were selected first. Second, the top four parameters that

highly affect NH3-N R2 and RMSE were selected and further assessed for calibration, see Error!

Reference source not found..

Unlike TSS, the NH3-N model performance indicators R2 and RMSE were always positively

correlated, i.e. parameter value change that results in increase of NH3-N R2 were also observed to

increase the of NH3-N RMSE. This is because parameter changes that result in NH3-N pattern similar

to the measured one but their values are far from the measure NH3-N. As a result, in the selection of

parameters for calibration more emphasis was given to NH3-N RMSE than R2.

Page 9: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

Figure 3: Percentage change in R2 and RMSE in simulating TSS

Figure 4: Percentage change in RMSE and R2 in simulating NH3-N

Since b_H was the most sensitive in all the measures at its lower bound, its lower bound range was

selected for calibration purposes. K_OH and K_S at their upper bound significantly increased model

performance indicators in simulating both TSS and NH3-N except they reduced R2 in simulating NH3-

N. Since the upper bound of these parameter improved model performance in most aspects, at a

higher degree, their upper bound range was selected for model calibration.

The maximum settling velocity in the secondary settlement tank (V0) at its upper bound and r_h at its

lower bound increases the model performance in simulating TSS but their impact on model

performance in simulating NH3-N is negative, especially the r_h. Due to their significant impact on

model performance in simulating TSS, these parameters were used for further calibration.

Page 10: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

Y_A at its lower bound reduced the RMSE in simulating NH3-N and TSS, and increased the R2 in

simulation of NH3_N and TSS. Y_H increased model performance in simulating TSS and NH3-N as

demonstrated by the R2 but it reduces the model performance due to a considerable increase in

RMSE at its lower bound. Since more emphasis was given to RMSE than R2 in analysing model

performance in simulating NH3-N, Y_A was selected for model calibration.

Selected parameters for model calibration and their respective four different values are listed in Table

3. The model has run for each unique combination of these parameters, which is discussed in the

next section.

Table 3: Most sensitive parameters selected for model calibration

Parameters used for calibration Values used for model calibration

b_H [d-1] 0.05 0.308 0.566 0.825

r_h [m3.(gSS)-1] 0.0003 0.0004 0.0005 0.0006

K_OH [g (-COD).m-3] 1.15 1.43 1.72 2.0

V0 [m.d-1] 474 553 632 711

K_S [g COD.m-3] 95 38.33 66.66 10

Y_A [g cell COD formed. (g COD oxidized)-1] 0.15 0.106 0.063 0.02

Calibration result for baseline model

The baseline model was calibrated using a semi-automated approach with the different combinations

of parameter values listed in Table 3. The model was run 4096 times to identify the best fit model

parameter combinations under calibration. Figure 5 shows that the model performance for the

baseline is not high with a maximum R2 value of 0.016 in simulating NH3-N and 0.015 in simulating

TSS. The parameter combinations which gave the highest R2 are not the ones which gave the lowest

RMSE, see Figure 6. The points with the same colour in figure 6 and 7 represents a simulation with

the same parameter values combinations. There were five combinations of parameters which gave R2

to be above 0.013 both for NH3-N and TSS. These combinations have RMSE between 3.6 – 3.8 for

NH3-N and 10 – 10.7 for TSS.

In absolute terms, the calibration accuracy indicators have low values. Since there is no consensus

on how to apply these performance measures nor has been their applicability over different value

ranges been established (Belia et al. 2009), it has not been possible to compare them with a

standard. Most often visual comparison is used as the only form of model performance assessment.

Ahnert et al. (2007) discussed the difficulties in finding a common basis for a standardised evaluation

of goodness-of-fit measures in wastewater treatment works modelling but unfortunately a benchmark

to compare model performance was not proposed.

Page 11: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

Figure 5: Calibration result baseline: model performance indicator R2 in simulating TSS

and NH3-N

Figure 6: Calibration result baseline: model performance indicator RMSE in simulating

TSS and NH3-N

Calibration result MLSS and DO scenario

Similarly this scenario was run 4096 times using the same parameter combinations. This scenario

resulted in more simulation runs having higher R2 values both in simulating NH3-N and TSS. There

were 55 simulations with R2 higher than 0.013 both for NH3-N and R2, see Figure 7. These

simulations also had the lowest RMSE ranging between 8.5 – 11mg/L for TSS and 2.6 – 2.85 for NH3-

N, see Figure 8. Higher numbers of accurate models were observed in this scenario than the baseline

scenario both in terms of R2 and RMSE.

Page 12: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

Figure 7: Calibration result MLSS and DO scenario: model performance indicator R2 in

simulating TSS and NH3-N

Figure 8: Calibration result MLSS and DO scenario: model performance indicator RMSE

in simulating TSS and NH3-N

Calibration results for the phenomenological influent generator scenario

In this scenario the model performance in simulating the NH3-N increased significantly but its

performance in simulating TSS did not change significantly. The R2 in simulating the NH3-N were

observed to be as high as 0.3 and 0.02 in simulating TSS, see Figure 9.

The parameter combinations which gave high R2 in simulated NH3-N were not the ones which

resulted in the highest R2 in simulating TSS, unlike the above two scenarios. One of the reasons for

this is the use of the same sensitivity analysis result that was used in the above two scenarios which

were built using different influent concentration inputs. Since a different input was used in this

scenario it would have been useful to run another sensitivity analysis and identify the most sensitive

parameter and define whether it is the upper bound or the lower bound which increases model

performance etc. However, model performance in simulating NH3-N increased much more than the

model performance in simulating TSS.

Page 13: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

Figure 9: Calibration result phenomenological influent generator scenario: model

performance indicator R2 in simulating TSS and NH3-N

Figure 10: Calibration result phenomenological influent generator scenario: model

performance indicator RMSE in simulating TSS and NH3-N

Conclusions

This paper investigated the performance of a WwTWs model that is potentially be used for active

control. The preliminary results obtained from the case study show the changes in model performance

under three scenarios of different levels of data availability in the model setup and calibration process.

The second scenario that uses measured MLSS and DO dataset in the calibration process shows a

slight model performance improvement.

The study showed that not all measured data increase model performance at equal level. The use of

a phenomenological influent generator achieved a higher improvement in model performance than the

use of MLSS/DO dataset. This is because a number of assumptions have to be made in the baseline

scenario in characterising the influent wastewater, due to lack of detailed influent quality data. The

low absolute model performance indicator values show a need for further model calibration and data

Page 14: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

monitoring, especially in characterising the influent wastewater. Thus, monitoring the influent quality at

a finer time scale and fractionating COD and total nitrogen can help modellers by avoiding

unnecessary data matching efforts in order to improve model performance.

Acknowledgments

This project is sponsored by EPSRC and Scottish Water as part of the STREAM Programme.

References

Abusam, A. and Keesman, K. (1999) Effect of number of CSTR's on the modelling of oxidation ditches: steady state and dynamic analysis. Mededelingen-Faculteit Landbouwkundige en Toegepaste Biologische Wetenschappen Universiteit Gent (Belgium).

Abusam, A., Keesman, K. J. and Van Straten, G. (2003) Forward and backward uncertainty propagation: An oxidation ditch modelling example. Water Research, 37(2), pp. 429-435.

Abusam, A., Keesman, K. J., Van Straten, G., Spanjers, H. and Meinema, K. (2001) Parameter estimation procedure for complex non-linear systems: Calibration of ASM No.1 for N-removal in a full-scale oxidation ditch. in. pp. 357-365.

Ahnert, M., Blumensaat, F., Langergraber, G., Alex, J., Woerner, D., Frehmann, T., Halft, N., Hobus, I., Plattes, M. and Spering, V. (2007) Goodness-of-fit measures for numerical modelling in urban water management–a summary to support practical applications. in Proceedings of 10th LWWTP Conference.

Belia, E., Amerlinck, Y., Benedetti, L., Johnson, B., Sin, G., Vanrolleghem, P. A., Gernaey, K., Gillot, S., Neumann, M. and Rieger, L. (2009) Wastewater treatment modelling: dealing with uncertainties. Water Science and Technology, 60(8), pp. 1929.

Benedetti, L., Batstone, D. J., De Baets, B., Nopens, I. and Vanrolleghem, P. A. (2008) Global sensitivity analysis of biochemical, design and operational parameters of the Benchmark Simulation Model no. 2. in Proceedings of iEMSs 2008: International Congress on Environmental Modelling and Software. pp. 7-10.

Dold, P. (1986) Evaluation of the general activated sludge model proposed by the IAWPRC task group. Water Science & Technology, 18(6), pp. 63-89.

Gernaey, K., Rosén, C. and Jeppsson, U. (2005) BSM2: A model for dynamic influent data generation: Technical report. Department of Industrial Electrical Engineering and Automation. Lund University, Lund, Sweden.

Gernaey, K. V., Flores-Alsina, X., Rosen, C., Benedetti, L. and Jeppsson, U. (2011) Dynamic influent pollutant disturbance scenario generation using a phenomenological modelling approach. Environmental Modelling & Software, 26(11), pp. 1255-1267.

Gernaey, K. V., Rosen, C. and Jeppsson, U. (2006) WWTP dynamic disturbance modelling - An essential module for long-term benchmarking development. in Water Science and Technology. pp. 225-234.

Hauduc, H., Gillot, S., Rieger, L., Ohtsuki, T., Shaw, A., Takács, I. and Winkler, S. (2009) Activated sludge modelling in practice: an international survey. Water Science and Technology, 60(8), pp. 1943.

Henze, M. (2002) Wastewater treatment: biological and chemical processes, Springer.

Henze, M. (2008) Biological wastewater treatment: principles, modelling and design, IWA publishing.

Page 15: THE IMPACT OF DATA AVAILABILITY ON PREDICTIVE … · settlement tanks. The final effluent from the settlement tank overflows over a weir into an outlet channel from where it gravitates

9th European Waste Water Management Conference

12-13 October 2015, Manchester, UK

www.ewwmconference.com

Organised by Aqua Enviro Limited

Henze, M., Gujer, W., Mino, T. and Van Loosdrecht, M. (2000) Activated Sludge Models ASM1, ASM2, ASM2d and ASM3, Scientific and Technical Report, Cornwall: IWA Publishing.

Horan, N. J. (1990) Biological wastewater treatment systems: theory and operation, John Wiley & Sons Ltd.

Jeppsson, U., Pons, M. N., Nopens, I., Alex, J., Copp, J. B., Gernaey, K. V., Rosen, C., Steyer, J. P. and Vanrolleghem, P. A. (2007) Benchmark simulation model no 2: General protocol and exploratory case studies. in. pp. 67-78.

Melcer, H. (2003) Methods for wastewater characterization in activated sludge modeling, IWA publishing.

Metcalf & Eddy, I. (2004) Wastewater engineering: treatment and reuse. Metcalf & Eddy, 4th ed., Inc., McGraw-Hill, New York.

Picioreanu, C., Van Loosdrecht, M. and Heijnen, J. (1997) Modelling the effect of oxygen concentration on nitrite accumulation in a biofilm airlift suspension reactor. Water Science and Technology, 36(1), pp. 147-156.

Rieger, L., Takács, I., Villez, K., Siegrist, H., Lessard, P., Vanrolleghem, P. A. and Comeau, Y. (2010) Data reconciliation for wastewater treatment plant simulation studies—planning for high-quality data and typical sources of errors. Water Environment Research, 82(5), pp. 426-433.

Rosén, C., Röttorp, J. and Jeppsson, U. (2003) Multivariate on-line monitoring: challenges and solutions for modern wastewater treatment operation. Water Science & Technology, 47(2), pp. 171-179.

Sweetapple, C., Fu, G. and Butler, D. (2013) Identifying key sources of uncertainty in the modelling of greenhouse gas emissions from wastewater treatment. Water Research, 47(13), pp. 4652-4665.

Takács, I., Patry, G. G. and Nolasco, D. (1991) A dynamic model of the clarification-thickening process. Water Research, 25(10), pp. 1263-1271.

UKWIR (2013) Role of Wastewater Process Control in Delivering Operating Efficiencies. London: UK Water Industry Research.