Model based river water quality assessment under current and future climate conditions

The second application of the WQ models is the study of the influence of model structure uncertainty on river water quality assessment.. The main contribution of this thesis is extending

Acknowledgements i

MODEL BASED RIVER WATER QUALITY

ASSESSMENT UNDER CURRENT AND

FUTURE CLIMATE CONDITIONS

Thanh Thuy Nguyen

Supervisor:

Prof dr ir Patrick Willems

Dissertation presented in partial fulfilment of the requirements for the degree of PhD in Engineering Science

August 2017

ARENBERG DOCTORAL SCHOOL FACULTY OF ENGINEERING SCIENCE

Cover image: photos/water-droplets-hd.html

http://all-free-download.com/free-MODEL BASED RIVER WATER QUALITY ASSESSMENT UNDER CURRENT AND

FUTURE CLIMATE CONDITIONS

Thanh Thuy NGUYEN

Prof dr ir P Willems partial fulfilment of the

requirements for the degree of Members of the Examination Committee: PhD in Engineering Science Prof dr ir H Heynen, chair

Prof dr ir.W Sansen, chair

Prof dr ir I Smets

Prof dr ir E Toorman

Prof dr ir A van Griensven

(Vrije Universiteit Brussel)

Assoc Prof dr ir T H L Pham

(Water Resources University, Hanoi, Vietnam)

August 2017

© 2017 KU Leuven, Science, Engineering & Technology

Uitgegeven in eigen beheer, Thanh Thuy Nguyen, Kasteelpark Arenberg 40 – bus 2448, B-3001 Heverlee (Belgium)

Alle rechten voorbehouden Niets uit deze uitgave mag worden vermenigvuldigd en/of openbaar gemaakt worden door middel van druk, fotokopie, microfilm, elektronisch of

op welke andere wijze ook zonder voorafgaandelijke schriftelijke toestemming van de uitgever

All rights reserved No part of the publication may be reproduced in any form by print, photoprint, microfilm, electronic or any other means without written permission from the publisher.

Water is key to life Language is key to society Knowledge is key to success

Acknowledgements

No one can live on other people but no one can succeed without support and cooperation from other people The day I finish my PhD research is not so far I would like to take a moment to thank all the people who, consciously or unconsciously, have contributed to this work

First of all, I would like to thank my promoter Prof Patrick Willems You brought me to the river water quality modelling since my Master thesis After my Master, you recommended me to keep working on the river water quality topic for my PhD At that moment, I thought there was not much work to do and everything was very clear During my PhD, I realized that nothing is perfect; the question was how we can mitigate the disadvantages and evaluate their effects

On the way to find the answers, I have received your guidance, comments and encouragements Especially, I highly appreciate the trust and freedom you gave

me to fill my research in the USA for the last 1.5 years I would like to thank Thuy Loi University for reserving my position at the university during my PhD research

I would like to thank Prof Ilse Smets and Prof Erik Toorman for the follow up

of my research and for providing comments and suggestions from the start of my PhD research to the final Your advice helped me to consider the research problems in a more comprehensive way In addition, I would also like to thank the other members of my jury for their feedback on my thesis text Prof Ann van Griensven, thank you for sharing your expertise to improve my future research Prof Thi Huong Lan Pham, thank you for making the long journey to attend my public defence Prof Willy Sansen and Prof Hilde Heynen, thank you for chairing my jury I also thank Prof Yoram Rubin and Dr Susan Hubbard for offering the opportunity to be a visiting student at UC of Berkeley I would like

Thank you Ingrid Keupers for your excellent work on river water quality modelling on which I could build further Thank you Els Van Uytven for your cooperation, and Vincent Wolfs for delivering the InfoWorks RS dongle to me during my maternity leave I also would like to thank all our group members who contributed to the development of WETSPRO and the climate change perturbation tool More thanks go to all of my friends who brought my home to Belgium and the USA in very colorful culture environments

A special word of thanks goes to my parents, siblings, niece and nephew You always encouraged and supported me to finish my PhD

I also would like to thank Phuong and Simba, who accompanied me since the very first day of my PhD We have shared the happy and sometimes though time together, even online or offline I am so proud of being with you, being your wife and your mama

Last but not least, I would like to say thanks to a very special person that I have admired since I was small Your love, your life and your work have been following me and lighting my life and my career up

Thuy Leuven, August 2017

Abstract

River water pollution is known as one of the major environmental issues in the world In order to achieve the sustainable development goals, it is crucial to provide quantitative information on river water quality Such information plays a key role in integrated river basin management in general and water quality management in particular Especially climate change with the increase in temperature, extreme flow conditions and changes of ecological systems is predicted to cause water quality degradation It requires policy makers to take right remediation actions Questions have to be answered such as how the water quality is, where surface waters are polluted, which sources cause stress to the rivers, how effective specific actions or measures would be Many existing methods have been proposed to answer these questions They make use of river water quality (WQ) observations in combination with models WQ measurements are limited to some discrete points along the river and at specific time moments They are also affected by measurement and sampling errors Therefore, it would

be incomplete, inaccurate and uneconomic if the WQ related decision-making would only rely on the available measurements By solving mathematical equations describing transport of pollutants and their biochemical processes in the aquatic environment, WQ models can provide a more complete picture of the

WQ state Therefore, the first main part of the thesis is applying models to simulate the transport of pollutants in the river system and analyse factors that influence this transport

In the first phase of the research, the three software packages MIKE 11, InfoWorks RS and InfoWorks ICM were used In these software packages, only MIKE 11 and InfoWorks RS allow to model the time variation in river bed roughness coefficient They, therefore, were selected for analysing the seasonal variation of river bed roughness impacts on water level The time series of roughness coefficient was calibrated to the observed water level in MIKE 11 The

iv Abstract

time series of roughness coefficient in InfoWorks RS was calibrated against the one calibrated in MIKE 11 due to the lack of measured data The results show that InfoWorks RS can simulate the influence better than MIKE 11, but the differences are relatively small The ignorance of vegetation growth on the roughness coefficient leads to underestimation in water levels in summer However, the influence is insignificant during periods of peak flows Because the highest determinant concentrations often occur during low water level periods, the influence is important for WQ modelling studies

The second application of the WQ models is the study of the influence of model structure uncertainty on river water quality assessment Firstly, the physico-biochemical processes were screened to obtain a preliminary assessment on the critical processes and to determine the processes that require a more detailed comparison Then, local sensitivity analysis was carried out to specify the sensitive parameters and processes Results show that the hydrodynamic results, heat transfer rate and reaeration simulations cause large differences in model simulation outputs for water temperature and dissolved oxygen (DO) concentrations The ignorance of processes related to sediment transport, phytoplankton and bacteria has a significant influence on the higher concentrations of organic matter and lower values of dissolved oxygen concentrations The three models show consensus on the main pollutant sources explaining organic matter and nitrate concentrations, but disagree on the main factors explaining the DO concentrations

The use of software packages as MIKE 11 and InfoWorks, which are based on full hydrodynamic models, show difficulties in quantifying model structure uncertainty due to the rigid structures Additionally, to get convergence in results,

a small time step is required The simulation for long periods and big systems requires very long simulation time Therefore, these models are inapplicable for studies in which many simulations or long simulation periods are needed One solution is the use of fast conceptual models The main contribution of this thesis

is extending the COnceptual RIver WAter Quality model based on the InfoWorks RS (CORIWAQ-RS) processes and results The research started from the code that Keupers (2016) developed for MIKE 11 The model conceptualizes rivers using cascades of reservoirs and lumps the advection-diffusion-physico-biochemical processes The hydrodynamic inputs are derived from the outputs of the full hydrodynamic models We performed comparative analysis on the CORIWAQ-RS and CORIWAQ-MIKE11 models Results indicate that the

Abstract v conceptual models perform equally well as the original MIKE 11 and InfoWorks

RS models, but with much shorter simulation time (104 times) The successful testing of the conceptual models opens a development avenue towards problem solving in the context of WQ control and management

Next, CORIWAQ-RS was considered on the basis of two types of applications: uncertainty analysis of river WQ input and assessment of climate change impacts

on river water quality In addition to the incomplete knowledge on the WQ processes, lack of pollution input data is known as one of the major uncertainty sources in WQ modelling Model improvement, hence, should focus on obtaining more information on the input/boundary conditions Given that there are so many pollutant sources along a river, model improvement actions should focus

on the most important sources to optimize the cost-benefit ratio of the actions Input uncertainties were modeled by the stochastic regression approach, i.e adding randomly time series of errors to estimated time series by regression equations Based on sensitivity indices related to the model output error variance, important pollution sources were identified The contribution of input uncertainties to total model output residuals was quantified by the total variance

of model output errors due to all inputs

The application of CORIWAQ-RS for climate change impact analysis involved: (i) quantifying climate change impacts on river flows and concentrations; (ii) determining and evaluating the input factors that primarily control the changes and (iii) evaluating the uncertainty of climate change scenarios on the water quality assessment Precipitation, potential evapotranspiration and air temperature

in 30 climate change scenarios were statistically downscaled from 30 Generalized Climate Model runs The influences are evaluated for the observed period 2000-

2010 and the target period 2050-2060 The hydrological model, regression and stochastic approaches were applied to transform the climate change signals for meteorological variables to changes in runoff, nitrogen loads from catchments,

DO concentrations and physico-biochemical rates It is shown that climate change may lead to highly negative impacts for DO concentrations The climate change impacts are, however, highly uncertain, especially for the high return periods

The case study selected to implement all objectives of the doctoral research is the Molse Neet river catchment in Belgium

Beknopte samenvatting

Pollutie van rivierwater staat bekend als één van de voornaamste milieuproblemen wereldwijd Om de duurzame ontwikkelingsdoelen te kunnen bereiken is het cruciaal om kwantitatieve informatie te analyseren m.b.t de rivierwaterkwaliteit Ze speelt een cruciale rol in het integrale stroomgebiedsbeheer

in het algemeen en in het waterkwaliteitsbeheer in het bijzonder Op vele locaties wereldwijd worden de waterkwaliteitsnormen nog niet gehaald Bovendien kan de klimaatverandering de waterkwaliteit van waterlopen negatief beïnvloeden Klimaatverandering zorgt immers voor een toename in de lucht- en watertemperatuur, voor meer extreme waterloopdebieten, zowel toenemende piekafvoeren als dalende laagwaterafvoeren, en gerelateerde wijzigingen aan het ecologisch systeem Er is daardoor nood aan sanerings- en klimaatadaptatiestrategieën Om de ontwikkeling of het ontwerp van zulke strategieën te ondersteunen wordt in modern waterbeheer gebruik gemaakt van simulatiemodellen Deze moeten toelaten de huidige waterkwaliteitstoestand te kwantificeren en te evalueren, de oorzaken na te gaan en meest efficiënte oplossingen voor te stellen De modellen worden complementair aan de beschikbare waterkwaliteitsmetingen en andere empirische informatie gebruikt Waterkwaliteitsmetingen zijn immers typisch slechts beschikbaar op een beperkt aantal locaties langs de rivier of op slechts een beperkt aantal tijdsmomenten Verder zijn deze waarnemingen sterk onderhevig aan meet- of analysefouten De beschikbare waarnemingen zijn dus onvolledig en onnauwkeurig Het zou inefficiënt zijn om waterkwaliteitsbeheer uitsluitend te baseren op zulke informatie Via wiskundige simulatiemodellen kunnen de waarnemingen fysisch onderbouwd geïnterpoleerd worden Ook kunnen ze via scenario-analyses geëxtrapoleerd worden, vb in de tijd – o.a via scenario’s inzake klimaatopwarming – of door simulatie van bepaalde wijzigingen aan de waterloop of aanpassingen aan het waterbeheer Deze inter- en extrapolaties gebeuren door gebruik te maken van kennis over de fysische processen en vervuilingsbronnen die aan de basis liggen van de waterkwaliteitstoestand van een waterloop De modellen

viii Beknopte samenvatting

lossen hiertoe de transportprocessen op, zoals de advectie-dispersievergelijking en

de voornaamste biochemische omzettingsprocessen Er zit evenwel nog veel onzekerheid vervat in zulke modellering Zeker in vergelijking met waterkwantiteitsmodellering zoals hydrologische en hydraulische modellering is er nog veel onderzoek nodig naar de nauwkeurigheid van waterkwaliteitsmodellering, zowel wat de modelinvoer (vervuilingsbronnen) betreft als de proceskennis en -schematisering Er is nood aan verder onderzoek inzake de invloed van de modelstructuur (de set van geïmplementeerde procesvergelijkingen) en het bepalen van de meest optimale modelstructuur voor een specifieke toepassing Dit vraagt een verbeterd inzicht in de onzekerheden betrokken bij de modellering zodat een goede afweging kan gemaakt worden tussen de modelgedetailleerdheid,

de bijhorende rekentijd en de nauwkeurigheid Vele scenario-analyses, zoals het simuleren van de impact van de klimaatverandering, andere externe invloedsfactoren en bepaalde waterbeheerstrategieën vragen lange-termijn simulaties (simulatie van tijdreeksen van meerdere tientallen jaren) of een groot aantal modelsimulaties Hetzelfde geldt ook voor onderzoek naar de modelonzekerheden Verder is de kennis over de impact van de hiervoor opgesomde typen scenario’s op de waterkwaliteitstoestand langs waterlopen nog ondermaats Dit doctoraatsonderzoek had als doel om voor een deel aan deze noden tegemoet te komen De modellering van de waterkwaliteit langs de Molse Neet in het Netebekken fungeerde daarbij als gevalstudie De waterkwaliteitsmodellering beperkte zich tot de opgeloste zuurstofcyclus inclusief organische verontreiniging, stikstofnutriënten en de watertemperatuur

Meer specifiek werden voor de gevalstudie drie bestaande en frequent gebruikte softwarepakketten bestudeerd: MIKE 11, InfoWorks RS en InfoWorks ICM De modelstructuur van elk van deze modellen werd geanalyseerd en geïmplementeerd

in het COnceptual RIver WAter Quality (CORIWAQ) conceptueel modelleringsplatform dat eerder aan de Afdeling Hydraulica van de KU Leuven werd ontwikkeld, maar toen beperkt tot de MIKE 11-procesvergelijkingen CORIWAQ conceptualiseert waterlopen via een cascade van reservoirmodellen

en voert een ruimtelijke aggregatie door van de advectie-dispersie en biochemische waterkwaliteitsprocessen per riviertak (per reservoir) Als modelinvoer wordt de uitvoer van hydrodynamische modellen en gegevens over

de verschillende vervuilingsbronnen langs de waterloop beschouwd Dit is dezelfde invoer als gebruikt in de MIKE 11, InfoWorks RS en ICM gedetailleerdere waterkwaliteitsmodellen

Beknopte samenvatting ix

Na vergelijking met de oorspronkelijke, gedetailleerdere modelresultaten en waterkwaliteitsmetingen bleken de drie conceptuele modellen een vergelijkbare nauwkeurigheid op te leveren De rekentijd van de conceptuele modellen was – in vergelijking met de gedetailleerdere modellen – een factor 104 kleiner, wat perspectieven opent voor heel wat toepassingen

Voor elk van de drie sets aan waterkwaliteitsmodelvergelijkingen (MIKE 11, InfoWorks RS en ICM) werd via een gevoeligheidsanalyse de kritische processen

en parameters geïdentificeerd Voor de modelresultaten van watertemperatuur en opgeloste zuurstof bleken dit de hydrodynamische modelresultaten (invoer voor

de waterkwaliteitsmodellen) te zijn, alsook de warmteoverdracht- en natuurlijke herbeluchtingsprocessen De verwaarlozing van sedimenttransport, phytoplankton

en bacteriën bleek een belangrijke impact te hebben op de hoge concentraties aan organische vervuiling en lage concentraties aan opgeloste zuurstof De drie modelstructuren bleken tot dezelfde conclusies te komen m.b.t welke vervuilingsbronnen vooral de concentraties aan organisch materiaal en nutriënten bepalen Ze bleken echter verschillende conclusies te geven over welke factoren

de opgeloste zuurstofconcentraties verklaren

Een andere belangrijke factor bleek de tijdsvariatie van de ruwheid van de rivierbedding, door de seizoensvariatie in begroeiing, op de waterpeilen langs de waterloop Van de drie bestudeerde modelleringspakketten laten enkel MIKE 11

en InfoWorks RS toe om deze invloed rechtstreeks te analyseren Er werd op basis van beide modellen geconcludeerd dat het niet inrekenen van de seizoensvariatie in de rivierbedruwheid een belangrijke invloed kan hebben op de waterpeilen in de zomer Tijdens periodes met rivierwassen is de invloed dan weer klein

De lage rekentijd van de CORIWAQ-modellen liet toe om een gedetailleerde onzekerheidsanalyse uit te voeren, wat vooral belangrijk bleek voor de onzekerheden op de modelinvoer Voor de verschillende inschattingen aan vervuilingsbronnen en andere modelinvoer werden de verschillende aannames geïnventariseerd en de mogelijke impact ervan op de invoer gekwantificeerd Deze modelinvoeronzekerheden werden via een stochastische modelleringsmethode beschreven en gepropageerd doorheen het conceptueel model Via gevoeligheidsindices werd het relatieve belang van elk van de bestudeerde onzekerheidsbronnen ingeschat Dit liet daarna toe om de totale onzekerheid via een variantie-decompositiemethode op te delen in de relatieve bijdragen van elk

de periode 2000-2010 en de klimaatveranderingsimpact werd bestudeerd voor het toekomsthorizon 2050-2060 De resultaten tonen voor bepaalde klimaatscenario’s een erg negatieve impact, vooral voor de concentraties aan opgeloste zuurstof en hoge terugkeerperioden

xi

Acronyms

APHA American Public Health Association

BELGAQUA Belgian Federation for the Water Sector

BOD Biological Oxygen Demand

CODA Centrum voor Onderzoek in Diergeneeskunde en

Agrochemie’

CORIWAQ COceptual RIver WAter Quality model

CORIWAQ-MIKE11 CORIWAQ developed based on MIKE 11

CORIWAQ-RS CORIWAQ developed based on InfoWorks RS DEFRA Department for Environment Food & Rural Affairs DHI Danish Hydraulic Institute

EPA Environment Protection Agency of United State

ETo potential EvapoTranspiration

Eurostat European Statistics

GWP Global Water Partnership

IE Inhabitant Equivalent

xii Acronyms

IQR InterQuartile Range

IRO Interfaculty Council for Development Co-operation

of KU Leuven

KjN Kjeldahl Nitrogen

KMI/IRM Royal Meteorological Institute of Belgium

LGP Length of Growing Period

LHS Latin Hypercube Sample

MATLAB MATrix LABoratory

NIST SEMITECH National Institute Standards & Technology

NN-N Sum of nitrate and nitrite

NSE Nash-Sutcliffe Efficiency

Nt-N Total nitrogen

PDF Probability Density Distribution

PDM Probability Distribution Model

PEGASE Planification Et Gestion de l’ASsainissement des

RWQM River Water Quality Model

SENTWA System for the Evaluation of Nutrient Transport to

Water

TAW Tweede Algemene Waterpassing

Tempbase Base Temperature

Acronyms xiii

USDA U.S Department of Agriculture

VHA Vlaamse Hydrografische Atlas

VHM Veralgemeend conceptueel Hydrologisch Model VLIR Flemish Interuniversity Council

VMM Vlaamse MilieuMaatschappij

WETSPRO Water Engineering Time Series PROcessing tool WHO/UNICEF World Health Organization/United Nation

Children’s Fund

xvi Table of contents

3.1.1 Probability Distribution Model 24 3.1.2 Generalized Hydrological Model 26

3.3.2 Physico-biochemical transformation processes 32

4 Development of a COnceptual RIver WAter Quality model

Table of contents xvii

5 Seasonal variation of river bed roughness impacts on water level 83

5.2.1 MIKE 11 calibration roughness coefficient 88 5.2.2 InfoWorks RS time-varying roughness coefficient 93

6 Influence of model structure uncertainty on river water quality

6.2.1 Graphical sensitivity analysis 102

7.2.1 Gap filling and stochastic modelling of input uncertainties 120

xviii Table of contents

8.1.3 Evaluation of the input factors controlling the climatic changes

8.1.4 Sensitivity of the hydrological model calibration to the climate

8.2.1 Propagation of climate change to river water quality input 143 8.2.2 Impacts of climate change on extreme flows and determinant

8.2.3 Evaluation of the input factors controlling the climatic changes

8.2.4 Sensitivity of the hydrological model calibration to the climate

9.1 Contributions of the doctoral research 158 9.1.1 Quantitatively evaluate and compare the reference river WQ models and increase the insights in the river WQ drivers 158 9.1.2 Develop a river water quality conceptual model based on

9.1.3 Uncertainty analysis and assessment of climate change impacts

List of Figures

Fig 1.1 Impact mechanism of CC on river WQ 11 Fig 1.2 Structure of the thesis 15 Fig 2.1 Molse Neet catchment 18 Fig 2.2 Livestock in Molse Neet catchment 18 Fig 2.3 Aquatic plants in Molse Neet river in August 2015 19 Fig 3.1 Model structure of The PDM rainfall runoff model (Moore et al 2007).24 Fig 3.2 Model structure of the VHM rainfall runoff model (Willems 2014) 26 Fig 3.3 Simulated and observed water temperature at station 333000, Molse Neet river 45 Fig 3.4 Parameters used to select nearly independent extreme flows (Willems 2009) 51 Fig 4.1 Flow chart of the CORIWAQ operation for MIKE 11 and RS 60 Fig 4.2 Reservoir determination in CORIWAQ for MIKE 11 and RS 60 Fig 4.3 Rainfall runoff from upstream sub-catchment to Molse Neet river 63 Fig 4.4 Calibration results at the location of station 333000 for the year 2008 for (a) water temperature, (b) DO, (c) NH4-N, (d) NO3-N, (e) BOD5, (f) PO4-P and (g) PP-P concentrations in MIKE 11 66 Fig 4.5 Calibration results at the location of station 333000 for the year 2008 for (a) water temperature, (b) DO, (c) NH4-N, (d) NO3-N, (e) BOD5, (f) ON-N and (g) NO2-N concentrations in RS 67

Fig 4.6 Box Whisker plots of calibrated f Tr (a) and f depth (b) at all reservoirs along the Molse Neet river for wet, mean and dry years corresponding to MIKE 11 (left) and RS (right) models 70 Fig 4.7 Scatter plots of water temperature and DO concentrations in MIKE 11 and CORIWAQ-MIKE11 or observed data at observation station 333000 along the Molse Neet river after applications of different calibrated parameter sets 72

xx List of Figures

Fig 4.8 Scatter plots of water temperature and DO concentrations in RS and CORIWAQ-RS or observed data at observation station 333000 along the Molse Neet river after application of different calibrated parameter sets 73 Fig 4.9 Scatter plots of NH4-N and NO3-N concentrations in MIKE 11 and CORIWAQ-MIKE11 or observed data at observation station 333000 along the Molse Neet river after application of different calibrated parameter sets 74 Fig 4.10 Scatter plots of NH4-N and NO3-N concentrations in RS and CORIWAQ-RS or observed data at observation station 333000 along the Molse Neet river after application of different calibrated parameter sets 75 Fig 5.1 Observed versus MIKE 11 simulated water depths at Meerhout, after use

of a constant Manning coefficient of 0.035 s/m1/3 88

Fig 5.2 Calibrated Manning’s n values per five-day period with RMSE for

calibrated value (a) and piecewise linear relation of the Manning’s n versus day of the year (b) 89 Fig 5.3 Observed versus MIKE 11 simulated water depths at Meerhout, afters use of a time-varying Manning coefficient 90 Fig 5.4 Cumulative probability plot for observed and simulated water depths for summer 2005, calibration (a) and summer 2007, validation (b) 90 Fig 5.5 Observed versus MIKE-11 simulated water depths, after use of a time-varying and flow-dependent Manning coefficient for summer 2007 91 Fig 5.6 Rating curves for Manning coefficients ranging from 0.03 to 0.08 s/m1/3

with an 0.005 interval (a) and relation of the rating curve parameters on the seasonal Manning coefficient (b) 92 Fig 5.7 Comparison of the estimated river discharges, after use of a fixed and time-varying Q – h relationship with the MIKE-11 simulated discharges and the precipitation time series for summer 2005 93 Fig 5.8 Correlation between independent water depths in MIKE-11 and InfoWorks RS; for low (a) and peak (b) water depths in calibration period (2004-2005) and for low (c) and peak (d) water depths in validation period (2006-2007);

at 8.034 km along the Molse Neet River 94 Fig 5.9 Local roughness, overall cross-section Manning’s n in InfoWorks RS and MIKE 11 at 8.034 km along the Molse Neet river on three specific days: 1 February, 15 June and 31 July 95 Fig 6.1 Procedure to implement model structure uncertainty analysis 101 Fig 6.2 Profiles of the maximum and minimum water temperature (a), BOD5 (b),

NH4-N (c), NO3-N(d) and DO (e) concentrations along Molse Neet river in the initial simulations 103

List of Figures xxi Fig 6.3 (left) 10-percentile DO and 90-percentile BOD5 concentration profiles based on RS, ICM and MIKE 11 results in comparison with the Flemish standard A; (right) concentrations versus return period at the observation station 333000 111 Fig 6.4 Comparison of CDF-relationships for simulated DO concentrations at station 333000 in MIKE 11, RS, ICM and fundamental intermittent standards with a return period of 1 month, 3 month and 1 year 112 Fig 7.1 Procedure for the input uncertainty analysis 119 Fig 7.2 Trend in discharge from factory Philips 124 Fig 7.3 Empirical and calibrated distributions of (a) KjN concentration from Ajinomoto, (b) KjN concentration and (c) BOD5 concentrations from Philips 126 Fig 7.4 Empirical and calibrated distributions for DO concentration from Ajinomoto 128

Fig 7.5 Sensitivity indices S ij of pollutant sources for (a) DO, (b) NH4-N, (c)

NO3-N, (d) BOD5 and (e) ON-N concentrations, for the river concentrations at the observed station 333000 at the Molse Neet river in the whole year, winter, spring, summer and autumn 131

Fig 7.6 S ij for different sample sizes for ON-N concentrations from Ajinomoto factory to river ON-N concentrations 132 Fig 7.7 Contribution of input uncertainty to total model residuals 133 Fig 8.1 Procedure to evaluate impact of uncertainty in CC scenarios on WQ and contribution of each input factors 142

Fig 8.2 Box Whisker plots of (a) changes in LGP and (b) changes in ip

coefficient values for specific crops 144 Fig 8.3 Box Whisker plots of the impact factors calculated for (a) monthly precipitation and (b) monthly total nitrogen losses 145 Fig 8.4 Box Whisker plots of the changes in (a) water temperature and (b) DOS concentration at the boundaries of sub-catchments in CC scenarios 147

Fig 8.5 Box-Whisker plots of the IP calculated for temperature coefficient (a) α=2.4, (b) α= 4.7 148

Fig 8.6 Impact factors for extreme (a) high flows and (b) low flows, versus return period 149 Fig 8.7 Impact factors for extreme (a) high NH4-N, (b) NO3-N, (c) BOD5, (d) ON-N and (e) low DO concentrations, versus return period 150 Fig 8.8 Impact factors calculated for extremely (a) high flows and (b) low flows with PDM model parameters calibrated to Grote Nete Varendonk station, versus return period 153

xxii List of Figures

Fig 8.9 Impact factors for extreme high (a) NH4-N, (b) NO3-N, (c) BOD5, (d) ON-N and (e) low DO concentrations at measurement station 333000 with PDM model parameters calibrated to Grote Nete Varendonk station, versus return period 154

List of Tables

Table 2.1 Manufactories discharging to the Molse Neet river 20 Table 2.2 Number of inhabitants discharging their wastewater to the Molse Neet without treatment and as point sources (UE-114363, UE-114372, UE-114373) The other inhabitants discharging their wastewater untreated (point sources with lower IE) are distributed over the hydrological zone VHA501 (distributed source UD-501) 21 Table 3.1 Physico-biochemical processes and equations in InfoWorks RS/ICM and MIKE 11 33 Table 3.2 WQ parameter values in InfoWorks RS, ICM and MIKE 11 41 Table 3.3 Summary of available data for different pollutant sources 42 Table 3.4 Initial (Init) and fine-tuned values (Cali) for the nitrogen fractions of different nitrogen components from agriculture 47 Table 3.5 Overview of 30 GCM runs considered 50

Table 4.1 Minimum values of NSE of CORIWAQ-RS vs RS calculated with

different values of 12 WQ parameters for the 7 WQ variables 78 Table 5.1 Transition depth in function of the total unit roughness 96 Table 6.1 Simulations conducted for local sensitivity analysis 104 Table 6.2 Sensitivity indices for the maximum and minimum temperatures and

DO concentrations (the white cells refer to RMSE and the grey cells to S) 108 Table 6.3 RMSE between the BOD5, NH4-N and NO3-N concentrations of the different MIKE 11 simulations and those of the RS_1 simulation 109 Table 7.1 Coefficients of regression equations for the model input variables of factory Ajinomoto (independent variables in the rows, dependent variables in the columns) The last row shows the error distributions (normal distributions indicating the mean and standard deviation as follows: N(mean, standard deviation) 122

xxiv List of Tables

Table 7.2 Coefficients of regression equations for the model input variables of factory Philips (independent variables in the rows, dependent variables in the columns) The last row shows the error distributions (normal distributions indicating the mean and standard deviation as follows: N(mean, standard deviation)) 123 Table 7.3 Estimated number of inhabitants that discharge to the Molse Neet river 125 Table 7.4 Estimated wastewater discharge by domestic households 125

Table 7.5 DW values for input variables of industrial pollution loads 126

Table 8.1 Standard deviation of the impact factors calculated for extreme values

of WQ variables due to changes in 4 input factors 152 Table 8.2 Standard deviation of the impact factors calculated for extreme values

of WQ variables due to changes in 4 input factors with PDM model parameters calibrated to the Grote Nete Varendonk station 153

1 Introduction

1.1 Problem statement

Clean water is an increasing concern to our society due to its strong direct and indirect influence on human’s health The United Nations (2006) reported that people with water-related diseases contribute up to 50 % of hospitalized patients

in the world Water-related diseases cause nearly one of every five deaths under years old (WHO/UNICEF 2009) Under impacts of climate change (CC), this effect might be more severe due to depletion of water quantity and degradation

5-of water quality Observations and CC scenarios derived for different economic and societal conditions indicate significant changes in air temperature and extreme precipitation (Ntegeka et al 2014) These two variables are driving forces

of the physico-biochemical processes in rivers As a result, in order to support river management plans that cope with CC, it is essential to quantitatively evaluate the changes of river water quality (WQ) under different CC conditions Such studies are, however, still very limited (Michalak 2016)

Research on CC impacts on river WQ is implemented by field experiments and/or numerical modelling (Moss et al 2003, 2004) They both indicate negative impacts of CC on river WQ Water temperature, biological oxygen demand and nitrogen loads to rivers increase while dissolved oxygen in water reduces (Wilby 1993; Whitehead et al 1997; Bouraoui et al 2002 and ; Monteith et al 2007) During the summer, the temperature increases, the flow decreases and the residence time increases, which are good conditions for algae blooms, especially

in low dissolved oxygen concentration and higher nutrient concentrations (i.e nitrogen and phosphorus) In the winter, the higher temperature, organic matter

1

2 Chapter 1: Introduction

decay and soil mineralization cause more organic matter and nitrogen to be transported to river during the storms However, the direction and magnitude of changes in determinant concentrations for specific case studies can be very different For instance, the simulation results for the Kennet river in the UK showed increasing trends in both ammonium and nitrate concentrations (Wilby et

al 2006) However, van Vliet and Zwolsman (2008) reported that in drought events, decreases in nitrate concentration and increases in ammonium concentration were observed at different locations in the Meuse river The DO concentrations at one station increase while they decrease at another station

To keep water clean, it is crucial to effectively control and manage the WQ, which requires quantitative knowledge on the water quantity and quality status both in time and space The WQ status can be obtained from monitoring systems (Shrestha and Kazama 2007; Gholizadeh et al 2016) However, such systems are expensive and usually insufficient to cover the high spatio-temporal variability of

WQ variables (Bartley et al 2012; Lessels and Bishop 2015) Therefore, it would

be incomplete, inaccurate and uneconomic if the WQ related decision-making would only rely on the available measurements In that context, mathematical

WQ modelling is a useful tool to provide such information WQ modelling also gives decision makers insights into the cause-effect relationships and predicts the future WQ states, and therefore, support the design of water resources management strategies (GWP 2013) By comparing simulation results for different scenarios, decision makers can easily derive evidence-based decisions

However, the main problem in using mathematical models is their accuracy The use of simple models exposes to many assumptions and simplification, which can reduce the accuracy of simulation results On other hand, complex models that integrate a full HydroDynamic (HD) model with a detailed description of the physico-biochemical processes allow to derive results with detailed temporal and spatial variations The WQ modelling software packages are different in given assumptions, number and equations of physico-biochemical processes that are taken into account These dissimilarities can lead to discrepancies in their simulation results Therefore, it is crucial to quantitatively compare WQ modelling packages to provide the users with insights of each package and support them in selecting a suitable model for their study area

The typical error in HD river studies, which reduces the modelling accuracy, is that the roughness coefficient is usually assumed to be constant, hence disregarding the seasonal effect of vegetation growth This assumption may lead to a bias in

Chapter 1: Introduction 3

the estimation of river flow velocities and water depth Where high nutrient

loadings are present in rivers with low velocities, aquatic plants can grow

abundantly and possibly mitigate the detrimental effects of this pollution stress

Indeed, a lower velocity means longer residence time and higher sedimentation

rates and thus a higher self-purifying capacity (Schulz et al 2003) Therefore, it is

necessary to consider the sensitivity of the flow regime to seasonal variation of

river bed roughness due to vegetation growth

Another challenge in WQ modelling and management is the scarcity and

heterogeneity of the WQ data available for model calibration and validation

Various types of point and diffuse pollutant sources are spatially spread over the

whole catchment and their concentrations and loads are highly variable in both

time and space The information about this variability is, however, often very

scarce Some pollutant sources even do not have any measurement of their

discharges and concentrations Several authors reported that the pollutant input

uncertainties are one of the major sources of uncertainty in river WQ modelling,

if not the most important one (Radwan et al 2004; Freni and Mannina 2010;

Willems 2012) For these reasons, it is crucial to carefully handle the missing data

and to assess the influence the lack of input data has on the simulation results

Finally, among many WQ models, one may pose the question which model is

suitable to conduct CC impact assessment with relatively high accuracy As

mentioned above, the detailed WQ models are able to simulate the temporal and

spatial variations of river WQ variables However, they have long calculation

times, which makes the use of these models impractical for many types of

applications The long simulation time poses difficulties on applications that

involve huge number of model runs, iterations and/or long-term simulations,

such as model uncertainty analysis, auto-calibration, real-time control,

optimization, etc Particularly, the assessment of CC on WQ, which involves

30-year or more long-term simulations for several scenarios, is impractical with

detailed WQ models Therefore, it is necessary to develop the conceptual model

that can derive simulation results similar to the WQ models but in shorter time

1.2 State of the art

Dym (1980) defined a mathematical model to be “a triplet (S, Q, M) where S is a

system, Q is a question relating to S, and M is a set of mathematical statements M

= {1, 2, , n} which can be used to answer Q” A river WQ model is the

4 Chapter 1: Introduction

mathematical model which depicts the river (S) to answer questions reg the space evolution of water quality variables along the river (Q) By solving mathematical equations (M) describing transport of pollutants and their biochemical processes

time-in the aquatic environment, WQ models can provide a more complete picture of the WQ state They allow one to better identify possible stress positions and important pollutant sources as well as to evaluate the effectiveness of remediation actions The slope of WQ profiles also helps the modelers to recognize the sensitive processes and to assess the reliability of their parameter values

Model types

The Streeter and Phelps equation (Streeter and Phelps 1925) for simulating dissolved oxygen (DO) and biochemical oxygen demand (BOD5) have formed the basis of many WQ models In terms of simulation complexity, WQ models can be classified into simple (e.g TOMCAT), intermediate (e.g QUAL2E) and complex (e.g Delft 3D) models (Cox 2003) Although the simple model requires less input, it considers only steady flow state and limited physico-biochemical processes The intermediate models are more complicated Yet, some important processes, e.g back flows, loops in river systems and lateral rainfall-runoff, are not accounted for The complex models do take these processes into account but are computationally expensive and may encounter the problem of parameter identifiability in calibration The choice of the most appropriate model depends

on the study area However, WQ modelling software packages are not always clear about their assumptions and application conditions As a result, it is difficult for users to judge whether a given model is suitable for their specific applications For example, differences in empirical roughness equations for main channels and hydraulic structures representation may cause important differences in the HD simulations and, as a consequence, also in the WQ results (Warmink et al 2011) Emulation modelling is known as a low-order approximation of the detailed physically-based models to reduce their computational complexity (Castelletti et

al 2012) In this manner, the most relevant variables are taken into account in the emulator The variables are identified by data-based or structure-based approaches In the data-based approach, the variables can be selected by a statistical measure of input-output relationship, e.g partial mutual information (Bowden et al 2005) and minimum redundancy maximum relevance (Hejazi and Cai 2009) In the structure-based approach, a model formulation is derived for each possible combination of the variable replacement by constants The model

Chapter 1: Introduction 5 performances are evaluated by criteria such as residual sum of squares, Akaike’s information criterion and Bayesian information criterion (e.g Cox et al 2006 and Crout et al 2009)

This research meets the above-mentioned needs and builds further on the recent advances by the development of a flexible river WQ model CORIWAQ (COnceptual RIver WAter Quality) CORIWAQ is a hybrid of conceptual and physically-based models to obtain more accurate simulations than the traditional lumped conceptual models and shorter computational time than detailed physically-based reference models Accordingly, HD information for CORIWAQ

is obtained from detailed physically-based models The biochemical transformation processes are explicitly simulated in the CORIWAQ model, similar to the physically-based models after applying correction factors With the data from the detailed models, the lumped model is implemented for the motions of determinant concentrations The advection and diffusion processes along river segments are conceptualized using the reservoir-type approach The physico-biochemical processes along the river segments are presented by a set of equations with the incoming concentrations at the first cross-sections and outcoming concentrations at the last cross-sections of the corresponding river segments With this approach, each river segment is characterized by one time series for each hydraulic characteristic and WQ variable This approach has been widely used to transform precipitation to runoff (Pedersen et al 1980; Te Chow

et al 1988; Weiler 2003; Chetan and Sudheer 2006 and Nourani et al 2009), groundwater recharge to discharge (Peters et al 2003) and for river or tidal river hydraulics (Wolfs et al 2015, Meert et al 2016) However, there are only few studies on the application of lumped conceptual models for river WQ modelling (e.g Whitehead et al 1997; Willems and Berlamont 2002; Radwan et al 2003, 2004; Willems 2008) Two detailed physically-based models, implemented in the software packages, MIKE 11 and InfoWorks RS (hereafter denoted shortly as

“RS”), with different numerical schemes and different equations to simulate biochemical transformations, are selected as reference models This research is a follow-up of the initial CORIWAQ developments based on MIKE 11 (CORIWAQ-MIKE11) by Keupers and Willems (2017) CORIWAQ-MIKE11 was applied to simulate the influence of combined sewer system overflows on river WQ (Keupers et al 2015) and to conduct global sensitivity analysis of WQ parameters (Keupers and Willems 2015)

6 Chapter 1: Introduction

Uncertainties in river WQ modelling

Both detailed and simplified conceptual WQ modelling involves several types of uncertainties These uncertainties are typically classified into model structure uncertainty, parameter uncertainties, input uncertainties and measurement errors These types of uncertainty collectively determine the total model output uncertainty Uncertainties related to WQ modelling have long been recognized but only few studies quantified these uncertainties For instance, Beck (1978) and Tung and Yen (2006) named the types of uncertainty as well as the means to quantify and apply uncertainties in general, but they did not discuss specific models The measurement uncertainty was considered by Harmel and Smith (2007) and parameter uncertainty was addressed by Mannina and Viviani (2009) There have been not many studies on the model structure and input uncertainties while they are known as two dominated uncertainty sources in river WQ modelling (Van der Perk 1997; Håkanson 2000; Van Griensven and Meixner 2006)

Total model output uncertainty

The total model output uncertainty can be determined by different approaches corresponding to available data for model validation tests (Refsgaard et al 2006; Van Griensven and Meixner 2006) When observation data are available, the model parameters can be achieved through calibration After split of the data into

2 periods, one for calibration and one for evaluation, the differences between the model simulation results and observed data can be analysed for the evaluation period, e.g by computing the model residual variance This total uncertainty may

be decomposed in its main contributing uncertainty sources by variance decomposition (Radwan and Willems 2008; Freni and Mannina 2010b; Willems 2012) This is done after quantifying and propagating the input and parameter uncertainties in the model, computing their contributions to the total model output variance, and considering the rest variance as the result of model structure uncertainty (apart from observation errors)

Model structure uncertainty

As explained in the previous section, the model structure uncertainty can be computed as the rest uncertainty after subtracting the contributions of model input and parameter uncertainties from the total model output uncertainty When

no observation data are available, an alternative approach is to consider several different plausible model structures and analyse the differences in results This

Chapter 1: Introduction 7 was done by Van der Perk (1997) considering 8 different model structures to assess the model structure uncertainty in simulating phosphorus concentrations Instead of using multiple models, Lindenschmidt et al (2007) considered the exchange data of sub-models The data in the sub-models were linked by linear regression equations and stochastic error terms were added to these equations In case the direct uncertainty quantification is inapplicable to insufficient empirical material, the expert judgement is used (Hora 1993)

The model structure uncertainty analysis has been intensively studied in hydrology, but there has been little effort on applying this analysis to river WQ modelling Willems (2008) and (Freni and Mannina 2010b) quantified the total uncertainties in urban WQ models and attempted to decompose these in its major contributing uncertainty sources The variance decomposition approach was also implemented by Radwan et al (2004) to quantify the structure uncertainty of a river WQ model

Input uncertainties

The input uncertainty is primarily caused by a lack of data Existing approaches to handle the missing data can be divided into two groups depending on whether the study area is measured/gauged or unmeasured/ungauged For the measured

WQ variables, they are classified into traditional methods and “modern” methods (i.e applying maximum likelihood estimation, Bayesian estimation and multiple imputation) (Enders 2010) The traditional methods deal with missing data by list-wise deletion or by filling (e.g using mean imputation and regression imputation (Shrive et al 2006; Kalteh and Hjorth 2009)) These methods are beneficial in computational cost but may cause bias in the filling data estimation when the missing values are not completely random For instance, the missing observed data for river flow during storm events can lead to underestimation when mean imputation is used As for the “modern” methods, the maximum likelihood estimation determines the filling data with a given probability distribution, but this distribution may be subject to a high secondary uncertainty when the available sample size is small (NIST SEMATECH 2012) The Bayesian method estimates the posterior distributions for inputs from a given prior distribution Multiple imputations involve combining stochastic regression and Bayesian estimation Such “modern” methods aim to obtain unbiased estimates for the filling input data distributions Stochastic regression is a traditional method but it can derive unbiased estimates with much less computational cost than the “modern” methods

8 Chapter 1: Introduction

For ungauged basins, regionalization approaches are often applied in hydrological studies In such studies, hydrological model parameters for ungauged basins are obtained from parameters of gauged basins typically using regression, spatial proximity or physical similarity In regression approach, the relationships between catchment characteristics and model parameters are obtained from large data sets (e.g Young 2006) The spatial proximity approach consists of transferring parameters of neighboring catchments which are similar to the ungauged catchment in term of climate and catchment conditions (e.g Parajka et al 2007) When the model parameters are undetermined for neighboring catchments, the parameters can be obtained from the donor catchments which have similar catchment descriptors to the ungauged catchment (e.g McIntyre et al 2005)

To assess the model uncertainty related to the lack of model input data, the worldwide used approach is sensitivity analysis (SA) (Saltelli et al 2008) SA is recommended in the guidelines for extended impact assessment by the European Commission (EC 2002) Based on the testing area, SA is classified into local and global methods In the local methods, the output variability is achieved by changing the input factors around reference values (Pianosi et al 2016) The sensitivity is then quantified as the partial derivatives (Hill and Tiedeman 2007) or explored by algebraic “no box” SA (Norton 2015) The local approaches involve

a very low computational cost However, they are only applicable for linear or additive models (Saltelli and Annoni 2010) and it is impossible to compare the sensitivity of the different inputs Additionally, these methods are not straightforward when the inputs are variable in time Meanwhile, global SA can consider the output variation across the entire space of input factors for nonlinear models Commonly applied global SA methods are the Elementary Effect Test (EET), or the variance-based methods (FAST and Sobol) The inputs considered in such analysis are not only the model input but also the model data (Hamm et al 2006) and their resolutions (Baroni and Tarantola 2014) The uncertainties considered in these researches are mostly for the model parameters (e.g Nossent et al 2011; Vanuytrecht et al 2014; Peeters et al 2014), steady input variables (Hamm et al 2006) and their resolutions (Baroni and Tarantola 2014) A major drawback is that the global methods are computationally expensive with huge number of model runs

Seasonal variation of river bed roughness impacts on water level

One of the assumptions in state-of-the-art WQ modelling and one of the contributors to the model uncertainty is the assumption that the river bed