An Integrated Modeling and Decision Framework to Evaluate ... · An Integrated Modeling and...

An Integrated Modeling and Decision Framework to Evaluate Adaptation Strategies

for Sustainable Drinking Water Utility Management Under Drought and Climate Change

PROJECT NO.4636

An Integrated Modeling and Decision Framework to Evaluate Adaptation Strategies for Sustainable Drinking Water Utility Management Under

Drought and Climate Change

Prepared by:

Joseph Kasprzyk, Jenna Stewart, Aaron Heldmeyer, Kelsey Reeves, William Raseman, Balaji Rajagopalan, R. Scott Summers, Fernando Rosario-Ortiz, and Ben Livneh

University of Colorado Boulder

Co-sponsored by:

U.S. Environmental Protection Agency

2019

ii The Water Research Foundation

The Water Research Foundation (WRF) is a nonprofit (501c3) organization that provides a unified source for One Water research and a strong presence in relationships with partner organizations, government and regulatory agencies, and Congress. WRF conducts research in all areas of drinking water, wastewater, stormwater, and water reuse. The Water Research Foundation’s research portfolio is valued at over $700 million.

WRF plays an important role in the translation and dissemination of applied research, technology demonstration, and education, through creation of research-based educational tools and technology exchange opportunities. WRF serves as a leader and model for collaboration across the water industry and its materials are used to inform policymakers and the public on the science, economic value, and environmental benefits of using and recovering resources found in water, as well as the feasibility of implementing new technologies.

For more information, contact: The Water Research Foundation

1199 North Fairfax Street, Suite 900 Alexandria, VA 22314-1445 P 571.384.2100

6666 West Quincy Avenue Denver, Colorado 80235-3098 P 303.347.6100

www.waterrf.org [email protected]

©Copyright 2019 by The Water Research Foundation. All rights reserved. Permission to copy must be obtained from The Water Research Foundation. WRF ISBN: 978-1-60573-452-1 WRF Project Number: 4636

This report was prepared by the organization(s) named below as an account of work sponsored by The Water Research Foundation. Neither The Water Research Foundation, members of The Water Research Foundation, the organization(s) named below, nor any person acting on their behalf: (a) makes any warranty, express or implied, with respect to the use of any information, apparatus, method, or process disclosed in this report or that such use may not infringe on privately owned rights; or (b) assumes any liabilities with respect to the use of, or for damages resulting from the use of, any information, apparatus, method, or process disclosed in this report.

Prepared by University of Colorado Boulder

The research on which this report is based was developed, in part, by the United States Environmental Protection Agency (EPA) through Cooperative Agreement No. R835865 with The Water Research Foundation. However, the views expressed in this document are not necessarily those of the EPA and EPA does not endorse any products or commercial services mentioned in this publication. This report is a publication of WRF, not EPA. Funds awarded under the Cooperative Agreement cited above were not used for editorial services, reproduction, printing, or distribution.

This document was reviewed by a panel of independent experts selected by The Water Research Foundation. Mention of trade names or commercial products or services does not constitute endorsement or recommendations for use. Similarly, omission of products or trade names indicates nothing concerning The Water Research Foundation's or EPA's positions regarding product effectiveness or applicability.

An Integrated Modeling and Decision Framework to Evaluate Adaptation Strategies for Sustainable Drinking Water Utility Management Under Drought and Climate Change iii

Acknowledgments Thank you to all the utilities, graduate students, research personnel, and other collaborators who participated in this study.

Research Team Principal Investigators: Rajagopalan Balaji, PhD Ben Livneh, PhD Fernando L. Rosario-Ortiz, PhD R. Scott Summers, PhD Joseph R. Kasprzyk, PhD University of Colorado Boulder Project Team: Ariel Retuta, MS Yun Yu, PhD Paul Wilkerson, BS Jenna Stewart, MS Leah Bensching, MS Aaron Heldmyer, BS Carli Brucker, BS Kelsey Reeves, MS William Raseman, PhD University of Colorado Boulder Project Advisory Committee William Becker, PhD Hazen and Sawyer Tirusew Asefa, PhD Tampa Bay Water Paul Conrads, MS USGS

WRF Staff John Albert, MPA Chief Research Officer

Kenan Ozekin, PhD Unit Leader – Research Services

iv The Water Research Foundation

Abstract and Benefits Abstract:

The project presents an integrated framework that enhances understanding of the processes that impact water quality under a suite of climate and natural hazards, quantifies the attendant uncertainty, and informs a decision support tool that enables evaluation and selection of mitigation strategies considering socio-economic impacts and tradeoffs. Four objectives were considered. The first was to understand flow and sediment generation from drinking water supply watersheds, especially given natural hazards such as wildfires. The second objective was to understand the mobilization and transport of dissolved organic matter (DOM) and sediments (i.e., turbidity), and in some cases nutrients (associated with algal growth) through the watershed and eventually to the water treatment plant. The third objective was to develop source water thresholds for turbidity and disinfection byproduct (DBP) precursors based on regulatory constraints in finished water, and to predict water quality threshold exceedances using stream water quality data with extreme value theory. The fourth objective was to develop a simulation-optimization decision support system for water treatment plants using data and methods generated in the other activities.

Benefits:

• Improved understanding of post-fire water quality through a new, physically based estimate of suspended sediment response to wildfire disturbance

• Guidance for utilities on expected water quality effects from a wildfire, including mobilization of TOC as a function of peak temperature

• A framework for assessing the exceedance of filtered water turbidity standards based on source water and settled water turbidity, allowing utilities to control turbidity events and their propagation through a treatment plant

• A decision support tool that addresses utilities’ need to adapt to changing source water quality and reduce costs of water treatment

Keywords: water quality, hydrology, sediment transport, water treatment, statistical modeling, decision support

An Integrated Modeling and Decision Framework to Evaluate Adaptation Strategies for Sustainable Drinking Water Utility Management Under Drought and Climate Change v

Contents Acknowledgments ........................................................................................................................................ iii Abstract and Benefits ................................................................................................................................... iv Tables ........................................................................................................................................................... vii Figures ......................................................................................................................................................... viii Acronyms and Abbreviations ........................................................................................................................ x Executive Summary ..................................................................................................................................... xiii

Chapter 1: Introduction ................................................................................................................................ 1

Chapter 2: Activity 1: Estimating Watershed-Scale Response of Post-Fire Sediment ............................... 3 2.1 Overview ............................................................................................................................. 3

2.2 Developing a Generalized Model of Watershed-Scale Streamflow and Suspended Sediment ............................................................................................................................. 3

2.3 Evaluation of the Sensitivity of Sediment Response to Wildfire in a Mountainous Watershed, and the Detectability of Post-Fire Sediment Signals ....................................... 9 2.3.1 Sensitivity of Sediment Response to Wildfire in a Mountainous Watershed ........ 9 2.3.2 Wildfire-Based Sediment Signal Detection in the West ...................................... 13

Chapter 3: Activity 2: Impact of Heating Temperature on the Character of Water-Soluble Constituents from Organic and Mineral Soils ............................................................................. 23

3.1 Introduction ...................................................................................................................... 23 3.2 Materials and Methods ..................................................................................................... 24

3.2.1 Soil and Litter Sampling ....................................................................................... 24 3.2.2 Sample Processing ............................................................................................... 26 3.2.3 Heating Simulation............................................................................................... 26 3.2.4 Soil and Litter Leaching ........................................................................................ 27 3.2.5 Analyses of Material and Leachates .................................................................... 33 3.2.6 Soluble Elements ................................................................................................. 34 3.2.7 Statistical Analysis ................................................................................................ 34

3.3 Results and Discussion ...................................................................................................... 34 3.3.1 Carbon and Nitrogen Mineralization ................................................................... 34 3.3.2 Solubility of Organic Carbon and Nitrogen .......................................................... 36 3.3.3 Soluble Elements ................................................................................................. 38

3.4 Conclusions ....................................................................................................................... 41

Chapter 4: Activity 3: Source Water Quality Thresholds and Exceedance Evaluation ............................ 43 4.1 Introduction ...................................................................................................................... 43

4.1.1 Activity 3 Motivation ........................................................................................... 43 4.1.2 Importance of Characterizing Thresholds for Source Water Quality .................. 43 4.1.3 Motivation for Modelling Surface Water Quality Using Climate and Land

Cover Predictors .................................................................................................. 43 4.1.4 Objectives of Activity 3: Source Water Quality Thresholds and Exceedance

Evaluation ............................................................................................................ 43 4.2 Source Water TOC and Bromide Thresholds .................................................................... 44

4.2.1 Overview .............................................................................................................. 44

vi The Water Research Foundation

4.2.2 Methods ............................................................................................................... 45 4.2.3 Results .................................................................................................................. 50 4.2.4 Modeling Using Extreme Value Theory ............................................................... 54 4.2.5 Conclusions .......................................................................................................... 56

4.3 Source Water Turbidity Thresholds .................................................................................. 56 4.3.1 Overview .............................................................................................................. 56 4.3.2 Data Collection ..................................................................................................... 57 4.3.3 Probability of Target Exceedances ....................................................................... 58 4.3.4 Turbidity Spikes throughout the Treatment Train ............................................... 62

4.4 Modelling Water Quality Using Climate and Land Cover Predictors ................................ 68 4.4.1 Overview .............................................................................................................. 68 4.4.2 Methods ............................................................................................................... 70 4.4.3 Results .................................................................................................................. 73 4.4.4 Conclusions .......................................................................................................... 76

Chapter 5: Activity 4: Decision Support Tool for Adapting to Variable Water Quality and Competing Objectives ..................................................................................................................................... 79 5.1 Stakeholder Input ............................................................................................................. 80 5.1.1 XLRM Framework ................................................................................................. 81 5.2 Iterative Problem Formulation ......................................................................................... 82 5.2.1 Relationships: Select an Appropriate Treatment Model ..................................... 82

5.2.2 Uncertainties: Quantify Uncertainty in Model Inputs ......................................... 83 5.2.3 Decision Levers: Identify Relevant Decisions Based on Model Inputs and Data

Availability ............................................................................................................ 84 5.2.4 Performance Measures: Identify Relevant Metrics Based on Model Outputs .... 84

5.3 Decision Support Tool Case Study .................................................................................... 85 5.4 Water Quality Scenarios ................................................................................................... 85 5.5 Simulation-Optimization ................................................................................................... 87

5.5.1 Decision Levers .................................................................................................... 89 5.5.2 Performance Measures: Objectives ..................................................................... 89 5.5.3 Performance Measures: Constraints ................................................................... 89 5.5.4 Pareto Optimal Solutions ..................................................................................... 90

5.6 Interactive Visualization.................................................................................................... 91 5.7 Summary and Conclusions ................................................................................................ 93 Chapter 6: Conclusion ................................................................................................................................ 95 References .................................................................................................................................................. 97

An Integrated Modeling and Decision Framework to Evaluate Adaptation Strategies for Sustainable Drinking Water Utility Management Under Drought and Climate Change vii

Tables 2-1 Parameters Used in the Multi-Objective Optimization Routine for Each Calibration ..................... 7 2-2 Summary of Mean and Maximum Differences between Forecast Mean and Observed SSL (in

kg/s) during Post-Fire Periods ........................................................................................................ 20 3-1 Summary of Soluble Element Trends and Peak Concentrations from Soil and Litter ................... 40 4-1 Data Summary for the Miller WTP, Harwood’s Mill WTP and the Betasso WTP ........................... 46 4-2 Monthly Source Water TOC Thresholds for the Betasso WTP ...................................................... 52 4-3 Data Collected from Partner Drinking Water Utilities ................................................................... 58 4-4 Probability of Target Turbidities Given the Turbidity at a Prior Process Point Less than a

Specified Value ............................................................................................................................... 61 4-5 Meaning of Zones Shown in Figure 4-9 in Terms of the Normalized Data in Each Zone

Compared to the Normalized Average Value of the Data ............................................................. 64 4-6 Summary Statistics, Data, and Normalized Data for an Example Illustrating Zones for

Individual Data Points .................................................................................................................... 65 4-7 Percentage of Data Points per Each of the 6 Zones for Each Plot of Normalized Data ................. 66 4-8 Description of Data Sets Used to Model Water Quality in the CLP at Fort Collins, CO ................. 72 4-9 Description of Water Quality Data Set for the Trinity River at Houston, TX ................................. 72 4-10 Models Built for the CLP, Fort Collins Case Study .......................................................................... 73 5-1 XLRM Framework Ideas Generated at the December 2016 Workshop ........................................ 82 5-2 A Revised XLRM Framework Based on Availability of Models and Data ....................................... 85 5-3 Summary of Decision Levers and Performance Measures ............................................................ 90

viii The Water Research Foundation

Figures 2-1 Portrayal of the Four Basic Factors that Control Soil Erosion, Which Motivate the Model

Development for This Study ............................................................................................................ 4 2-2 illustration of the Five-Algorithms Considered in This Study within the VIC Hydrologic Model,

as a Consistent Framework .............................................................................................................. 4 2-3 Study Domain ................................................................................................................................... 5 2-4 Critical Areas Computed for Using Slope Threshold of 15˚ and Above, Vegetation Types

Including Shrub, Grassland, Bare Ground (SGB), and Forest, and Stream Proximities of 100 m and 500 m. Critical area Included Four Estimated Regions ............................................................. 6

2-5 (a) Hydrograph for CC-Late with VIC Streamflow Computed Over the Calibration Period 1992 – 1997 for the Joint Algorithm Calibration with a Spin-Up Year in 1991. (b) Absolute Value of Bias Performance for Calibration, Validation and Transfer Periods for All Algorithms in CLP-L, CC-Early and CC-Late ......................................................................................................... 8

2-6 Application of the Model Ensemble for the Clear Creek Catchment, from 1950 to 2013 Using a Top Performing Parameter Set from CC-Late ............................................................................... 8

2-7 Depiction of the 2012 High Park Fire Scar (Red) atop the CLP Basin ............................................... 9 2-8 Sobol Sensitivity Analysis Results for Sediment Models DHSVM (SHESED), HSPF, and MUSLE .... 11 2-9 Hill Gulch (Top) and Skin Gulch (Bottom) Estimated Total Sediment Yields ................................. 13 2-10 Western U.S. Study Domain ........................................................................................................... 14 2-11 USGS Gage 08353000 (Rio Puerco near Bernardo, NM)................................................................ 15 2-12 Spatial Summary of Coefficient ‘a’ from the MRC Curve Fitting Process ...................................... 17 2-13 Spatial Summary of Exponent ‘b’ from the MRC Curve Fitting Process ........................................ 17 2-14 Spatial Summary of R2 from the MRC Curve Fitting Process ........................................................ 18 2-15 Autocorrelation Function (Left) and Partial Autocorrelation Function (Right) for SSL .................. 19 2-16 Model-Predicted SSL versus Observations Following Five Observed Wildfires at the Study Site

Near Bernardo, NM ........................................................................................................................ 20 2-17 Cross-Correlation between Streamflow and Precipitation ............................................................ 21 3-1 Sampling Locations (Indicated by Stars in the Boulder Creek Watershed) ................................... 26 3-2 DOC Kinetics Plots for Soils, Subsamples A, B, D, and E ................................................................ 28 3-3 TDN Kinetics Plots for Soils, Subsamples A, B, D, and E ................................................................. 29 3-4 DOC Kinetics Plots for Litter, Sites NED (Left) and GROSS (Right) ................................................. 30 3-5 DOC Kinetics Plots for Litter, Sites NED (Left) and GROSS (Right) ................................................. 30 3-6 DOC Linearity Plots for Soils, Subsamples A, B, D, and E ............................................................... 31 3-7 TDN Linearity Plots for Soils, Subsamples A, B, D, and E ............................................................... 32 3-8 DOC (Left) and TDN (Right) Linearity for Litter from Sampling Site, NED ...................................... 32 3-9 Carbon (Top Panel) and Nitrogen (Bottom Panel) Remaining in the Mineral and Organic

Layers after Heating at Various Temperatures Organized in Ascending Order ............................. 35 3-10 Fraction of Water Extractable Organic Carbon (WEOC) for Organic and Mineral Layer Samples

(Left), and Fraction of Water Extractable Organic Nitrogen (WEON) for Organic and Mineral Layer Samples (Right) ..................................................................................................................... 36

3-11 Water Soluble Elements for Mineral Layer (Soil, Left Panel) and Organic Layer (Litter, Right Panel) Material for NED ................................................................................................................. 39

4-1 Monthly Source Water Temperature for the Miller and Harwood’s Mill WTPs ............................ 47 4-2 Monthly Source Water TOC Concentrations for the Miller and Harwood’s Mill WTPs ................. 48 4-3 Conventional Treatment Modeling Validation .............................................................................. 50 4-4 Comparison of Monthly Source Water TOC Concentrations and TOC Thresholds ........................ 52

An Integrated Modeling and Decision Framework to Evaluate Adaptation Strategies for Sustainable Drinking Water Utility Management Under Drought and Climate Change ix

4-5 Comparison of Monthly Source Water Concentrations and Thresholds for Bromide and TOC Based on the Miller WTP Data ....................................................................................................... 54

4-6 Observed TOC Concentrations with 2-year, 20-year, and 100-year Return Periods for the Stationary and Non-Stationary GPD Models with a TOC Threshold of 2.2 mg/L .......................... 55

4-7 Conventional Surface Water Treatment Process with Turbidity Checkpoints .............................. 57 4-8 Plots of Turbidity Data at the Houston, TX Water Facility ............................................................. 60 4-9 Example Zoning of a Normalized Settled Turbidity versus Normalized Raw Turbidity Plot .......... 64 4-10 Example Data Points from Table 4-6 Plotted into Zones ............................................................... 65 4-11 Figures Showing Zoning of the SRWTP, Sacramento, CA Turbidity Data ....................................... 66 4-12 Time Series of Raw and Settled Turbidity for the EAFWTP, Sacramento, CA Utility ..................... 68 4-13 HUC 8 Watersheds for Fort Collins (left) and Houston (Right) ...................................................... 72 4-14 Scatterplots of Observed versus Modelled Turbidity for Models 1 and 2 ..................................... 73 4-15 Box Plots of R Squared from Performing the Drop 10% Cross Validation Method 500 Times ...... 74 4-16 Model 3 on Turbidity in the CLP .................................................................................................... 74 4-17 Model 4 on TOC in the CLP River ................................................................................................... 75 4-18 Scatterplots of Observed versus Modelled Turbidity for Models 5 and 6 ..................................... 76 4-19 Box Plots of R Squared from Performing the Drop 10% Cross Validation Method 500 Times ...... 76 5-1 Overview of Water Treatment Decision Support Tool for Choosing Treatment Alternatives

That Are Resilient to Water Quality Variability: 1) Gather Stakeholder Input to Inform Problem Formulation, 2) Refine Problem Based on Additional Information, 3) Generate Influent Water Quality Scenarios, 4) Use Simulation-Optimization Techniques to Generate Alternatives, and 5) Select among Alternatives Using Interactive Visualization Techniques ........ 80

5-2 Map of Utilities Participating in December 2016 Workshop and the Influent Water Quality Data Collection Effort ..................................................................................................................... 81

5-3 A Representative Example of Source Water Temperature Based on Data from Denver Water in Colorado ..................................................................................................................................... 83

5-4 A Representative Example of Total Organic Carbon Based on Data from Houston, TX ................ 84 5-5 Quarterly Boxplots of the Observed Record (nyears = 11) and Simulated Data (nsimulations = 500

Each of nyears = 11) Simulations for the Fort Collins Cache la Poudre Dataset (2007-2017) .......... 87 5-6 A Conceptual Diagram of the Optimization Using a Multi-Objective Evolutionary Algorithm

(MOEA) and the WTP Model to Evaluate Operational Decisions for Treating Cache la Poudre River Water .................................................................................................................................... 88

5-7 Conventional Treatment Train Simulated Using the WTP Model ................................................. 88 5-8 Parallel Coordinates Plot of Pareto Optimal Operating Policies Produced by the Simulation-

Optimization Outlined in Figure 5-6 and Table 5-3 ....................................................................... 91 5-9 A Parasol-Based Web Application for Interactive Visualization of Water Treatment Operating

Policies Generated by the Borg MOEA .......................................................................................... 92

x The Water Research Foundation

Acronyms and Abbreviations BIC Bayesian Information Criterion CART Classification and Regression Trees CC Clear Creak at Golden, CO CFE Combined filter effluent CLP Cache La Poudre CLP-F Cache La Poudre River at Mouth of Canyon near Fort Collins, CO CLP-L North Fork Cache La Poudre River at Livermore, CO CSWTP Conventional surface water treatment plant CT Contact time DBP Disinfection byproduct D/DBPRs Disinfectants and Disinfection Byproducts Rules DHSVM Distributed Hydrology Soil Vegetation Model DIN Dissolved inorganic nitrogen DOC Dissolved organic carbon DOM Dissolved organic matter DON Dissolved organic nitrogen DS Distribution system EVT Extreme value theory FLG Flagstaff Mountain sample GAC Granular activated carbon GCV Generalized Cross Validation criterion GEV Generalized Extreme Value GLM Generalized Linear Regression GPD Generalized Pareto distribution GROSS Gross Reservoir sample HAA5 Five haloacetic acids HSPF Hydrologic Simulation Program - Fortran k-NN K-Nearest Neighbors LOADEST Load Estimator MCL Maximum contaminant level MOEA Multi-Objective Evolutionary Algorithm MRC Monovariate Rating Curve MUSLE Modified Universal Soil Loss Equation NDVI Normalized Difference Vegetation Index NED Nederland, CO sample OC Organic carbon OM Organic matter ON Organic nitrogen PDSI Palmer Drought Severity Index POT Peaks Over Threshold SOM Soil organic matter SSL Suspended sediment load TDN Total dissolved nitrogen TOC Total organic carbon TTHM Total trihalomethanes

An Integrated Modeling and Decision Framework to Evaluate Adaptation Strategies for Sustainable Drinking Water Utility Management Under Drought and Climate Change xi

WEOC Water extractable organic carbon WEON Water extractable organic nitrogen WTP Water treatment plant VIC Variable Infiltration Capacity model XLRM Uncertainties, Decision Levers, Relationships, Measures

An Integrated Modeling and Decision Framework to Evaluate Adaptation Strategies for Sustainable Drinking Water Utility Management Under Drought and Climate Change xiii

Executive Summary Climate change and its manifestation in the hydrologic cycle, along with extreme events, can have profound impacts on water quality and water quantity. These impacts will be felt by drinking water systems - especially by utilities charged with the task of providing safe and reliable supplies of water to the public. The impact is felt through variability in a suite of water quality variables that are relevant to stream ecosystems and drinking water treatment. This is further exacerbated by natural hazards such as fire in the water supply watershed combining with climate extremes. For example, droughts produce lower water supply, but if punctuated by extreme wet events and fires, they mobilize a lot of sediment (turbidity), nutrients, and dissolved organic content – all of which have significant impact on water utilities as they seek to deliver safe and reliable drinking water. There are several studies investigating the impact of climate change on water quantity; however, there is a paucity of tools relating them to water quality. Such an understanding and set of tools would enable efficient planning and management strategies and also aid in setting more-sensible regulatory policies. With this main motivation, the research team proposed an interdisciplinary framework with four broad activities: (i) understanding watershed response to post-fire sedimentation; (ii) understanding the heating due to fires on mobilization of organic carbon; (iii) understanding water quality thresholds, exceedances and modeling using climate and land surface variables; and (iv) developing a decision support tool to adapt to varying water quality. The outcomes from these activities are summarized below.

ES.1 Activity 1: Estimating Watershed-Scale Response of Post-Fire Sediment A generalized model of watershed-scale streamflow and suspended sediment was developed to simulate the impacts of climate and land cover disturbances on flow volumes and sediment transport. Five erosion and suspended sediment load algorithms were applied, showing larger year-to-year variation in suspended sediment than previous estimates. The simulated sediment is sensitive to climate conditions; therefore, the model results provide new insights into how suspended sediment is likely to respond to changing climate conditions and extremes. The sensitivity of sediment response to wildfire was also evaluated within the model, showing a large and distinct increase in sediment loading after wildfire. An exclusively data-driven analysis of post-fire streamflow conducted over the western U.S. suggests that post-fire signals are often masked by data limitations, underscoring the necessity of a well-validated model to predict post-fire sediment responses.

Towards the end of the project, a laboratory-scale rainfall simulator and burning apparatus was constructed to measure sediment responses across a range of terrain slopes, wildfire intensities, and rainfall intensities. This apparatus allows for detailed measurements of slope, fire, and storm impacts on sediment transport. Preliminary analysis suggests that wildfire severity will be the largest driver of suspended sediment, with slope and rainfall intensity playing secondary roles.

ES.2 Activity 2: Impact of Heating Temperature on the Character of Water-Soluble Constituents from Organic and Mineral Soils A wildfire is a natural and ubiquitous phenomenon that often leaves behind a perturbed environment. Post-fire landscapes become susceptible to enhanced erosion, decreased infiltration capacity, and soil hydrophobicity, which facilitate the transport of post-fire residue into surface waters that often serve as potable water sources. The risks posed by post-fire residue, or ash, deposition to raw water sources and subsequent impairment of water quality is difficult to understand due to the complex nature of wildfire

xiv The Water Research Foundation

effects. To address the impacts of ash on water quality, surface soils and litter were progressively heated to a range of temperatures (150 to 550˚C) and subsequently leached in water to evaluate changes in the release of dissolvable constituents as a function of burn temperature. Water quality parameters assessed include dissolved organic carbon (DOC), dissolved organic nitrogen (DON), dissolved inorganic nitrogen (DIN), fractions of water extractable organic carbon and nitrogen (WEOC and WEON), and soluble elements. Following heating, both the quantity and solubility of DOC and DON from soil were enhanced at 250˚C to 350˚C while that from litter decreased drastically after material was heated above 150˚C. DOC contributions from litter across all temperatures exceeded that from soil while DON contributions for both materials were comparable. Soluble elements ranged in trends; however, litter concentrations were magnitudes higher than soil in general, except for Al and Mn, which were comparable between materials. This work contributes to the growing understanding of the impacts of wildfires on water quality in general and helps identify major contributors (soil or litter) to water quality risk based on fire severity.

ES.3 Activity 3: Source Water Quality Thresholds and Exceedance Evaluation Spatio-temporal variability of surface water quality impacts the regulatory compliance of the finished water from drinking water utilities, consequently impacting public health. Of particular interest are total organic carbon (TOC), turbidity, and bromide. TOC reacts with common disinfectants, such as chlorine, to form regulated and unregulated disinfection byproducts (DBPs). Bromide also increases the formation of DBPs and their health effects. Turbidity is an indicator of the potential presence of pathogenic microorganisms. In many cases, turbidity and TOC and bromide concentrations vary significantly seasonally and regionally. Thus, understanding and modeling the thresholds and the variability of these constituents in the influent waters of the treatment plant is crucial in developing mitigation strategies.

To this end, this activity of the project had two broad objectives: (i) determine source water thresholds for TOC, bromide, and turbidity based on regulatory constraints in the finished water, and use source water quality data with extreme value theory to model water quality threshold exceedances; and (ii) develop models that relate surface water quality concentrations to historic climate (precipitation, temperature, drought index) and land surface (vegetation index) variables. The researchers developed thresholds for TOC and bromide at representative sample locations with diverse climates. Statistical learning methods based on regression trees and nonlinear regression were used in the development of skillful models for TOC and turbidity using climate and land variables, bypassing streamflow. This is a significant contribution in that streamflow data is often hard to obtain, while climate and land surface variables are readily available. Extreme Value Theory was applied to model threshold exceedances of TOC, which will be of immense help in developing a meaningful regulatory regime to ensure safe drinking water. A framework for demonstrating the probability of target level turbidities being exceeded given increases in prior process turbidity levels was developed and applied to utility data. These probabilities of target exceedances provide utilities with opportunities to perform risk-based assessments of their utility operations. Relationships between source water turbidity levels to settled and filter effluent turbidities, and settled water turbidity levels to combined filter effluent (CFE), are uniquely described via the novel zoning methodology presented in this activity. Each zone provides insight into how signals of turbidity propagate throughout the treatment train.

An Integrated Modeling and Decision Framework to Evaluate Adaptation Strategies for Sustainable Drinking Water Utility Management Under Drought and Climate Change xv

ES.4 Activity 4: Decision Support Tool for Adapting to Variable Water Quality and Competing Objectives The fourth objective was to develop a decision support tool to help utilities adapt to changing source water conditions and improve the reliability and efficiency of treatment. To integrate the preferences of water managers, the researchers developed this tool based on input from water utilities across the United States. Based on this feedback, the researchers designed a simulation-optimization tool capable of generating innovative operating policies for existing treatment facilities. The benefit of this approach was illustrated via a disinfection byproduct (DBP) management case study in northern Colorado. In this case study, the tool generated a suite of chemical dosing strategies that improved the reliability and adaptability of treatment plant operations. Moreover, these results offered the utility with insights into the tradeoffs among competing treatment objectives, such as operational costs and risks of regulatory violation.

The development of the decision support tool was made possible through innovations in stochastic water quality methods, water treatment simulation, multi-objective optimization, and interactive visualization techniques. In this project, the researchers advanced stochastic water quality methods to generate realistic scenarios of source water quality and characterize water quality uncertainty. Regarding simulation and optimization, the researchers coupled the EPA Water Treatment Plant Model—developed to estimate DBP formation and DBP precursor removal—with a multi-objective evolutionary algorithm to discover safer and more efficient operational decisions. Lastly, the researchers developed interactive visualization techniques to enhance data exploration and treatment decision making.

ES.5 Related WRF Research • Advanced Techniques for Monitoring Changes in NOM and Controlling DBPs under Dynamic

Weather Conditions (project 4422) • Impacts of Climate Change on Honolulu Water Supplies and Planning Strategies for Mitigation

(project 4637) • Implications of Climate Change for Adaptation by Wastewater and Stormwater Agencies (project

1334) • Water Utilities and Climate Change: A Research Workshop on Effective System Adaptation (project

4228) • Water/Wastewater Utilities and Extreme Climate and Weather Events: Case Studies on Community

Response, Lessons Learned, Adaptation, and Planning Needs for the Future (project 1338)

An Integrated Modeling and Decision Framework to Evaluate Adaptation Strategies for Sustainable Drinking Water Utility Management Under Drought and Climate Change 1

CHAPTER 1

Introduction The main objective of this project is to develop an integrated framework that enables understanding the processes that impact water quality under a suite of climate and natural hazards, quantifies the attendant uncertainty, and informs a decision support tool that enables evaluation and selection of mitigation strategies considering socio-economic impacts and tradeoffs. To accomplish this goal, four main areas were evaluated:

1. Understand the flow and sediment generation from drinking water supply watersheds in response to scenarios of hydroclimatological extremes and natural hazards.

2. Understand the mobilization and transport of DOM and sediments (i.e., turbidity), and in some cases nutrients (associated with algal growth) through the watershed and eventually to the water treatment plant.

3. Develop source water thresholds for turbidity and DBP precursors based on regulatory constraints in the finished water and using stream water quality data with extreme value theory predict water WQ threshold exceedances.

4. Evaluate a suite of adaptation and operation strategies (e.g., watershed management, wildfire mitigation, WTP modifications, etc.) along with their economic, societal and policy implications - with multi-objective optimization and multi-criteria analysis tools.


CHAPTER 2

Activity 1: Estimating Watershed-Scale Response of Post-Fire Sediment

2.1 Overview Changes in climate and land-cover represent key uncertainties for the quality and quantity of influent water. Soil erosion adds constituents to streams, altering water chemistry and streambed morphology, which can impact drinking water treatment and water resources infrastructure. The overarching goal of Activity 1 is to model post-fire erosion and sediment transport, to understand loading relevant for water treatment. This work is divided into three phases with the following objectives:

Objective 1: Develop a generalized model of watershed-scale streamflow and suspended sediment capable of modeling the impacts of climate and land cover disturbances on both quantities.

Objective 2: Evaluate the sensitivity of sediment response to wildfire and our ability to detect sediment signals from the observational record alone.

Objective 3: Measure sediment responses across a range of topographic slopes, wildfire intensities, and rainfall intensities in the laboratory.

Addressing the above objectives has led to several key insights that are described in the following sub-sections. The main focus of this analysis was towards the response of mountainous source watersheds, looking at both simulated and observed data from the past several decades.

2.2 Developing a Generalized Model of Watershed-Scale Streamflow and Suspended Sediment The research presented in this section represents the work of GRA Jenna Stewart under the guidance of co-PI Livneh, resulting in an MS thesis and journal publication (Stewart, 2017; Stewart et al. 2017a, 2017b). A large amount of research (Morgan, 2009; Toy et al., 2002) suggests that the key factors in driving Suspended Sediment Load (SSL) are conditions of climate, topography, soil texture, vegetation and land use, as shown in Figure 2-1.

The overarching modeling framework was developed to realistically portray these factors. A consistent hydrologic structure was implemented, within which a set of diverse sediment transport algorithms could be evaluated, as a way to acknowledge the structural uncertainty in predicting SSL amidst the differences across widely-used algorithms.

4 The Water Research Foundation

Figure 2-1. Portrayal of the Four Basic Factors that Control Soil Erosion, Which Motivate the Model Development

for This Study.

In total, we applied five erosion and suspended sediment load algorithms to predict SSL. These algorithms were provided with boundary conditions and dynamic meteorological and hydrological inputs from a common hydrologic framework to quantify uncertainty and evaluate predictability within mid-sized, forested catchment settings (> 1,000 km2). The algorithms were chosen from among widely used sediment models, including empirical models: Monovariate Rating Curve (MRC; Gray and Simoes, 2008) and the Modified Universal Soil Loss Equation (MUSLE; Arnold et al., 1998; Wischmeier and Smith 1960), a stochastic model: the Load Estimator (LOADEST; Runkel et al., 2004), a conceptual model: the Hydrologic Simulation Program—Fortran (HSPF; Bicknell et al., 1996; Johanson and Davis, 1980), and a physically based model: the Distributed Hydrology Soil Vegetation Model (DHSVM; Burton and Bathurst, 1998; Wicks and Bathurst, 1996). We coupled the algorithms with the Variable Infiltration Capacity Model (VIC; Liang et al., 1994), using hydrologic and meteorological inputs and fluxes generated from VIC. This five-algorithm structure is shown in Figure 2-2.

Figure 2-2. Illustration of the Five-Algorithms Considered in This Study within the VIC Hydrologic Model, as a

Consistent Framework.

The model framework was developed and evaluated over three catchments on the Colorado Front Range: a sub-basin of the Cache La Poudre catchment delineated to USGS gage 06751490 North Fork Cache La Poudre River at Livermore, CO (CLP-L); a larger sub-basin of the Cache La Poudre catchment delineated to USGS gage 06752000 Cache La Poudre River at Mouth of Canyon near Fort Collins, CO (CLP-F); and a sub-basin of the Clear Creek catchment delineated to USGS gage 06719505 Clear Creek at Golden, CO (CC) (Figure 2-3).

Erosion

Topography

Climate

Soil Type

Vegetation

VIC Streamflow

MRC MUSLE LOADEST HSPF DHSVM


The VIC model was implemented at a 1/16˚ spatial resolution (~ 6 km) on a daily time-step, leveraging an existing gridded dataset (Livneh et al., 2015). This dataset includes meteorological forcing for precipitation, maximum temperature, minimum temperature and wind speed, as well as basic soil, vegetation, and topographic inputs for VIC. The RVIC module was used to route simulated streamflow through each catchment by solving the linearized Saint-Venant equations using a unit-hydrograph approach for each grid cell (Lohmann et al., 1996).

Figure 2-3. Study Domain. North Fork of Cache La Poudre at Livermore (CLP-L, USGS 06751490, 1,393 km2), Cache

La Poudre at Mouth of Canyon near Fort Collins (CLP-F, USGS 06752000, 2,017 km2) and Clear Creek at Golden, CO (CC, USGS 06719505, 1,024 km2) Catchments Located within the Colorado Front Range Overlaid by 1/16˚

Resolution VIC Grid Cells. Calculated upstream reservoir storage was 35%, 25% and 8% of mean annual streamflow, respectively. Shaded

regions indicate catchment areas, red dots denote USGS streamflow and SSL gauges, and yellow dots show nearby cities.

Erosion processes generally occur at the hillslope scale, e.g., on the order of meters. Yet, there is a need to apply erosion equations at catchment scales, e.g., on the order of kilometers. Such an application is computationally expensive and inefficient. Therefore, we developed a critical area approach, to isolate model computations over productive regions where soil erosion is most likely to occur, thereby minimizing computational expense, while capturing important erosion processes at the appropriate scales. We calculate erosion from the MUSLE, HSPF and DHSVM using the characteristics of individual hillslopes (e.g., calculated from a 10 m resolution DEM) incorporating hydrologic inputs from the VIC grid cell, but scaled down to the hillslope dimensions. We subsequently up-scaled the hillslope sedimentation rates on the basis of their relative areas within the catchment.

Critical areas were defined by three general criteria: (i) areas exceeding a slope steepness threshold of greater than 15˚ (Larsen et al., 2016), land-cover, (ii) either forest cover or grassland, as well as (iii) proximity to the channel, exploring 100 m and 500 m buffering distances. Figure 2-4 illustrates how the intersection of the above factors produces a set of critical areas that greatly reduce the total area and hence a considerable reduction in computational expense.

CLP-L: North Fork of Cache La Poudre at Livermore, CO

CLP-F: Cache La Poudre at Mouth of Canyon near Fort Collins, CO

CC: Clear Creek at Golden, CO

VIC Grid Cell


Figure 2-4. Critical Areas Computed for Using Slope Threshold of 15˚ and Above, Vegetation Types Including Shrub, Grassland, Bare Ground (SGB), and Forest, and Stream Proximities of 100 m and 500 m. Critical Area

Included Four Estimated Regions.

A calibration was conducted using the Borg multi-objective evolutionary algorithm (Hadka and Reed, 2013) using the parameters listed in Table 2-1. The resulting performance is shown in Figure 2-5. Performance of optimized parameter sets from the calibration were validated over an ancillary period, as well as in an inter-basin transfer to a separate catchment to explore parameter robustness. Results highlight the tradeoffs in sediment prediction across a range of algorithm structures and catchments. Model performance showed consistent decreases when parameter sets were applied to time periods with greatly differing SSL magnitudes than the calibration period.

An interesting tradeoff was observed, where the algorithms in the joint calibrations performed worse in NSE and Bias than the individual calibrations, which we attributed to the dynamic dependence of SSL on streamflow, as different equilibrium states with streamflow can be optimal for different algorithms. Solutions from the joint algorithm calibration favored simulated streamflow partitioning into runoff and baseflow that differed from the streamflow-only calibration. Inter-basin transferability performance was highest in algorithms with lower dependence on streamflow performance, the HSPF and the DHSVM. We therefore consider these more sophisticated algorithms to be more suitable for predicting SSL response under different climate conditions due to their inclusion of physical conditions, precipitation rates and vegetation coverage, rather than solely relying on streamflow as in the case of the MRC.

The jointly calibrated model was applied over the period 1950-2013, to explore the range of simulated uncertainty in SSL and to evaluate long-term sediment variability. Figure 2-6 shows the range in annual sediment loading across the five algorithms and contrasts this with a traditional historical sediment yield estimate (Jansson, 1988). The dynamic multi-algorithm estimate of sediment loading is deemed superior to the traditional static estimate, as it is able to capture the impact of large climate events, such as the one that occurred in the early 1980s, on SSL. This exposition supports the future application and development of the multi-algorithm ensemble as a useful tool for evaluating SSL response to key drivers.


Table 2-1. Parameters Used in the Multi-Objective Optimization Routine for Each Calibration. Given that sediment modules are dependent on streamflow (VIC flow), sensitive VIC soil parameters were

incorporated into the individual module calibrations as well.

Individual Calibrations Joint Calibration

Parameter Definition Range VIC flow DHSVM MUSLE HSPF All

Binf Infiltration Capacity 0.0001 – 0.4

Ds Fraction of DsMax where non-linear baseflow occurs 0.0001 – 1.0

DsMax Maximum baseflow velocity 0.1 – 30.0

Ws Fraction of maximum soil moisture where non-linear baseflow occurs 0.1 – 1.0

C Baseflow curve exponent 1.0 – 2.0

Layer 1 Soil layer depth 1 0.1 – 0.3



K Factor Erodibility factor 0.02 – 0.6

C Factor Cropping management factor 0.0001 – 0.5

P Factor Conservation practice factor 0.0 – 1.0

K Index Erodibility index 19.0 – 32.0

D50 Median grain size 0.5 – 2.0

Soil Cohesion Soil cohesion 0.00075 –

15.0

JR Detachment exponent 1.0 – 3.0

AFFIX Attachment fraction 0.01 – 0.50

KS Transport coefficient 0.1 – 10.0

JS Transport exponent 1.0 – 3.0

KG Scour coefficient 0.0 – 10.0

JG Scour exponent 1.0 – 5.0

Critical Area Catchment critical area

CC: 0.002 – 0.04

CLP-L: 0.003 – 0.02


Figure 2-5. (a) Hydrograph for CC-Late with VIC Streamflow Computed Over the Calibration Period 1992 – 1997

for the Joint Algorithm Calibration with a Spin-up Year in 1991. (b) Absolute Value of Bias Performance for Calibration, Validation and Transfer Periods for All Algorithms in CLP-L, CC-Early and CC-Late.

Error bars represent the range in performance of the three solutions. Values of |Bias| > 200% are plotted as 200% for visualization purposes.

Figure 2-6. Application of the Model Ensemble for the Clear Creek Catchment, from 1950 to 2013 Using a Top

Performing Parameter Set from CC-Late. Error Bars represent the multi-algorithm spread, while the highlighted box shows a traditional estimate (Jansson

1988).

(a) (b)

Susp

ende

d Se

dim

ent L

oad

(tons

/km

2 /yr

)

Year


2.3 Evaluation of the Sensitivity of Sediment Response to Wildfire in a Mountainous Watershed, and the Detectability of Post-Fire Sediment Signals 2.3.1 Sensitivity of Sediment Response to Wildfire in a Mountainous Watershed In the first part of this evaluation, GRA Aaron Heldmyer and co-PI Ben Livneh investigated post-fire erosion modeling on the Colorado Front Range. Using observational data collected during the year following the 2012 High Park Fire in the Cache La Poudre (CLP) Basin near Fort Collins, CO, a reduced ensemble of sediment algorithms (relative to Section 2.2) were applied to two affected sub-regions, known as Hill Gulch and Skin Gulch, within the greater burned area. Three different algorithms (DHSVM, HSPF, and MUSLE) were applied, unified within the VIC model, as described in Section 2.2.

2.3.1.1 Study Area and Wildfire Description The CLP is a mid-sized (2,017 km2) basin that is forested, predominantly by ponderosa pine (Pinus ponderosa) and other mixed conifers, with minimal land use affecting water quality (Heath and Oropeza, 2016). It was here that a wildfire, triggered by a lightning strike and fueled by dry stands of trees produced by beetle infestation, burned over 87,284 acres from 9 June to 1 July 2012, becoming the second-largest fire in recorded Colorado history in terms of spatial extent (Figure 2-7).

In the years following the fire, collaborators at Colorado State University (CSU) installed and maintained an observation network within two affected watersheds of the CLP, referred to as Hill Gulch (14.3 km2 drainage area) and Skin Gulch (15.5 km2). Rainfall was monitored with a network of tipping bucket rain gauges placed throughout the watersheds, and single or double sediment fences were placed to trap eroded sediments derived from converging hillslopes across the study area (Kampf et al., 2016). Regular observations of vegetation cover were also made during the post-fire recovery period. These data were used to inform the sediment models applied in this study.

Figure 2-7. Depiction of the 2012 High Park Fire Scar (Red) atop the CLP Basin.

2.3.1.2 Model Descriptions The sediment models used in this study (DHSVM, MUSLE, and HSPF) were developed for the studies described in Section 2.2, estimating sediment loading at several locations on the Colorado Front Range (Stewart et al., 2017b). While DHSVM and HSPF were applied in the same fashion as Section 2.2, MUSLE


was modified to facilitate a more direct comparison to another model: the Revised Universal Soil Loss Equation (RUSLE), applied by CSU collaborators. One factor, R, which accounts for runoff, is calculated as 𝑅 = 11.8(𝑄 × 𝑞 ) . (2-1)

where Qvol is the volumetric runoff from a storm event, and qpeak is the peak runoff from a storm event. The coefficient and exponent values were developed by (Williams, 1975) by fitting MUSLE to sediment data from 18 study watersheds in southern USA. These fitted parameters yielded sediment output 3 orders of magnitude higher than what was expected from previous simulations and observations in the CLP, and so the exponent was adjusted to 0.15 by fitting the equation to data obtained from the CLP.

2.3.1.3 Methodology An initial sensitivity analysis was conducted to identify parameters within each sediment model that have a large influence on the magnitude of sediment production using the Sobol Method (Sobol’ et al., 2007). Parameters associated with each sediment model were varied across their effective range determined by (Stewart et al., 2017b) with a sample size of n = 10,000 parameter combinations. Results from this analysis are shown in Figure 2-8.

Following the sensitivity analysis, a model simulation was run for the time period 1 June 2013 to 31 October 2013 over Hill Gulch and Skin Gulch. It is widely known that sediment models can suffer from scale-dependency issues, that is varying the size of the model spatial subdivisions can result in different total sediment loads when the subdivisions are aggregated to the overall area. In order to uncover these scale dependencies, different subdivisions were tested and compared. The subdivision sizes, hereafter referred to as Critical Source Areas (CSAs), were chosen to be 0.5, 1, 2.5, 5, 10, and 25 ha resolution, and the sediment models were run separately for each of these CSAs.

Modifications to default parameters were made based on observations gathered in Hill Gulch and Skin Gulch, as well as adapted from the Kinematic Runoff and Erosion Model (KINEROS2) and RUSLE parameter values set during previous model runs by CSU collaborators. Vegetation cover was scaled using bare ground percentage estimations made at the site. Parameters such as hillslope area, mean elevation, hillslope length, saturated hydraulic conductivity, and soil cohesion were adapted from KINEROS2 parameters provided by CSU collaborators, while MUSLE parameters were adapted from the prior experimentation with RUSLE as well as data collections such as the National Resource Conservation Service (NRCS) Soil Survey Geographic Database (SSURGO). Five-minute rain gauge data from the nearest rain gauge to each CSA were combined with average daily wind speed, maximum daily air temperature and minimum daily temperature from the (Livneh et al., 2015) dataset to generate a complete forcing dataset.


Figure 2-8. Sobol Sensitivity Analysis Results for Sediment Models DHSVM (SHESED), HSPF, and MUSLE.

The main effect is obtained by varying one parameter in isolation, while the total effect includes all possible synergistic terms with all other variables.

Additionally, parameters known to have an outsized influence on simulated sediment production, but whose values were not known or were unobtainable, were varied to account for uncertainty. These parameters include the Binf and the second soil layer depth for the VIC model, as well as the coefficient and exponent for scour of the matrix soil used in HSPF. A total of 100 model runs using combinations of these parameters were performed.


2.3.1.4 Results Results from these model runs are organized in Figure 2-9. These results show a clear increase in sediment yield for both domains with coarsening spatial resolution. These estimates and their spatial dependence are comparable to as-yet-unpublished model results from CSU collaborators over the same domains, as well as observational data from (Kampf et al., 2016). The inter-comparison of SSL simulations represents a means to identify robust response patterns, as well as an opportunity to evaluate best-practices for accurate post-fire sediment modeling at the basin scale.


Figure 2-9. Hill Gulch (Top) and Skin Gulch (Bottom) Estimated Total Sediment Yields.

CSA sizes—the resolution at which the domains were modeled—are shown on the horizontal axis, and the total sediment yield across the entire gulch and modeled time period, are shown on the vertical axis in log-scale.

2.3.2 Wildfire-Based Sediment Signal Detection in the West In the second part of this analysis, we sought to evaluate whether post-fire sediment signals could be detected using exclusively observations, by characterizing streamflow and sediment relationships through commonly used rating curve parameters at a diverse set of gaged locations across the western

1e+03

1e+05

1e+07

0.5ha 1ha 2.5ha 5ha 10ha 15ha 25haCSA

Sedi

men

t Loa

d (M

g)

ModelDHSVM

HSPF

MUSLE

Hill Gulch Total Sediment Yields

1e+03

1e+05

1e+07

0.5ha 1ha 2.5ha 5ha 10ha 15ha 25haCSA

Sedi

men

t Loa

d (M

g)

ModelDHSVM

HSPF

MUSLE

Skin Gulch Total Sediment Yields


U.S. We identified a relatively undeveloped basin from the GAGES-II (Falcone, 2011) dataset (the Rio Puerco near Bernardo, NM) that has experienced five observed fire events between 1999 and 2014, and used this basin as a testbed for evaluating a set of increasingly data-intensive approaches for detecting a post-fire sediment response signal.

2.3.2.1 Study Area and Data Products The western U.S. was selected as the study domain for evaluation of regional variations in hydrologic and geomorphologic responses to wildfire, as well as in relationships between streamflow and sediment transport (Figure 2-10).

Figure 2-10. Western U.S. Study Domain.

All wildfire ignition points (recorded 1985-2016) are shown in red. All GAGES-II basins are also displayed (Falcone, 2011). Note that some basins are nested.

Monitoring Trends in Burn Severity (MTBS; Eidenshink et al., 2007) data were coupled with USGS GAGES-II basin information, and USGS National Water Information System (NWIS) stream gage data to support this regional analysis. Merging these datasets yielded a total of 255 wildfires across 187 gaged basins that have been observed for both streamflow and sediment, representing the wildfires and basins analyzed here.

A test-bed basin was isolated in order to evaluate and refine the methodology for quantifying post-fire sediment signals. This site was selected from the total 187 gaged basins based on several criteria including: minimal upstream regulation (e.g., dam density and storage) from GAGES-II, number of observed fires, size of observed fires as a percentage of basin size, and excluding snowmelt-dominated basins given complexities associated with snowmelt timing and transport uncertainty. This process yielded the USGS Gage 08353000 for Rio Puerco near Bernardo, NM as the most suitable candidate. Located west of Albuquerque, NM, this basin (Figure 2-11) contains little infrastructure development,


with a basin area of 15,725 km2, 23 dams total (0.15 dams/km2), and a dam storage density of 3.9 ML/km2.

Figure 2-11. USGS Gage 08353000 (Rio Puerco near Bernardo, NM).

Basin extent is shown in blue. Fires observed for both sediment and streamflow are shown in red, and the gage location can be seen towards the south as a black and white circle.

A total of five fires (two prescribed burns and three wildfires) were observed within the drainage area during the period 1999-2014 when the gage was collecting both streamflow and sediment data. The largest fire, which occurred June 12th, 2004, consumed approximately 37.4 km2 (0.23%) of the basin areal extent. The second and third largest fires occurred May 30th, 2008 and June 3rd, 1999, respectively, consuming 17.8 km2 (0.11%) and 13.1 km2 (0.08%) of the basin extent.

2.3.2.2 Methodology The Monovariate Rating Curve (MRC), also known as the sediment rating curve, is an empirical method for estimating sediment loading exclusively as a function of streamflow. The most common form is that of a power relationship:

SSL = aQb (2-2)

where SSL is suspended sediment loading, Q is streamflow, a is a coefficient for the intercept, and b is an exponent for slope (Gray and Simoes, 2008). USGS daily streamflow and sediment data were fit to this model to create a summary relationship characterizing the expected sediment loading per unit streamflow for all 187 suitable basins with available data.

A cascade of increasingly data-intensive techniques for detection of a post-wildfire signal was applied to the single site in New Mexico, with the intent of both conclusively attributing, and later accurately predicting, the presence and magnitude of SSL response at a basin outlet due to a wildfire event. This methodology is driven by two key motivations: (i) the necessity for actionable post-fire sediment response information under conditions of data scarcity, and (ii) the need to identify influences on sediment response that may be generalized to the western U.S., such that inferences about post-fire sediment loading may be drawn for basins without an abundance of observational data. The optimal


result would achieve an accurate prediction of sediment response magnitude due to a wildfire event using as little input data as possible.

The first method analyzed stream gage time-series data alone using a statistical technique called intervention analysis. Intervention analysis is commonly used to uncover the effects of an intervention, or an impactful event, on a time-series. Typically, an Auto Regressive Integrated Moving Average (ARIMA) model (Box et al., 2015) is applied to pre-event time-series data and used to forecast theoretical post-event data. This essentially contrasts the actual post-event observations with the model scenario in which pre-event data are used to forecast post event response as if the event did not occur. An ARIMAX model is an extended version of ARIMA, and additionally includes one or more exogenous predictor variables. The equation for ARIMAX can be written as follows: 𝑌 = 𝐶 + 𝑣(𝐵)𝑋 + 𝑁 (2-3)

where 𝑌 represents the dependent variable, 𝑋 is the independent variable, C is the constant term, 𝑁 is the stochastic disturbance (i.e., the ‘noise’), 𝐵 is the backshift operator, and 𝑣(𝐵)𝑋 is the transfer function that allows X to influence Y through a distributed lag (Peter and Silvia, 2012). The transfer function can then be described as: 𝑣(𝐵)𝑋 = (𝑣 + 𝑣 𝐵 + 𝑣 𝐵 + ⋯ )𝑋 (2-4)

This model was applied individually to each of the five wildfires at the study basin to train the model. The ‘pre-fire’ period used for training was defined as the start of the observational record up to the fire date of ignition. Determination of model parameters, including the lag order (number of lagged observations), degree of differencing (the number of times raw observations were differenced to remove non-stationarity), and the moving average order (the size of the moving average window), hereafter referred to as the variables p, d, and q, respectively, were made using the entirety of the observational record through step-wise Akaike Information Criterion (AIC) model ranking (Akaike, 1998). A forecasted period of one year was selected to encapsulate a full post-fire season.

The second technique added daily precipitation data from Daymet to create a basin mean areal precipitation as another predictor variable for the ARIMAX model. Several studies (Knapen et al., 2007; Momm et al., 2018; Moody et al., 2008) identify precipitation intensity thresholds that must be overcome for the initiation of concentrated erosion and debris flows. Identifying this threshold may be particularly important for the prediction of suspended sediment following a fire, as a signal may not be readily detectable for storm events below these thresholds, whereas disproportionately high volumes of sediment may be transported during the first post-fire storm above the threshold. Initially, precipitation intensity was plotted against post-fire stream gage SSL to find evidence of this threshold by inspection. Then, precipitation data were used to filter out days during which no precipitation occurred (as it is assumed no overland flow was present to transport sediment into the channel), and subsequently added as a predictor variable for the ARIMAX model.

The third technique retains the spatial information of the precipitation data, rather than aggregating over the basin area. Fire extents from MTBS for the five events were compared against the gridded storm extents to find storms co-located with affected areas. A detectable signal may be more easily found and isolated when knowledge of whether a storm precipitated over a burned area is added. Thus, periods of time when no precipitation occurred over a wildfire-affected area were filtered out, and the remaining data were again modeled to detect pre- vs. post-fire differences in SSL.


2.3.2.3 Results The results of fitting the MRC to observational data are summarized below. Figure 2-12 summarizes coefficient ‘a’ from Equation 2-2 across the domain.

Figure 2-12. Spatial Summary of Coefficient ‘a’ from the MRC Curve Fitting Process.

GAGES-II basins across the U.S. West are shown. Grayed-out basins are those with not enough available data.

The spatial variation of exponent ‘b’ in Equation 2-2 is shown below in Figure 2-13.

Figure 2-13. Spatial Summary of Exponent ‘b’ from the MRC Curve Fitting Process.



Finally, the correlation coefficient, or R2, of the fit between MRC-predicted and observed sediment is shown below in Figure 2-14.

Figure 2-14. Spatial Summary of R2 from the MRC Curve Fitting Process.


The MRC fit exhibited an R2 of greater than 0.75 for 140 out of 187 basins across the domain. Coefficient ‘a’ remained relatively low across all basins, likely indicative of the ‘flashy’ nature of streams located in the arid West. Results for exponent ‘b’ are arguably the most interesting: high relative variability in its magnitude points towards basin-level differences that may be affecting its value. However, comparing ‘b’ against several basin metrics such as relief, mean slope, basin size, and mean flow (not shown) did not reveal any strong explanatory skill.

Overall, the regional analysis revealed trends in the rating curve exponent ‘b’ that may be worthwhile in future work to investigate further in order to identify correlations with additional basin characteristics not analyzed here. Future work will explore other predictive variables, such as land cover, geology, and climate, and their covariance with the value of ‘b.’

For the single site analysis at Rio Puerco near Bernardo, NM, flow was multiplied with SSC to obtain SSL, which represents the volume of sediment flowing through the channel. Before fitting to an ARIMA model, these time-series were first examined for the presence of long-term trends and seasonality. An Augmented Dickey-Fuller (ADF) test was applied to test for stationarity. The null hypothesis of non-stationarity over the time period was rejected with p < 0.01, indicating the absence of any longer-term trends that would need to be included in the model. To identify seasonal cycles in the data, plots of the auto-correlation function (ACF) and partial auto-correlation function (PACF) for SSL observations at the gage were examined (Figure 2-15).


Figure 2-15. Autocorrelation Function (Left) and Partial Autocorrelation Function (Right) for SSL.

The horizontal axis shows lag in days, and the vertical axis shows correlation.

Significant auto-correlation exists for lags up to approximately one week and, to a lesser extent, for a one-year lag, denoting the presence of an annual cycle for the time-series. A model optimization method using the R function ‘auto.arima’ was employed, which identifies a best model fit using the AIC method of ranking a model based on relative quality. Using the ARMIAX model with streamflow as an additional predictor, a non-seasonal AR order, degree of differencing, and MA order (p, d, q) = (2, 1, 2) was identified as the top performer through AIC best model selection, using the full record of observations.

For each wildfire, the model was fit to the pre-fire data and used to forecast the full year following the event. This was then compared against the true observed series for those 12 months, and the mean difference was taken to be the effect of the wildfire event. The results of this process are presented below in Figure 2-16.


Figure 2-16. Model-Predicted SSL versus Observations Following Five Observed Wildfires at the Study Site Near

Bernardo, NM. Pre-fire data used to inform the model is shown in black, the observed post-fire data is in red, model-predicted

values are shown in blue, and the 95th percentile upper confidence bound is shown in light blue.

Differences between post-fire observational and forecasted SSL during each post-fire, year-long period are summarized below in Table 2-2.

Table 2-2. Summary of Mean and Maximum Differences between Forecast Mean and Observed SSL (in kg/s) during Post-Fire Periods.

Ignition date and fire spatial extent are presented in the first column. Percent differences between forecasted and observed are also shown. An asterisk (*) denotes prescribed fires.

Fire Event Forecasted Mean SSL

Observed Mean SSL

Percent Difference

Forecasted Maximum SSL

Observed Maximum SSL

Percent Difference

1999-06-03

(13.15 km2) 70.82 77.16 8.22% 1656.86 1744.32 5.01%

2004-06-12

(37.40 km2) 20.02 20.63 3.00% 1525.73 1842.66 17.20%

2008-05-30*

(17.83 km2) 12.82 39.05 67.17% 340.74 4144.65 91.78%

2012-04-23

(6.4 km2) 7.57 13.23 42.78% 288.75 439.03 34.23%

2014-05-02*

(5.77 km2) 49.07 56.83 13.65% 997.22 1299.03 23.23%


Next, the additional incorporation of precipitation data was tested to first determine if a specific post-fire erosion threshold could be found from the data, then added as a predictor to the ARIMAX model. First, a cross-correlation was applied between streamflow and precipitation, shown in Figure 2-17.

Figure 2-17. Cross-Correlation between Streamflow and Precipitation.

The horizontal axis shows the lag in days between the two series of streamflow and precipitation. The vertical axis shows correlation.

Prior literature (Valois et al., 2017) have taken the time of maximum correlation to be representative of the time of concentration for the watershed. The magnitude of the maximum correlation can also be taken as an indicator for the strength of the relationship between streamflow and precipitation. This basin shows a maximum lag time of 7 days, and a maximum correlation of 0.31, indicating a relatively slow concentration time and somewhat weak relationship between streamflow and precipitation. These initial findings point to the possibility of a groundwater-fed watershed and/or the presence of other attenuating sub-basin processes that may store precipitation for a period of time before reaching the basin outlet.

Precipitation was incorporated into the ARIMAX model alongside streamflow as an exogeneous variable and used to forecast over the same post-fire periods. Adding precipitation as an additional parameter did not yield any significantly different forecast predictions, reflecting the weakness of basin-averaged precipitation as a predictor for SSL at the basin outlet. Among all fires, the largest difference between post-fire predicted mean SSL from the model with and without precipitation added was 0.71 kg/s. The largest difference in maximum SSL was 6.52 kg/s.

Finally, only precipitation that fell over a fire-affected area was analyzed. For each fire event, a unique time series of precipitation was used that represented precipitation falling within the extent of the fire perimeter provided by MTBS data. In order to determine if a sediment signal at the basin outlet could be attributed to sediment delivered from a burned area, the top 10 largest instantaneous SSL magnitudes were examined for each post-fire period.

Of the 50 maximum instantaneous sediment loadings (10 for each of 5 observed fires in the basin), 20


showed no prior precipitation over the burned area, indicating that at least 40% of these large sediment signals could not be attributed to sediment originating from burned areas.

Replacing precipitation over the entire basin with precipitation over burned areas alone again yielded few differences from the original model. Among all fires, the largest difference between post-fire predicted mean SSL from the model with and without burned area precipitation added was 1.28 kg/s. The largest difference in maximum SSL was 21.07 kg/s.

Overall, the gage-only method revealed a post-fire sediment signal for several of the fires, based on the difference between the observed and forecasted time series. The presence of these differences are encouraging, and underscore the potential for establishing an estimate of post-fire sediment response based on gage data alone.

Importantly, data limitations would limit the applicability of this method, as most gages do not provide consistent sediment measurements. Including precipitation did not significantly affect the model. However, examining the link between cross-correlation in streamflow and precipitation versus sediment response for other basins may be worthwhile to investigate, as it could be posited that a shorter time of concentration, paired with a high correlation between precipitation and streamflow, may be indicative of a well-connected basin capable of delivering sediment with relative efficiency. Underpinning this hypothesis is the notion that a well-connected basin with little attenuation of streamflow from storm to outlet may also offer fewer obstacles for sediment as well. In these cases, strong post-fire SSL magnitudes may be more common.


CHAPTER 3

Activity 2: Impact of Heating Temperature on the Character of Water-Soluble Constituents from Organic and Mineral Soils 3.1 Introduction Wildfires play an integral role in the ecological productivity and health of forested areas (Kondolf et al., 2014; Reneau et al., 2007). However, wildfires can impact public safety by degrading water quality and severely altering several components of forest biomass, including vegetation, detritus litter, and mineral soil. By impacting surface conditions in watersheds, wildfires have been shown to alter hydrologic pathways, suppress subsurface flow, and produce overland runoff as the dominant flow path (Murphy et al., 2018). Entrainment of both sediment and ash is especially prevalent when wildfire events have sufficient proximity to waterways.

As a result, water quality is inextricably impacted by the post-fire landscape, particularly when less-dilute, lower order streams, are proximal to wildfire events. Moreover, wildfire regularity is anticipated to persist or worsen in the next century. Conditions for the pervasiveness of wildfires have become exacerbated by climate change (Fried et al., 2004; Westerling et al., 2006), changes in land use, and the over-densification of forests through prolonged wildfire prevention management (Sexton, 2006). Within the last year alone, nearly 71,500 fires have consumed over 10,000,000 acres of forested land in the United States (National Interagency Figure Center, 2018). Marked increases in the frequency and season length of forest wildfires in the last 40 years (Westerling et al., 2006) are the impetus for a large push in the research community towards understanding the impacts and implications of these environmental perturbations.

The introduction of ash into surface waters after a wildfire has been identified as contributing to elevated dissolved constituents in water bodies. Accordingly, there is a necessity for a clear and fundamental study regarding the capacity for organic litter and soil to release ecology-altering and water quality degrading constituents into the environment after a burn event. Ash is defined as a mixture of charred organic and inorganic material, that is derived from the heating of soil and plant material (Bodí et al., 2014). However, the specific impact of ash on water quality in forest catchments is difficult to characterize because rarely is ash decoupled from mineral soil as they are both delivered into streams and water bodies following a post-fire storm event (Santín et al., 2015). Due to the complications of defining ash, this study will focus primarily on the release of organic matter, nutrients, and soluble bio and lithogenic elements.

The changes in quantity and character of dissolved organic matter (DOM) resulting from wildfires is likely a mixture of both physical and chemical processes. Moreover, there is evidence that suggests heating chemically alters soil organic matter (SOM), thus creating smaller organic compounds that are more readily mobilized by surface runoff (Santos et al., 2019). These alterations to the A and O soil horizons within a burned catchment invariably effect water quality by increasing downstream DOC and disinfection byproduct precursor loads.

Previous studies have revealed that inorganic nutrient levels increase within surface waters receiving runoff from wildfire impacted areas. The Hayman fire, which burned the Upper South Platte watershed


of Colorado’s Front Range, has been linked to increased inorganic nutrient levels in streams throughout the catchment. Monitoring of the affected streams revealed that peak inorganic nutrient levels persisted several years after the burn event, suggesting that alterations to soil nutrient cycling have long-term impacts on surface water quality (Rhoades et al., 2011). Furthermore, understanding the amount of organic and inorganic nitrogen released after a simulated heating event will provide critical information on the effects of low to high intensity wildfires on water quality within a burned watershed.

This study introduces a highly controlled, systematic, and deconstructed study to decouple the impacts of heat-altered forest floor litter (organic layer) and soil (mineral layer) on water quality. While the complete matrix of wildfire dynamics is not reflected in the proceeding simulated wildfire technique, the value in performing a controlled experiment to make visible processes that are otherwise lost cannot be overstated. Mineral and organic layer samples were collected in the Colorado Front Range within the Boulder Creek Watershed. Collected materials were dried and heated to five temperatures using a muffle furnace. Changes in the properties of the heated material were assessed, including total C and total N. Heated material was then leached in ultra-pure water and the resulting leachates were analyzed for the following parameters: DOC, total dissolved nitrogen (TDN), water extractable organic carbon and organic nitrogen (WEOC and WEON), and soluble bio- and lithogenic elements. Three mineral soil replicate samples were collected at three different geographical locations within the Colorado Front Range, totaling nine samples. One composite litter sample, representing the organic layer, was collected at each location, totaling three samples.

3.2 Materials and Methods 3.2.1 Soil and Litter Sampling Mineral and organic layer samples were collected near Boulder, Colorado (Figure 3-1) in the summer of 2016. The Boulder Creek Watershed, located East of the Continental Divide within the Front Range of the Colorado Rocky Mountains, spans approximately 1160 square kilometers and comprises headwater regions that supply water for over 300,000 residents (U.S. Census Bureau, 2018). Elevations in the Boulder Creek Watershed range from 1480 meters to 4120 meters above sea level, encapsulating several climatic and vegetal zones including plains, foothills, montane, subalpine, and alpine (Murphy et al., 2003) . The watershed spans two physiographically distinct regions. The upper basin, described by steeply sloping valleys and mountainous terrain (Murphy et al., 2003) is part of the Southern Rocky Mountain Province. The lower basin, described by gentle sloping terrain to the east, is encompassed by the Southern Rocky Mountain Province. The mountainous regions receive precipitation mainly in the form of snow in the winter and spring months; however, this area also experiences episodic high-intensity convective storm events during summer months, which is one of the major drivers contributing to post-fire water quality impairment (Murphy et al., 2015). Summer convective storms are known to produce rainfall intensities greater than 10 mm/h, eliciting substantial runoff responses in fire-impacted Colorado watersheds. Wildfire represents one of many factors that produce variation in water quality within the Boulder Creek Watershed and the increasing prevalence, size, and severity of fires are compelling reasons to gain more information about how they impact raw water sources.

Sampling locations were chosen based on their proximity to one Denver-owned and one Boulder-owned water storage reservoir. Site 1 is located close to Lakewood Reservoir, which is one of two reservoirs that supply 40 percent of the city of Boulder’s water supply and is diverted to Betasso Water Treatment Plant through the Lakewood Pipeline. Site two is proximal to Gross Reservoir, which serves as a combined storage and water regulating facility for water that flows under the continental divide and diverts water to the Denver Metropolitan area through Moffat Tunnel. Site three was chosen as an intermediate site representing an area of lower elevation. Sites 1 and 2 are located in the upper basin


while Site 3 was meant to represent the lower basin. At the time of sampling in September 2016, Boulder, CO had a recorded mean monthly precipitation of 114 mm and a mean annual precipitation of 1724 cm in 2016 (NOAA). The mean monthly temperature was 18.3˚C with a low of 6.1˚C and a high of 27.3˚C in September 2016 and a mean annual temperature of 11.1˚C.

Three mineral soil subsamples were collected at each location. Triplicate samples were taken within 10 m from one another from locations of varying canopy coverage. Soils A, B, and C were collected north of Nederland, Colorado (NED) at an elevation of 2601 m. This sampling area had no closed canopy with understory vegetation characterized by blue grama grass (Bouteloua gracilis), needle-and-thread grass (Hesperostipa comate), and western wheatgrass (Pascopyrum smithii). The Nederland soil series is moderately permeable and well-drained, characterized by a cobbly sandy loam. Soils D, E, and F were collected near the summit of Flagstaff Mountain outside of Boulder, CO, at an elevation of 1849 m elevation (FLG). All three sampling replicates were taken under closed canopy, which was characterized by coniferous forest stands comprised of ponderosa pine, Douglas-fir, and subalpine fir-Engelmann spruce (Picea engelmanni, Abies lasiocarpa). There was no prominent understory vegetation; however, a layer of fallen detritus consisting chiefly of pine needles was present and was removed prior to soil excavation. Soils G, H, and I were collected south of Gross Reservoir in Boulder Country, at an elevation of 2,222 m (GROSS). Soil G had no overstory canopy, Soil H had 30% canopy coverage, while Soil I had a closed canopy. The surrounding overhead vegetation varied in density comprising of four main coniferous tree species: ponderosa pine, lodgepole pine (Pinus contorta), limber pine, and Douglas-fir. Understory vegetation for the sampling area include a mix of grasses, forbs, and shrubs; largely dominated by cheatgrass (Bromus spp.) and Canada thistle (Cirsium arvense).

Organic layer litter samples were obtained in December of 2017. Forest floor material, referred to as the organic layer, is the uppermost soil horizon that comprises organic debris and some vegetation. The organic layer is highly susceptible to consumption and combustion during a wildfire. Because understory vegetation litter was relatively spatially homogenous, one large composite sample was taken at each site. For NED, where there was minimal overhead canopy cover and the understory was dominated by grass that was still intact during the fall and winter months, organic layer sampling included some grass, pine needles, pine cones, and small twigs. Litter contents at both FLG and GROSS were relatively similar, comprised of mostly pine needles, pine cones, pollen cones, and small twigs.


Figure 3-1. Sampling Locations (Indicated by Stars in the Boulder Creek Watershed).

3.2.2 Sample Processing Mineral and organic layer samples were processed immediately after sampling. Mineral soils were distributed on metal trays to a depth of approximately 1 cm and oven-dried at 100˚C for two hours to eliminate moisture and to suppress the survival of microbial communities present in the soil (Dunn et al., 1985) that may compromise sample integrity during storage and to avoid any impacts of antecedent soil moisture. Soil was then passed through both a 2 mm (No. 10) stainless steel sieve to remove large rocks and plant matter and through a 0.841mm (No. 20) sieve to remove smaller plant matter before storage.

Processing the organic layer samples involved a drying procedure similar to that of the mineral soil, in which litter was uniformly, and of minimal density, distributed over metal trays and dried for two hours at a temperature of 100˚C. The purpose of drying the organic layer samples was twofold: first, to achieve an even level of dryness between litter materials to allow the material to later be heated as consistently as possible and second, to optimize sample integrity for storage.

3.2.3 Heating Simulation Mineral soils were heated in an electric muffle to temperatures of 150, 250, 350, 450, and 550ºC in 90-mL porcelain dish crucibles using a Lindberg/Blue Box Furnace Model BF51442C with a Lindberg Furnace Power Supply Controller Model 59344. Organic layer litter was heated to temperatures of 150, 250, 350, and 450ºC in loaf pans using the same furnace. Results for 550 ºC were excluded due to limited sample volume, as litter heated to this temperature had significant mass loss as well as anticipated negligible levels of mobilized carbon and nitrogen. For this experiment, CTRL/150ºC, 250/350ºC, and 450/550ºC were considered.

All batches of samples were held at each temperature in aerobic conditions for two hours to ensure combustion completeness, as wildfire severity is often characterized by organic matter consumption


(Keeley, 2009) . Once cooled, burned samples were stored in either amber glass 40 mL vials (soils) or half-pint glass jars (organic litter) at room temperature. To ensure uniformity in mineral soil heating, 10 g soil per crucible (approximately 0.5 cm high) was heated in batches of 10 crucibles. Intact organic layer material was loosely placed in the loaf pans to ensure uniform heating. After heating, the organic layer samples were mechanically ground using an 8150 Enclosed Shatterbox for 15 seconds, producing a fine powder. The purpose of grinding the organic layer material was to homogenize the samples, which maximized the reproducibility and visibility of trends resulting from leaching tests, which are driven by mass. Because organic layer parent material is so heterogeneous in both composition and density, homogenization was required.

3.2.4 Soil and Litter Leaching Two experiments were conducted for sites A, B, D, and E (mineral layer) and NED and GROSS (organic layer) to determine an optimal leaching duration and leaching concentration (solid-to-solution ratio) as well as to investigate the leaching capacity of forest floor material. A kinetics test was performed to determine an optimal leaching time for the subsequent experiments. Processed mineral soil samples were leached in 100 mL of ultra-pure Milli-Q water separately for a contact time of one, two, four, six, eight, 10, 24 and 48 hours. DOC and TDN dissolution curves are shown in Figures 3-2 and 3-3. These results were obtained while being agitated on a VWR Standard Analog Shaker table. A fixed concentration of 5 g soil/L water was used throughout all the kinetics experiments so that both the DOC and TDN concentrations at the end of the reaction would not exceed the maximum detection thresholds of the total carbon and nitrogen analyzer. Processed organic layer samples were leached in the same way; however, contact times were one, two, four, six, eight, and 24 hours and a fixed concentration of 2 g litter/L water was used (Figures 3-4 and 3-5). The linearity of solubility behavior was tested to ensure appropriate scaling in which 0.1, 0.25, 0.5, 1.0, 2.5, 5.0, 10, and 15 g of each soil and 0.1, 0.25, 0.5, 1.0, 2.5 and 5.0 g of organic litter was leached separately under the same conditions as the kinetics test at a fixed contact time of six hours. Linearity plots for the mineral and organic layers are depicted in Figures 3-6, 3-7, and 3-8.


Figure 3-2. DOC Kinetics Plots for Soils, Subsamples A, B, D, and E. Soil/water contact times of 1, 2, 4, 6, 8, 10, 24, and 48 hours with a soil concentration of 0.5 g soil per 100 mL of

ultra-pure Milli-Q water were used for all temperatures.


Figure 3-3. TDN Kinetics Plots for Soils, Subsamples A, B, D, and E.

Soil/water contact times of 1, 2, 4, 6, 8, 10, 24, and 48 hours with a soil concentration of 0.5 g soil per 100 mL of ultra-pure Milli-Q water were used for all temperatures.


Figure 3-4. DOC Kinetics Plots for Litter, Sites NED (Left) and GROSS (Right).

Due to limited material, litter/water contact times of 1, 2, 4, 6, 10, and 24 with a litter concentration of 0.2 soil per 100 mL of ultra-pure Milli-Q water were used for temperatures up to 350°C.

Figure 3-5. DOC Kinetics Plots for Litter, Sites NED (Left) and GROSS (Right).

Due to limited material, litter/water contact times of 1, 2, 4, 6, 10, and 24 with a litter concentration of 0.2 soil per 100 mL of ultra-pure Milli-Q water were used for temperatures up to 350°C.


Figure 3-6. DOC Linearity Plots for Soils, Subsamples A, B, D, and E.

Soil masses of 0.1, 0.25, 0.5, 1.0, 2.5, 5.0, 10, and 15 g per 100 mL of ultra-pure Milli-Q water were used for all temperatures.


Figure 3-7. TDN Linearity Plots for Soils, Subsamples A, B, D, and E.

Soil masses of 0.1, 0.25, 0.5, 1.0, 2.5, 5.0, 10, and 15 g per 100 mL of ultra-pure Milli-Q water were used for all temperatures.

Figure 3-8. DOC (Left) and TDN (Right) Linearity for Litter from Sampling Site, NED.

Due to limited material, masses of 0.1, 0.25, 0.5, and 1.0 g per 100 mL of ultra-pure Millie-Q water were used for CTRL and 250°C only.

All leachates were passed through a 25-mm, 0.45-micron Whatman Puradisc Polyethersulfone syringe filter. Each filter was prewashed with 500 mL of Milli-Q water and field rinsed with 100 mL of leachate


solution in order to saturate adsorption sites on the filter and avoid potential loss of dissolved organic matter through filtration (Karanfil et al., 2003). Leachates were filtered directly into 20-mL glass vials for immediate analysis. The same leaching procedure was then scaled up, using a mineral soil mass of 5.0 g of soil or 0.2 g of homogenized organic detritus material leached in one L of ultra-pure Milli-Q water. These masses were chosen based on targeted DOC and TDN concentrations, which reflect commonly observed levels in surface waters. Because a larger volume of leachate was being filtered, large scale filtration was done through a 47-mm, 0.50-micron EMD Millipore Express PLUS Membrane disc filter using a pre-washed glass vacuum filtration apparatus. The discrepancy in filter pore size was accepted in order to keep the filtration material constant, as it was not possible to acquire this size filter in a 0.45-micron pore size. Leachates for these experiments were stored at 4˚C in 1-L amber Wheaton bottles. All containers used for any experiments and storage had been Liquinox-soaked, acid washed, and pre-furnaced to prevent contamination.

3.2.5 Analyses of Material and Leachates 3.2.5.1 Carbon and Nitrogen Content Carbon and nitrogen fractions in mineral soil was determined using a Thermo Scientific Flash EA1112 Nitrogen and Carbon Analyzer, which employs the Flash Dynamic Combustion method. These analyses were done in replicates of two. Total carbon percentage was assumed to be total organic carbon content. These data were reported as percent organic carbon (OC) or organic nitrogen (ON) remaining as a function of heating temperature using Equation 3-1. %𝑂𝐶 = 100 × , % , (% ) (3-1)

where: mgsolid,T = theoretical mass of the starting material at temperature, T,

%Mloss,T = percent mass loss due to heating at temperature, T, calculated using the difference between the original sample mass and the sample mass after burning,

%OCT (or %ONT) = carbon (or nitrogen) content at temperature, T,

This equation allowed for a successive tracking of OC or ON degradation as a function of heating temperature.

3.2.5.2 Dissolved Organic Carbon and Total Dissolved Nitrogen Both DOC and TDN concentrations in the resulting leachates were determined using a Shimadzu TOC-V CSN Total Organic Carbon Analyzer with a TN unit in replicates of four. DOC was determined as non-purgeable organic carbon (NPOC), detected as CO2 by a nondispersive infrared detector (NDIR). TraceMetalTM Grade hydrochloric acid was used as used as the purging agent, which was added to each sample manually.

DON was calculated by the subtraction of DIN from TDN. Nitrate and nitrite were measured together using a Lachat QuikChem 8500 Flow Injection Module where nitrate is quantified as reduced nitrite by passage through a copperized cadmium column. Nitrite (reduced nitrate plus pre-existing nitrite) was then quantified calorimetrically (Arikaree Environmental Laboratory, INSTAAR, CU Boulder). Nitrate was then back-calculated through subtraction of nitrite values. Ammonium was measured calorimetrically using a BioTek Synergy 2 Multi-Detection Microplate Reader (Arikaree Environmental Laboratory, INSTAAR, CU Boulder). Both analyses were done in replicates of four.


3.2.5.3 Fraction of Water Extractable Carbon and Nitrogen With both total carbon and nitrogen compositions in the soil and total dissolved concentrations of carbon and nitrogen characterized for each soil, the respective fractions of water extractable organic carbon and organic nitrogen (WEOC and WEON) SOM were calculated using Equation 3-2. 𝑊𝐸𝑂𝐶 𝑜𝑟 𝑊𝐸𝑂𝑁 = ( ) × × % ( % ) (3-2)

where: DOCT (or DONT) = leachate OC (or ON) concentrations resulting from the solid burned at temperature, T, (mg L-1)

Vleachate = leaching volume (L),

Msolid leached = mass of either soil or litter that was leached (mg),

%OCT (or %ONT) = carbon (or nitrogen) content at temperature, T,

3.2.6 Soluble Elements All soluble element concentrations were determined using Inductively Coupled Plasma Mass Spectrometry (ICP-MS) using a Thermo Finnigan Element2 magnetic sector inductively-coupled plasma mass spectrometer by Fredrick Luiszer at the Laboratory for Environmental and Geological Sciences (LEGS), CU Boulder. Mineral and organic layer leachate samples were prepared in replicates of 4 for soil and 2 for litter and acidified (pH < 1.5) using TraceMetalTM Grade nitric acid before analysis. This analysis was conducted for mineral layer samples from NED and FLG and for organic layer samples from all sites.

3.2.7 Statistical Analysis Statistical analysis was conducted using the open source programming language, R. All values are reported as the arithmetic mean of four replicates plus/minus one standard deviation, unless otherwise indicated. For calculated parameters, standard deviation is represented as a propagated error. Statistically significant differences between heating temperature and sites were assessed using a 2-tailed, 2-sample t-test with a Satterthwaite approximation of degrees of freedom. Each parameter was compared for each temperature to the control (e.g., CTRL-150, CTRL-250, CTRL-350, CTRL-450, CTRL-550). The linearity of desorption behavior for increasing soil leaching concentrations were established by linear regression.

Because mineral layer samples were taken in triplicate at each site, values for each parameter were averaged to represent a global value for its respective site. Standard deviations for each parameter at a given site represents the spatial heterogeneity of the soil over small distances. All trends identified in this study from site-averaged soil data are consistent among soil sub samples.

3.3 Results and Discussion 3.3.1 Carbon and Nitrogen Mineralization Carbon and nitrogen cycling in forested environments depends on several interplaying factors; among them, fuel type, soil moisture, productivity rates, and decomposition (Harden et al., 2002). Within the previous decades, wildfire has been incorporated as a major player in C and N sequestration and cycling, namely as a process causing the mineralization and subsequent emissions of NOx and CO2 into the atmosphere. Understanding the relationship between heating temperature and OM mineralization provides insight into how much material is lost after a burn event. Our results suggest that the magnitude of atmospheric emissions will be a function of burn severity with almost complete


mineralization at and above 450°C for mineral and organic layer carbon and at 550°C for mineral and organic layer nitrogen. At the 350°C threshold for NED, approximately 80% of mineral and organic layer carbon and 40% of mineral and organic layer nitrogen was lost as a result of heating, likely in the forms of CO2 and NOx, respectively.

The results in Figure 3-9 present the loss of total carbon (top panel) and nitrogen (bottom panel) in the organic layer (red) and mineral layer (blue) as a function of heating temperature for the NED site. As expected, both carbon and nitrogen for the mineral and organic layers showed a decreasing trend with increasing burning temperature. This was true across all sites and soil sub-samples. Compared to carbon, more nitrogen was remaining in both heated mineral soil and organic litter for temperatures at and above 250°C. As shown in Figure 3-9, the soil organic carbon remaining at 350°C was just above 20%, while the remaining soil organic nitrogen was 60%. A similar trend was observed in the organic layer, in which organic carbon remained at 25% and organic nitrogen remained at 50% after heating.

Figure 3-9. Carbon (Top Panel) and Nitrogen (Bottom Panel) Remaining in the Mineral and Organic Layers after

Heating at Various Temperatures Organized in Ascending Order.

NA

CTRL 150 250 350 450 5500%

20%

40%

60%

80%

100%

Nitr

ogen

(%)

CTRL 150 250 350 450 550

Mineral Layer N Organic Layer N

NACTRL 150 250 350 450 550

0%

20%

40%

60%

80%

100%

Car

bon

(%)

CTRL 150 250 350 450 550

Organic Layer C Mineral Layer C

Heating Temperature (°C)


3.3.2 Solubility of Organic Carbon and Nitrogen We evaluated the solubility of carbon and nitrogen from the studied mineral and organic layers. Figure 3-10 presents the results for the mobilization experiments. This figure shows the mobilization of carbon and nitrogen as a function of temperature, reported as a fraction of both C and N mobilized from the soil.

Figure 3-10. Fraction of Water Extractable Organic Carbon (WEOC) for Organic and Mineral Layer Samples (Left), and Fraction of Water Extractable Organic Nitrogen (WEON) for Organic and Mineral Layer Samples (Right).

The same amount of dried material yielded higher DOC from organic layer material than from mineral soil because the organic layer has a higher abundance of detritus. DOC originating from the organic layer also had consistent trends between all three sites, showing relatively high DOC concentrations for litter heated up to 150˚C and a subsequent, sharp, and step-wise decrease with increasing burn temperatures to levels below 1.0 mgC L-1 at 450 ˚C. At the 250˚C threshold, there was a DOC decrease between 7 - 10 times less than that of the CTRL and 150˚C. In general, organic carbon originating from litter contributes successively less DOC with increasing maximum burn temperature while soil contributes comparable DOC concentrations up to a moderate burn temperature (350˚C) and subsequently decreases at higher temperatures.

A very different trend was demonstrated by the dissolution of organic carbon from mineral layer soil. DOC release increased three-fold between CTRL and 250˚C (0.62 to 3.3 mgC L-1 for NED) followed by a progressive decrease with increasing temperature. TDN profiles for the organic and mineral layers looked very similar to the DOC profiles, with a sharp decrease after 150˚C for organic layer material and a clear peak for mineral soils at 250˚C.

Mineral layer leachates from 250˚C yielded 4.5 – 7.5 times more TDN than leachates from CTRL while litter leachates from 250˚C yielded 4.0 – 4.5 times less TDN than leachates from CTRL. Unlike DOC, there was no marked discrepancy between mineral soil and organic litter for TDN concentrations throughout the entire temperature range as a whole; both materials yielded TDN concentrations below 1.0 mgN L-1. For both materials, the majority of TDN was comprised of organic nitrogen, which could indicate that what is coming off of both soil and litter material could potentially be precursor material for nitrogenous DBPs. No other inorganic nitrogen species (𝑁𝑂 , 𝑁𝑂 , 𝑁𝐻 ) showed statistically significant change from the CTRL leachate for either material, except for ammonium for soil samples. Ammonium increased 4 – 7 times for leachates of soil heated to 250˚C from NED, FLG, and GROSS relatively, but still

NA

1.3%

2.9%

9.5%

6.6%

4.6%4.7%6.4% 5.8%

1.2% 1.1%1.8%

CTRL 150 250 350 450 5500%

3%

6%

9%

12%

15%

WEO

N (%

)CTRL 150 250 350 450 550

Mineral Layer Organic Layer

NA

0.99%

2.9%

6.5%

7.5%

4.2%

0.69%

5.1%5.7%

0.76%0.62% 0.32%

CTRL 150 250 350 450 5500%

3%

6%

9%

12%

15%

WEO

C (%

)

CTRL 150 250 350 450 550

Mineral Layer Organic Layer

Heating Temperature (°C)


made up a small portion of TDN. Soil and litter 𝑁𝑂 /𝑁𝑂 release was almost negligible. In contrast, Rhoades et al. (2011) reported post-fire stream 𝑁𝑂 /𝑁𝑂 concentrations that exceeded the baseline by one order of magnitude for nearly a decade after an ‘extreme extent’ portion of the Hayman fire (Rhoades et al., 2011). Possible explanations included a higher N availability from lowered N demand because of reduced vegetation from the fire and an accelerated N cycling via nitrification stimulated by heat and changes in pH. Our study indicated that heat impacted material from the mineral and organic layers exhibited no statistically significant increase in nitrate/nitrite leaching as a function of heating temperature. It is possible that this process is stimulated only at temperatures higher than the experimental temperature range in our study or that elevated stream nitrate/nitrite concentrations were a result of in-stream or terrestrial biological processes that are not reflected in our study. The solubility of soil organic nitrogen peaked at a lower temperature than did soil organic carbon (at 250˚C for nitrogen instead of 350˚C for carbon). In addition, solubility fractions were slightly higher, most notably after 250˚C where solubility leveled off instead of significantly dropping. This could be an artifact of the higher thermal recalcitrance in nitrogen compared to carbon (Figure 3-10). For organic layer material, the trend of nitrogen solubility as a function of heating temperature also mirrored that of litter carbon with slightly higher values. Again, this reflects the fact that there is more nitrogen available to be solubilized, as organic layer nitrogen was more thermally recalcitrant than mineral layer carbon (Figure 3-9).

The nature of SOM is inherently different to that of plant material (Knicker et al., 1996). SOM, a complex mixture of organic bio molecules of several morphologies and stages of decomposition, is bound to a mineral surface (Baldock and Skjemstad, 2000). This continuum can be complexed by their interactions with mineral surfaces and typically exist within matrices that can further stabilize otherwise labile moieties of organic matter. This interaction has been reported to act as a stabilizing barrier against biological oxidation (Baldock and Skjemstad, 2000); in a similar way, we suspect that the same protective barrier can occur against heat. At low temperatures, we hypothesize that the dominating mechanism of carbon release is a physical, heat-induced cleaving of the mineral/organic matter bonds. At moderate temperatures, in addition to a physical mechanism of release, we suspect that the SOM undergoes a thermally induced chemical transformation as well that renders the remaining organic matter more water soluble in water. The simultaneous occurrence of different processes during heating, among them the creation of polar moieties through incomplete combustion mechanisms and a general decrease in molecular weight (Cawley et al., 2017), may be a major factor in the increased solubility of heat-impacted organic matter. Other studies have demonstrated this same trend in SOM (Cawley et al., 2017; Santos et al., 2016). Albalasmeh et al. (2013) suggested that the heating process may weaken soil aggregates either as soil pore water is vaporized or as adhesive organic matter is combusted (DeBano et al., 1998), allowing a greater release of SOM from soil pores that were formerly inaccessible. This study suggested a predominantly physical mechanism of SOM desorption.

This hypothesis is supported by the solubility trends exhibited by soil organic carbon and nitrogen (Figure 3-10). The water-solubility trend of organic carbon as a function of heating temperature from soil mimicked that of DOC; however, the peak of solubility occurred at a slightly higher temperature of 350˚C instead of 250˚C. At 350˚C, less DOC is released (Figure 3-10) and 80% of the carbon is lost or mineralized (Figure 3-9). Therefore, even though there is less carbon available to be solubilized, the solubility not only remains steady but increases significantly. This is consistent with trends of DOC dissolution that, at 350˚C, a behavior-changing chemical transformation is occurring that is making the carbon remaining more water-soluble than carbon that is unburned or heated at lower temperatures. This explanation is especially coherent given that there is even less carbon remaining at these temperatures. For organic layer material, the trend paralleled that of mineral soil DOC showing a successively decreasing carbon solubility with increasing heating temperature. In contrast, the organic


layer is comprised of carbonaceous bio macromolecules only and lacks the protective mineral surface barrier, which makes this material more susceptible to mineralization through heat. This is supported by the results depicted in Figure 3-10 where the fraction of water-soluble carbon and nitrogen plummets to around and below 1% after 150˚C.

Soil-derived DOM encompasses a heterogeneous mixture of several organic materials including higher molecular weight polysaccharides and proteins, and sugars. As plant matter is a large component of SOM, it is conceivable that the lignin fraction of SOM contributes to much of the aromaticity of the DOM fraction. During the combustion processes at temperatures above 300°C, the lignin structures may break down the carbon-based “arms,” freeing the aromatic centers. Moreover, high temperatures have been reported to induce the formation of fused-ring structures (Knicker, 2007). This is also consistent with Santos et al. (2016), whose results demonstrated a greater abundance of conjugated C-structure through liquid-state 1H NMR spectroscopy (Santos et al., 2016). As temperature increases and these now-free arms volatilize, the carbon content will decrease, but aromaticity may still be preserved as aromatic structure are much less susceptible to breakage than alkyl outcroppings. This simultaneous decrease of carbon content with the preservation of aromaticity may help to explain the observed increase of SUVA254 values at and above 250°C. The spike at 350°C indicates a significant change in DOC character from soil extracts, which corroborates with the observed heightened solubility observed at the same temperature.

3.3.3 Soluble Elements Visible trends of soluble element release as a function of heating temperature were identified for most element species between materials. For Mg, Fe, Na, K, and Si, the scale of peak release for organic layer material was one order of magnitude greater for the release from mineral soil. Conversely, the opposite was true for Al, where mineral soil release of this element dominated. Scales were comparable for Ca and Mn between soil and litter. All trends are presented in Figure 3-11.


Figure 3-11. Water Soluble Elements for Mineral Layer (Soil, Left Panel) and Organic Layer (Litter, Right Panel)

Material for NED. Values for mineral soil are represented by the average of each site for each respective temperature (n=3) and the

average of their standard deviations. Litter values represent the average of two analytical duplicates and two sample duplicates (n=4) with one standard deviation.


Mineral soil exhibited a much higher (10x) Na release than organic layer derived litter, with the highest concentrations occurring at higher temperatures; nearly 350 µg g-1Dry Material at 450°C and 550°C. Litter released a maximum of 45.6 µg g-1Dry Material at 350°C. There was a 30-fold release of Ca from litter at 450°C compared to the maximum release of soil, which also occurred at 450°C. Ca became mobile in litter starting at 350°C (2676 µg g-1Dry Material) whereas Ca in soil became mobile starting at 250°C and sustained a relatively constant release for subsequent temperatures (961 – 1236 µg g-1Dry Material). Almost no Fe was mobilized from soil, except for at 250°C (5.1 µg g-1Dry Material). Litter-derived Fe peaked at 350 and 450°C (51.6 and 69.1 µg g-1Dry Material). Fe release from lower temperatures, including CTRL, was stable (between 12.3 and 16.3 µg g-1Dry Material). Litter released nearly 20x more K than did soil, with a peak at 5109 µg g-1Dry Material at 350°C. Concentrations from soil remained relatively stable from CTRL to 550°C, ranging between 193 and 271 µg g-1Dry Material. Litter had similar trend, which remained constant at a notably higher concentration compared to soil across all temperatures. Data from 450°C was unusable; therefore, it was excluded. For FLG peak K concentrations occurred at 450°C and was 4383 µg g-1Dry

Material. Litter released almost 14x more Mg at its peak (2419 µg g-1Dry Material at 450°C) than that of soil (175 µg g-1Dry Material at 250°C). Mg release from soil was relatively constant until a sharp increase at 250°C and then a slower, step-wise decrease from 250 to 550°C. Litter remained constant at CTRL, 150°C, and 250°C (ranging in concentration from 328 to 398 µg g-1Dry Material) and experienced a notable increase for subsequent temperatures. There was no large difference between Mn dissolution from soil compared to that of litter. Mn concentration spiked to about 50 µg g-1Dry Material at 250°C but was otherwise negligible for all other temperatures. Litter showed higher Mn concentrations for

CTRL and 150°C (20.7 – 29.6 µg g-1Dry Material) relative to 250°C and 350°C (12.5 – 15.4 µg g-1Dry Material). Na dissolution from litter was a full order of magnitude higher than soil for all temperatures, except for 250°C and 350°C, where concentrations were 5 and 6.5 times higher. Trends for both types of material were similar with relatively stable concentrations until a spike at 550°C for soil and 450°C for litter. Si concentrations up to 250°C for both littler and soil were comparable ranging from 22.6 – 39.5 µg g-1Dry

Material for soil and 41.2 – 37.4 µg g-1Dry Material for litter. Si dissolution from litter far exceeded that of soil, showing a 32- and 51-fold concentration compared to soils heated to 350°C and 450°C.

Table 3-1. Summary of Soluble Element Trends and Peak Concentrations from Soil and Litter. Data for NED was not usable; peak for FLG and GROSS occurred at 450˚C and was 4382.7 ± 549.5 μg 𝑔 and 4212.5 ± 833.5 μg 𝑔 .

Soil Litter

Element

Temperature of peak release (˚C)

Peak Concentration (μg 𝒈𝑫𝒓𝒚 𝑴𝒂𝒕𝒆𝒓𝒊𝒂𝒍𝟏 )

General Trend

Temperature of peak release (˚C)

Peak Concentration (μg 𝒈𝑫𝒓𝒚 𝑴𝒂𝒕𝒆𝒓𝒊𝒂𝒍𝟏 )

General Trend

Al 550 231.8 ± 4.9 ↑ C with ↑ T 350 45.6 ± 1.7 ↑ C with ↑ T

Ca 450 1236.1 ± 475.7 ↑ C with ↑ T 450 3686.4 ± 43.6 ↑ C with ↑ T

Fe 250 5.1 ± 1.9 Peak at 250 450 69.1 ± 0.7 ↑ C with ↑ T

K - - Constant 450 *- ↑ C with ↑ T

Mg 250 175.4 ± 58.6 Peak at 350 450 2418.7 ± 69.6 ↑ C with ↑ T

Mn 250 49.5 ± 14.8 Peak at 250 150 29.6 ± 0.8 Peak at 150

Na 550 24.9 ± 4.7 ↑ C with ↑ T 450 252.3 ± 31.8 ↑ C with ↑ T

Si 550 81.5 ± 32.7 ↑ C with ↑ T 450 3532 ± 81.4 ↑ C with ↑ T * K exhibited no discernable trend and remained relatively constant across the entire temperature range


Table 3-1 summarizes peak concentrations of each soluble element for organic layer (litter) and mineral layer (soil) material and at which temperatures the peak occurs. For litter, most elements (Al, Ca, Fe, K, Mg, Na, and Si) showed an increasing trend with increasing temperatures with peaks at the highest temperature reached, 450˚C. Release of Mn from litter peaked at a relatively low temperature of 150˚C. With the exception of Al, Fe, and Mn, the dissolution of the remaining elements yielded concentrations one order or magnitude (x10) higher than that of soil. In general, the elements investigated in this study, will originate predominantly from litter across the experimental temperature range and especially at temperatures at and above 450˚C. Most element species from soil had a similar trend, increasing release with increasing temperatures, except for Fe, Mg, and Mn, which peaked at 250˚C, 350˚C, and 250˚C respectively, and K, which exhibited no discernable trend and remained relatively constant across the entire temperature range (indicated by a * in the table).

3.4 Conclusions Findings of this study demonstrate that dissolution behavior across burn severities differ between the mineral and organic layers of soil. The most critical temperature for mineral layer-derived organic carbon and nitrogen release magnitude and solubility profile was 250°C and 350°C, showing a peak of DOC at 250°C and a peak of solubility at 350°C. For organic layer material, peaks occurred for CTRL and 150°C and dropped sharply for all subsequent temperatures. Additionally, organic layer material is the chief contributor of most constituents across all temperatures, with the exception of two of the investigated soluble elements and DON, suggesting that charred and combusted plant material poses a greater threat to water quality impairment following a wildfire with regard to the magnitude of DOC and soluble element export. This detail, together with the peak of solubility at 350°C may suggest a critical molecular transformation mechanism that renders wildfire-impacted soil dissolved organic matter a-typical in structure and therefore chemical behavior that warrants further investigation.

Organic matter originating from the mineral and organic layers of soil are not decoupled in natural systems following a wildfire; so, the contributions of their respective dissolved constituents will be combined. One would expect different dissolution behavior between them because SOM and organic matter derived from ‘fresh’ plant material are fundamentally different in structure; therefore, mechanisms driving their solubilities are not the same. Additionally, their fate, capacity to be metabolized, and the reactivity of the dissolved constituents from either soil or plant matter impacted by wildfire will likely be different. Notwithstanding, trends across parameters for material from mineral and organic layers showed very consistent respective patterns, making it potentially feasible to estimate water quality impacts under different burning conditions if maximum temperatures can be elucidated and further, if maximum temperature reached can be used as a surrogate parameter for wildfire severity. The implications of this work to land managers can inform decision making regarding how prescribed fires are administered. For example, it may not be advisable to control fuels by igniting moderate-severity wildfires. These practices could facilitate an unintentional release of altered organic matter and elements after a weather event, especially if there is sufficient proximity to water ways. If controllable, a higher severity fire would be advisable. This research revealed that land and water managers can expect contributions of dissolved constituents from both mineral and organic layer material following a wildfire with contributions from each differing in magnitude depending on the severity and thermal regime of the wildfire event.


CHAPTER 4

Activity 3: Source Water Quality Thresholds and Exceedance Evaluation 4.1 Introduction 4.1.1 Activity 3 Motivation Temporal and spatial variability of surface water quality poses challenges to drinking water treatment decision makers tasked with providing safe, reliable water. Variability in source water quality can result in utilities being in regulatory noncompliance for their finished depending on the water quality constituent of concern, the concentration range of the constituent, and the robustness of the water treatment plant to handle the changes in concentration. Utilities lack a method for understanding concentration ranges that are within their capacity to treat effectively. Climate change, droughts, and extreme events on the hydrologic cycle further exacerbate the problem of fluctuating source water quality. Utilities require an understanding of the future water quality of their source water given climatic influences on surface water quality and climate change.

4.1.2 Importance of Characterizing Thresholds for Source Water Quality Source water quality thresholds indicate the maximum allowable concentration of a water quality constituent that a water utility is capable of treating. Defining a threshold allows utilities to anticipate the potential of being out of regulatory compliance based on their source water quality. Several utilities have experienced peaks in source water concentrations, but lack an understanding of how those peaks will propagate through the treatment system and the likelihood of those peak concentrations resulting in regulatory violation. Characterizing a threshold, paired with an analysis of the probability of that threshold being exceeded enables water utilities to plan in advance when their utility is likely to become noncompliant with their regulations.

4.1.3 Motivation for Modelling Surface Water Quality Using Climate and Land Cover Predictors Surface water quality modelling using only climate and land cover predictors allows utilities insight on critical meteorological processes in their region influencing their source-water water quality. Utilities experience changes to source water quality due to weather events, such as precipitation, but do not have a means to assess what climatic or land cover factors most influence their water quality. This modelling approach provides opportunities for informed decision making by identifying key relationships between climate, land cover, and water quality. Further, climate change is expected to increasingly exacerbate challenges with variable water quality on utility operations. This modelling approach enables analyses of future water quality given uncertainties in future climate, allowing for risk-based planning that considers climate change. Lastly, this methodology relies on predictor data sets that are widely available, making it an applicable approach for water quality modelling in data-poor regions that lack catchment scale data.

4.1.4 Objectives of Activity 3: Source Water Quality Thresholds and Exceedance Evaluation The goal of Activity 3 was to enhance the ability of water utilities to assess and model variability in their surface-water source-water quality, allowing for decision making that ensures safe, reliable drinking


water and utility compliance with regulations. To achieve this goal, thresholds for key source water quality constituents were investigated, statistical tools for modelling threshold exceedances employed, and source water quality modelling frameworks using climate and land cover predictors were developed.

Activity 3 focuses on critical water quality constituents: total organic carbon (TOC), bromide, and turbidity. TOC reacts with a common disinfectant, chlorine, to form carcinogenic disinfection byproducts (DBP). TOC is controlled for by water utilities as mandated by DBP regulations. Bromide is a concern for utilities as it significantly increases the formation of DBPs. Turbidity, controlled for disinfection regulations, is used as a proxy measurement in drinking water science to indicate the potential presence of pathogenic microorganisms. Water utilities require an understanding of their current and future source water turbidity, TOC, and bromide concentrations to inform decisions that protect public health.

The objectives of Activity 3 are as follows:

1. Determine source water thresholds for TOC, bromide, and turbidity based on regulatory constraints in the finished water and use stream water quality data with extreme value theory to predict water quality threshold exceedances.

2. Develop models that relate surface water quality concentrations to historic climate (precipitation, temperature, drought index) and land surface variables (vegetation index).

4.2 Source Water TOC and Bromide Thresholds 4.2.1 Overview The EPA Stage 1 and Stage 2 Disinfectants and Disinfection Byproducts Rules (D/DBPRs) aimed to reduce disinfection byproduct (DBP) occurrence in public drinking water. DBPs are formed in the treatment process when disinfectants, most commonly chlorine, react with dissolved organic matter (DOM) and bromide, present in the water. Many DBPs are of health-concern, and two class sums, total trihalomethanes (TTHM) and five haloacetic acids (HAA5), are currently regulated with maximum contaminant levels (MCLs) of 80 μg/L and 60 μg/L, respectively (EPA, 1998). DBP formation continues in the distribution system (DS) as utilities are required to maintain a detectable chlorine residual throughout and the MCLs are enforced at the maximum level in the DS on a running annual average basis. Utilities must also remove a required percent TOC based on source water TOC and alkalinity concentrations (EPA, 1998). TTHM and HAA5 formation are functions of DOM concentration and character, bromide concentration, pH, temperature, contact time, and chlorine dose and residual. Higher source water DOM concentration, typically measured as total or dissolved organic carbon (TOC or DOC), results in increased DBP formation. Organic DBP precursor removal is the most effective method to control DBPs (Clark et al., 1994).

A conventional surface water treatment plant (CSWTP) with coagulation-flocculation-sedimentation-filtration processes only provides partial TOC removal, typically less than 50%, based on the DOM character and pH (Edwards, 1997). Thus, a utility using free chlorine for disinfection can only meet DBP regulations if its source water has a TOC concentration below a site-specific threshold. Additionally, increased source water bromide concentrations increase TTHM formation and impact DBP speciation; increased levels of bromide shifts DBP formation to higher ratios of brominated species (Cowman and Singer, 1996; Heller-Grossman et al., 1993; Krasner et al., 1989; Liang and Singer, 2003; McTigue et al., 2014; Summers et al., 1993). Brominated DBPs are more of a health concern than their fully chlorinated counterparts (Cantor et al., 2010; Richardson et al., 2007). Bromide is not removed by conventional SW treatment, thus meeting DBP regulations may be limited by the source water bromide concentration. In this study, source water TOC and bromide thresholds were investigated for three utilities.


Systems that use conventional treatment and were in compliance with Stage 1 D/DBPR may be experiencing greater challenges complying with the Stage 2 D/DBPR and may be considering advanced treatment options. Additionally, climate change threatens to further challenge water utilities in meeting DBP regulations as air temperatures rise, and increases in DOM have been observed in surface waters during periods of increased temperatures (Delpla et al., 2009; Freeman et al., 2001; Worrall et al., 2003). The risk of increasing source water temperature and DOM due to climate change, and increasing bromide levels from anthropogenic activities, may force water utilities to consider advanced treatment options, based on their current ability to meet DBP regulations.

The objective of this study was the development of a systematic methodology that will allow utilities to set site-specific source water thresholds for TOC and bromide based on DBP regulatory limits. The methodology is applied to source water TOC concentration thresholds using source water quality, treatment and distribution data for three case study locations, and to source water bromide concentration thresholds for one case study location.

The approach taken was to calculate the source water TOC or bromide thresholds as the maximum source water TOC or bromide concentration for which a CSWTP, using the minimum coagulant dose to meet the TOC removal requirement, can meet the DBP MCLs at the end of the DS.

4.2.2 Methods 4.2.2.1 Case Studies Data from three WTPs are utilized in this study, the Richard Miller WTP in Cincinnati, Ohio, the Harwood’s Mill WTP in Newport News, Virginia, and the Betasso WTP in Boulder, Colorado, representing different geographical locations and source waters. Source waters include the Ohio River for the Miller WTP, and the Chickahominy River and five reservoirs for the Harwood’s Mill WTP. Although the Betasso Plant receives source water from both the Lakewood Reservoir and Barker Reservoir, this study used only data for Betasso’s primary source, the Lakewood Reservoir. Table 4-1 summarizes the dataset for each WTP.

The Miller WTP utilizes granular activated carbon (GAC) as an advanced treatment after the conventional treatment process. In this study, settled water TOC, not yet treated with GAC, is used to represent finished water from the conventional treatment process. The Harwood’s Mill WTP uses chloramines for secondary disinfection, an advanced treatment method to reduce DBP formation. Source water TOC thresholds developed in this study are for conventional treatment; treatment with GAC or chloramines was not modeled.


Table 4-1. Data Summary for the Miller WTP, Harwood’s Mill WTP and the Betasso WTP.

Miller WTP Harwood’s Mill WTP Betasso WTP

Observed Data Range

Observation Date Range

Observed Data Range


Observed Data Range


Source Water

Total Alkalinity (mg/L) 28 – 348 1/1/1988 –

4/4/2007 20 – 72 1/2/2001 – 3/12/2012 6.9 – 25.3 1/10/1995 –

4/8/2013

TOC (mg/L) 1.23 – 4.66 1/1/1988 – 4/4/2007

3.38 – 10.70

1/2/2001 – 3/12/2012 1.16 – 8.53 1/10/1995 –

4/19/2013

Turbidity (NTU) 1.0 – 938 1/1/1988 – 4/6/2007 0.27 – 16.6 1/2/2001 –

3/12/2012 0.35 – 8.60 1/10/1995 –4/8/2013

UV254 (1/cm) 0.01 – 0.84 1/4/2000 – 3/6/2007 0.06 – 0.39 1/2/2001 –

3/12/2012 0.02 – 0.29 3/14/1995 –4/8/2013

Water Temperature

(°C) 0.6 – 31.1† 10/4/1992 –

3/21/2007 3.0 – 32.2 1/2/2001 – 3/12/2012 2.75 – 19.6 1/10/1995 –

4/8/2013

pH - - - - 6.16 – 9.09 1/10/1995 –4/8/2013

Bromide 0.00 – 0.30 1/3/1995 – 1/2/2016

0.00 – 0.530

4/4/2002 – 6/22/2015 - -

Treat-ment

Process

Alum dose (mg/L)

5.80 – 39.33

2/1/1996 – 4/6/2007 32-101 1/2/2001 –

3/12/2012 3.3 – 53.6†† 1/1/2003 –4/30/2013

Settled Water TOC (mg/L) 1.09 – 4.04 5/11/1988 –

4/4/2007 - - - -

Finish. Water

Finished Water TOC (mg/L) 0.23 – 2.01 10/15/1992 –

4/4/2007 1.76 – 4.35 1/2/2001 – 3/12/2012 0.42 – 3.26 1/10/1995 –

4/19/2013 Finished Water

pH - - - - 6.8 – 9.05 1/10/1995 –4/8/2013

Dist-ribution System

TTHM (𝛍g/L) - - - - 0.5 – 122.3†††

3/21/2005 – 4/2/2013

HAA5 (𝛍g/L) - - - - 9.1 – 166††† 3/21/2005 – 4/2/2013

† The Miller WTP provided water temperature data for the finished, effluent water, which is used as source water temperature with the assumption that treatment did not significantly alter the water temperature. †† The Betasso WTP provided daily coagulant doses of alum and polyaluminum chloride (PACl). An alum equivalence for the PACl was determined and added to the alum dose to give the results presented in this table. †††TTHM and HAA5 concentration data are for 28 locations in the Betasso WTP DS.

4.2.2.2 Seasonal Variability Source water temperature and TOC concentration are key factors in predicting a WTP’s ability to meet DBP regulation. Higher water temperatures increase the kinetics of the reaction between chlorine and OM and therefore increase DBP formation. Monthly variability of water temperature and TOC concentrations data from the Miller WTP from January 1988 to April 2007 and from the Harwood’s Mill WTP from January 1996 to April 2007 are shown in Figure 4-1, parts A and B, respectively and Figure 4-2, parts A and B, respectively. This data for the Betasso WTP from January 1995 to April 2012 is shown in Table 4-1. For all boxplots the bottom and top of the box represent the 25th and 75th percentile values, respectively, and the horizontal lines below and above the box, also known as the “whiskers,” represent the 5th and 95th percentile values, respectively. Open circles represent values less than the 5th percentile and above the 95th percentile, respectively. The horizontal line inside the box represents the median. Utilities must understand the seasonal variation of source water temperature and TOC and resulting DBP formation to adjust their treatment process to meet DBP regulations. This also suggests that for a given water utility, the source water TOC or bromide threshold, for below which a water utility could meet DBP regulations, varies monthly.


Figure 4-1. Monthly Source Water Temperature for the Miller and Harwood’s Mill WTPs.


Figure 4-2. Monthly Source Water TOC Concentrations for the Miller and Harwood’s Mill WTPs.

4.2.2.3 The EPA WTP Model This study utilizes the EPA WTP Model version 2.0 (Center for Drinking Water Optimization, 2001), which predicts the behavior of water quality factors that impact DBP formation (Solarik et al., 2000). The model provides a tool for water utilities to predict DBP formation based on their current treatment process and to estimate changes in DBP formation with different treatment options by altering the treatment train in the model. The most important aspects of the EPA WTP Model for this study are the post coagulation TOC concentration and the DBP formation throughout the DS. The DOM coagulation model (Edwards, 1997) is utilized in the EPA WTP Model for predicting dissolved organic carbon (DOC) concentration remaining after coagulation. In this study TTHM and HAA5 formation was predicted using the WTP model where chlorine was added to water treated by enhanced coagulation, coagulation to meet the required TOC removal. The DBP formation equations use bromide concentration, pH, water temperature, chlorine dose and reaction time, as well as the combined DOC and UVA (DOC*UVA) parameter, which takes into account the characteristics of OM, including reactivity, and the impact of treatment on OM removal. TTHM formation in treated waters is calculated as:


TTHM = 23.9(DOC*UVA)0.403(Cl2)0.225(Br)0.141(1.027)(T-20)(1.156)(pH-7.5)(t)0.264 and HAA5 formation in treated waters is calculated as:

HAA5 = 41.6(DOC*UVA)0.328(Cl2)0.585(Br-)-0.121(0.9216)(pH-7.5)(1.022)(T-20)(t)0.150

where DOC, UVA, Cl2, Br, T and t are equal to the dissolved organic carbon (mg/L), UV absorbance at 254 nm (1/cm), the applied chlorine dose (mg/L), the bromide concentration (𝜇g/L), the temperature (°C) and reaction time (hr), respectively. The data ranges for algorithm development were 1.00 ≤ DOC ≤ 7.77, 0.016 ≤ UVA ≤ 0.215, 1.11 ≤ Cl2 ≤ 24.75, 23 ≤ Br ≤ 308, 15 ≤ T ≤ 25, 6.5 ≤ pH ≤ 8.5, and 2 ≤ t ≤ 168 (Center for Drinking Water Optimization, 2001; Solarik et al., 2000).

4.2.2.4 EPA WTP Model Verification The WTP Model’s accuracy in modeling the treated water quality of the three case studies is paramount in developing monthly TOC and bromide thresholds. Data from the three case studies were used to test the model’s accuracy in estimating TOC removal by alum coagulation on a monthly basis. For each month, the monthly median source water temperature, alkalinity, turbidity and UV254 were used as model inputs, as well as a source water pH of 7.7 and representative maximum DS residences; 10 days in the winter to 7 days in the summer for the Miller WTP, 15 days in the winter to 10 days in the summer for the Harwood’s Mill WTP and 9 days in the winter to 6 days in the summer for the Betasso WTP. An internal validation method was used to verify how well each monthly median source water TOC concentration, settled (for the Miller WTP) or finished water TOC concentration and alum dose represented the actual values. This was done by randomly selecting five years within the range of each dataset, removing data from those five years, calculating the median of each reduced dataset, comparing the median values of the datasets missing five years of data and the median values of the full datasets, and determining the percent error between the two corresponding medians for each month. This process was repeated three times for each dataset; the average percent error between each pair of corresponding monthly medians was below 5%, suggesting that the monthly median values used were representative.

The modeled post-coagulation TOC concentration for each month was compared to the monthly median settled water TOC concentration, shown in Figure 4-3. For the Miller WTP (Figure 4-3, part A), the monthly medians of the alum dose data were multiplied by a scale factor of 2.3 to calibrate the model, because the Miller WTP adds coagulant in the pre-sedimentation basin, providing extra holding time, hours to days, after the dose is added. The WTP model does not have a direct way to model this, but the calibration factor simulates the increased efficiency of the alum dose in removing TOC. The model well predicted the Harwood’s Mill WTP (Figure 4-3, part B) finished water with the historic alum dose data. The Betasso Plant uses both alum and polyaluminum chloride (PACl) as coagulants and the model alum coagulant doses were adjusted a conversion factor of 1.13 based on PACl aluminum contribution (Samson, 2016). The model with this adjustment systematically underestimates the TOC removal, and therefore the model’s adjusted alum dose was calibrated, similar to the Miller WTP case study, using a calibration factor of 2.0. With the calibration, the TOC removal is modeled accurately, except in the month of June when the model overestimates TOC removal; the results are shown in Figure 4-3, part C. The average difference between the median of the settled water TOC concentration data and the modeled settled water TOC concentration for each month is 0.10 mg/L, 0.28 mg/L and 0.09 mg/L for the Miller, Harwood’s mill and Betasso WTPs, respectively.

The Betasso Plant’s data were also used to verify the model’s ability to predict DBP formation in the DS. Based on DBP data from 28 locations within the DS, the 7 locations with quarterly DBP concentration data spanning at least 6 years were selected to represent the end of the DS. For each of the 7 locations,


the median TTHM and HAA5 concentrations for the months of March, June, September and December were determined, and the range of these quarterly medians were compared with the model’s predicted TTHM and HAA5 concentrations using the minimum alum dose to meet the required TOC removal. The average difference between the predicted TTHM and HAA5 concentration and each location’s median quarterly TTHM and HAA5 concentration was 8.0 𝜇g/L and 7.5 𝜇g/L, respectively. In general, the model under-predicted DBP formation at the end of the DS, and in June, when the TOC removal was overestimated, the modeled TTHM and HAA5 concentrations were approximately equal to, and lower than, the median finished water TTHM and HAA5 concentrations, respectively.

Figure 4-3. Conventional Treatment Modeling Validation.

4.2.3 Results 4.2.3.1 TOC Thresholds Monthly TOC thresholds were developed based on the use of conventional treatment with alum for enhanced coagulation. Monthly medians for source water TOC, water temperature, alkalinity, turbidity and UV254 data were used in the WTP model. Source water bromide concentrations was set at 0.05 mg/L. Monthly TOC removal requirements were determined using monthly median source water TOC and alkalinity. They were 35% in December-May and 25% in June-November for the Miller WTP, 45% in all months for the Harwood’s Mill WTP, and 35% in April-July and 0% in August-March, because the median source water TOC concentration was below 2.0 mg/L, for the Betasso WTP. In August-March months, the monthly median alum dose from the Betasso WTP’s data was used in the model to ensure appropriate turbidity removal. In all other cases, the alum dose was adjusted to find the minimum dose


capable of removing the required percent TOC. For the Miller WTP in the months of December-May, the alum dose was determined based on a post-coagulation TOC concentration of 1.9 mg/L, as the required percent removal results in TOC values lower than 1.9 mg/L. The source water TOC concentration in the model was then adjusted to find the maximum TOC concentration, or source water TOC threshold, in which TTHM and HAA5 formation did not exceed their respective MCLs at the end of the DS. For the Miller and Harwood’s Mill WTP in all 12 month, the thresholds were TTHM-controlled, where TTHM concentrations reached 80 μg/L before HAA5 reached 60 μg/L at the end of the DS as the source water TOC concentration was increased, demonstrating that the model shows higher TTHM formation than HAA5 formation at a pH of approximately 8, consistent with past results (Liang and Singer, 2003). For the Betasso WTP, where the modeled finished water pH was adjusted based on their monthly median finished water pH, ranging from 7.5-7.7, the thresholds in February-May and in December were HAA5-controlled.

The chlorine dose was adjusted such that the chlorine residual at the end of the DS was equal to 0.3 mg/L, the size of the contact basin in the treatment train immediately after the chlorine dose was adjusted such that the Giardia contact time (CT) ratio was equal to 1.2, and the caustic soda dose, added after the contact basin, was adjusted such that the effluent pH was equal to 8.0. For the Miller WTP, the monthly source water SUVA values calculated using monthly median source water TOC and UV254 range from only 3.82 – 3.85 Lmg-1m-1 and a constant value of 3.84 Lmg-1m-1 was used. For the Harwood’s Mill and Betasso WTPs, the UV254 was adjusted such that the SUVA value for each month equaled the value calculated with the monthly median source water TOC concentration and UV254, ranging from 2.51-3.08 Lmg-1m-1 for the Harwood’s Mill WTP and from 2.60-3.82 Lmg-1m-1 for the Betasso WTP. Samson (2016) provides details of the approach and model inputs for each month for each case study. Monthly source water TOC thresholds are compared with source water TOC data for the Miller WTP and the Harwood’s Mill WTP in Figure 4-4, parts A and B, respectively, and in Table 4-2 for the Betasso WTP to provide insight to each WTP’s capacity to meet Stage 2 DBP regulations using conventional surface water treatment.

The Miller WTP’s monthly median source water TOC concentration exceeds the monthly source water TOC threshold in all 12 months, as shown in Figure 4-4: part A. Exceedance rates for all source water TOC concentrations in January – April ranged between approximately 52-65%, between 97-100% in May – November, and are approximately 81% in December. Meeting DBP regulations in the summer and early fall is particularly difficult because source water TOC concentrations are at their highest and TOC thresholds are at their lowest, primarily due to high water temperatures. The results suggest that the Miller WTP requires advanced treatment to meet Stage 2 DBP regulation, and in fact the Miller Plant uses GAC, which significantly reduces the TOC concentration before the point of chlorination.

All of the Harwood Mill’s WTP’s historic source water TOC concentrations exceeded the monthly source water TOC thresholds; monthly median concentrations were all significantly higher than the thresholds (Figure 4-4, part B), suggesting the Harwood’s Mill WTP also requires alternative treatment to meet Stage 2 DBP regulations, which they do in the form of chloramines for secondary disinfection.

The Betasso Plant utilizes conventional treatment, as modeled in this study. As shown in Table 4-2, no source water TOC thresholds exceedances have been observed in August – March. Less than 25% of the historic source TOC concentrations in the months of April and July have exceeded their monthly thresholds, however, approximately 57% and 55% of the data exceeded the thresholds in May and June, respectively. During these spring months, source water TOC concentrations peak due to snowmelt and spring runoff conditions.


Figure 4-4. Comparison of Monthly Source Water TOC Concentrations and TOC Thresholds.

Table 4-2. Monthly Source Water TOC Thresholds for the Betasso WTP.

Month Source Water Alkalinity (mg/L)

Source Water Temperature (°C)

Source Water TOC (mg/L)

Source Water TOC Threshold (mg/L)

January 13.3 3.6 1.6 4.1

February 15.4 4.4 1.6 4.2

March 17.7 5.3 1.8 4.0

April 17.8 6.4 2.9 3.0

May 15.5 8.6 4.6 3.2

June 12.9 11.6 3.4 3.1

July 10.8 15.7 2.2 2.2

August 11.4 17.1 1.7 3.1

September 12.5 15.1 1.8 3.2

October 14.2 11.1 1.7 3.9

November 14.7 5.9 1.7 4.0

December 14.7 4.2 1.6 3.9

Source Water TOC Thresholds with Monthly Bromide Concentrations A second set of source water TOC thresholds were developed for the Miller and Harwood’s Mill WTPs using monthly median source water bromide concentrations and the same methodology for the development of source water TOC thresholds with constant bromide concentrations. For the Miller WTP in February – April, and for the Harwood’s Mill WTP in February, these TOC thresholds were HAA5-controlled, and the remaining months were TTHM-controlled. Source water TOC thresholds with constant bromide concentration of 0.05 mg/L and with monthly median bromide concentrations are shown for the Miller WTP in Figure 4-4, part A, and in Figure 4-4, part B, for the Harwood’s Mill WTP. The two sets of thresholds are very similar for both utilities. For the Miller WTP, the greatest differences of 0.2 mg/L occur in April, May and July, when the TOC thresholds with monthly bromide concentrations


are higher, and in November, when the TOC threshold with a constant bromide concentration of 0.05 mg/L is higher. For the Harwood’s Mill WTP, the median source water bromide concentration is below 0.05 mg/L for all 12 months, and therefore, the TOC thresholds with monthly bromide concentrations are consistently higher by 0.2-0.3 mg/L than the thresholds developed with a constant bromide concentration of 0.05 mg/L.

4.2.3.2 Bromide Thresholds The Miller WTP case study was used to develop source water bromide thresholds for three scenarios: (1) monthly source water bromide thresholds when both source water and finished water TOC concentrations are held constant, (2) monthly source water bromide thresholds when source water TOC concentrations are set equal to the monthly median source water TOC concentration, and (3) monthly source water bromide thresholds when the post-coagulation TOC concentration in the model was set equal to the monthly median finished water TOC concentration. Bromide thresholds were not developed for the Harwood’s Mill WTP because the high TOC concentrations exceed the TOC thresholds in all cases; bromide concentrations below the model boundary conditions would still result in exceedances. Bromide thresholds were not developed for the Betasso WTP, because no source water bromide data were available. The Betasso WTP source water bromide concentrations are estimated to be below the lower bromide concentration boundary condition for the DBP formation equations, < 0.02 mg/L, after analyzing DS trihalomethane (THM) concentration data, which was predominately chloroform, with very low to undetectable concentrations of the brominated THM species.

Source Water Bromide Thresholds with Constant TOC Concentrations To develop monthly source water bromide thresholds with constant TOC concentrations, the source water TOC concentration in the WTP model was set to the average of the 12 monthly source water TOC concentration medians, 2.6 mg/L. The alum dose was adjusted in the WTP model each month to yield a post-coagulation TOC concentration of 1.9 mg/L to meet the TOC removal requirement of < 2.0 mg/L. The chlorine dose, the contact basin size and the caustic soda dose were adjusted for each month as described above for the development of source water TOC concentrations. The source water bromide concentration in the model for each month was adjusted until the maximum bromide concentration, or source water bromide threshold, was determined in which the TTHM concentration at the end of the DS did not exceed its MCL of 80 μg/L. Due to the boundary conditions of the DBP formation equations used in the WTP, bromide thresholds below 0.02 mg/L could not be determined accurately and are identified as < 0.02 mg/L. The results are shown in Figure 4-5. Monthly medians of source water bromide concentration data were below the corresponding monthly bromide threshold in three months, January-March, and above the thresholds in the remaining nine months. In October and November, 100% of the bromide concentration data was above the threshold. These results suggest that if the Miller WTP used conventional treatment, it would likely experience TTHM formation at the end of the DS in exceedance of the MCL approximately 75% of year or more, when the finished water TOC was 1.9 mg/L.

In the months of April and December, when the source water bromide concentration in the model was set equal to the bromide threshold, the HAA5 concentration at the end of the DS exceeded its MCL of 60 μg/L. In these months, increases in bromide concentration above the threshold increased TTHM concentrations, but actually lowered HAA5 concentrations. This is because increased bromide concentrations shift the formation of haloacetic acid (HAA) species to a higher proportion of brominated species (Cowman and Singer, 1996). Out of the five HAA species included in HAA5, only two are brominated, while the remaining four species that are not included in HAA5 are all brominated.


Figure 4-5. Comparison of Monthly Source Water Concentrations and Thresholds for Bromide and TOC Based on

the Miller WTP Data.

Source Water Bromide Thresholds with Monthly TOC Concentrations Source water bromide thresholds with monthly median source water TOC concentrations for meeting the TTHM MCL at the end of the DS were also developed. The minimum alum dose for meeting the required TOC removal, used for determining the source water TOC thresholds, were used. Resulting monthly source water bromide thresholds are shown in Figure 4-5, with the source water bromide thresholds with constant TOC concentration. The difference in the two sets of monthly source water bromide thresholds is within 0.001-0.006 mg/L. Similar to the bromide thresholds developed with constant TOC concentrations, the source water bromide thresholds with monthly median TOC concentrations are < 0.02 mg/L for the months of May – November, and in April and December, the HAA5 concentration exceeds 60 μg/L at the end of the DS when bromide concentrations are equal to the monthly threshold.

In reality, the Miller WTP utilizes GAC to significantly reduce the TOC concentration at the point of chlorination in the treatment process. A third set of monthly source water bromide thresholds were developed by setting the post-coagulation TOC concentration in the model equal to the monthly median of the Miller WTP’s finished water TOC concentration data. For all 12 months, the bromide threshold exceeded the upper bromide concentration boundary condition for the DBP formation equations of 0.300 mg/L, which is equal to the maximum source water bromide concentration provided in the Miller WTP data, thus the threshold would never be exceeded.

4.2.4 Modeling Using Extreme Value Theory Extreme value theory provides a framework for analyzing and modeling extreme values (Coles, 2001). Two forms are proposed in this theory. In the first, maxima of a block (e.g., seasonal or annual maximum rainfall, streamflow, turbidity etc.) is modeled using a Generalized Extreme Value (GEV) probability


distribution function. This has the drawback of excluding data that are closer to the block maxima - this is alleviated in the Peaks Over Threshold (POT) approach, the second form within the extreme value theory. In the POT approach data exceeding a threshold are modeled as a generalized Pareto distribution (GPD) – which is well suited to model water quality variables most of which have thresholds for regulatory compliance. In both the block maxima and the POT approach, the distribution parameters can be modeled as function of set of covariates, which could include physical variables, time trends etc., to capture temporal non-stationary – see Katz et al. (2002) for detailed exposition on the nonstationary modeling framework using GEV and GPD distributions. The EVT is finding its utility in water and wastewater quality modeling in recent years – that we have helped to pioneer. For example, in Towler et al. (2010) nonstationary GEV was used to model influent turbidity at a water treatment plant and its projections under climate change. Nonstationary GPD was used in Haagenson et al. (2013) to model water demand exceeding desired thresholds and in Samson (2016) for threshold exceedances of TOC and bromide, Suchetana et al. (2019) developed nonstationary GPD models for modeling wastewater quality constituents. We refer the reader to these references for details. A sample result from Samson’s research, which was partially supported by this grant, is shown below in Figure 4-6. This figure shows the source water TOC concentration observations in April – July from the Betasso plant (described in previous sections) over time with the stationary and non-stationary GPD model return levels of 2-year, 20-year, and 100-year return period. The nonstationary GPD model included land surface and climate variables as covariates. The stationary GPD model return periods are constant, while the return periods for the non-stationary model fluctuate with the observed TOC concentration data, indicating the ability to capture temporal variability of the extremes in TOC.

Figure 4-6. Observed TOC Concentrations with 2-year, 20-year, and 100-year Return Periods for the Stationary

and Non-Stationary GPD Models with a TOC Threshold of 2.2 mg/L.

Samson (2016) also modeled threshold exceedances of bromide with very good results - indicating the broader applicability of extreme value theory to water quality modeling.


4.2.5 Conclusions This study develops a methodology that establishes a site-specific TOC and bromide source water threshold above which the TTHM or HAA5 MCL would be exceeded. It utilizes the EPA WTP model and the site-specific source water, treatment plant and DS data. Utilities could use this model to assess the need for advanced treatment options based on current DBP regulations and potential future changes. The model could be used to predict treatment outcomes based on changing conditions, such as increasing source water temperatures or TOC concentrations as a result of climate change, and increasing source water bromide loading from anthropogenic activities.

Understanding relationship between projected climate scenarios and source water quality may be critical for utilities to make informed decisions on treatment process. Climate change threatens to alter water quality conditions through changing temperatures and precipitation levels and frequencies, in addition to other possible indirect environmental influences. Environmental conditions, such as increased temperatures, could affect source water TOC concentrations, and in turn, affect the formation of DBPs. Other environmental conditions, such as soil moisture and vegetation, may affect TOC concentrations through their influence on OM production and transport to source waters; these are additional factors that could be affected by climate change. Coupling the methodology presented in this paper for developing source water TOC thresholds with a model that predicts TOC concentrations based on future climate projections (Samson et al., 2016) would provide water utilities with a comprehensive understand to make informed treatment decisions for reducing DBP occurrence to meet regulations and protect the health of its consumers.

Extreme Value analysis offers robust options to modeling extremes and threshold exceedances of water quality variables that are crucial in developing robust strategies for regulatory compliance. Furthermore, these methods can enable good understanding of how potential future climate change can impact source water quality and source water threshold exceedances is integral for water utilities in future decision-making for treatment processes and source water selection for drinking water regulatory compliance and for the protection of public health.

4.3 Source Water Turbidity Thresholds 4.3.1 Overview The EPA utilizes filtration, a treatment technique, as part of the approach to controlling microbial contaminants. Utilities using conventional or direct filtration treatment are required to produce a combined filter effluent (CFE) of less than 0.3 NTU in at least 95% of measurements taken per month, with a maximum allowable level of 1 NTU. Utilities with slow sand and diatomaceous earth filtration systems are regulated to meet a CFE of less than 1 NTU in at least 95% of measurements per month, never exceeding a level of 5 NTU. For plants employing alternative filtration technologies, states can set the CFE limit, permitting that the CFE never exceed 5 NTU and 95% of monthly samples are less than 0.5 NTU. Generally, most surface water treatment plants use conventional treatment (coagulation/flocculation/sedimentation/filtration) and many strive for the Partnership for Safe Water Goal of a CFE less than 0.1 NTU (Partnership for Safe Water, 2014).

Figure 4-7 illustrates the process train for conventional surface water treatment. To facilitate that utilities meet CFE regulations, most utilities target a settled water turbidity Partnership for Safe Water Goal of 1 to 2 NTU . This target also enables utilities to extend the run time of their filters by reducing filter loadings. Variable source water turbidity, termed raw turbidity herein, can be problematic in meeting this target. Utilities only have a few alternatives, such as changing the coagulant dose (a primary operating cost), to meet the target settled water turbidity. Extreme events limit the ability of


utilities to meet the target by affecting residence times and sludge removal capacity.

Figure 4-7. Conventional Surface Water Treatment Process with Turbidity Checkpoints.

There is a need for utilities to understand how source water turbidity levels propagate throughout treatment processes, especially considering anticipated future increases in source water quality variability from climate change. Understanding to what extent signals of variable turbidity in source waters result in variability in settled water and CFE turbidities is necessary to inform operational strategies that ensure regulatory compliance. To this end, there were two objectives of this study. First was to develop probabilities assessing the likelihood of the CFE and settled water targets being exceeded given the turbidity as a prior process point. Secondly, this study aimed to characterize how spikes in source water turbidity propagated through the treatment process, resulting in spikes to the settled and CFE turbidities.

4.3.2 Data Collection Surface water based drinking water utilities provided turbidity data for three locations within the treatment plant: (1) the source water (i.e., raw, untreated, surface water entering the treatment plant); (2) water after the sedimentation process, referred to as settled water; and (3) combined filter effluent (CFE). Quality control checks were made on the data. Data was checked to ensure all three locations had turbidity values for the same time stamp. Data was also inspected for negative turbidity values and missing dates/times of data. For utilities that provided settled turbidity from multiple sedimentation basins, the average settled turbidity was used. Table 4-3 summarizes the data collected from partnering water utilities used in this activity. Two locations have two treatment plant datasets associated with them: Boulder City, NV and Sacramento, CA. For Boulder City, NV, both treatment plants, the RMWTF and AMSWTF, were direct filtration plants (no sedimentation) and use the same source water. For Sacramento, CA, the two utilities use differing source waters; the SRWTP is fed by the Sacramento River whereas the EAFWTP uses the American River. Throughout this report, results are referred to by the location where the data was obtained. For Boulder City, NV and Sacramento, CA, results are given per treatment facility name.


Table 4-3. Data Collected from Partner Drinking Water Utilities. All plants are conventional surface water treatment plants, except for the Boulder City, NV plants which are

direct filtration plants. Note that data was collected Fort Collins, CO, Austin, TX, Albuquerque, NM, and Denver, CO, but analysis could not be performed due to data limitations.

Location Treatment Facility Name Water Source

Turbidity Ranges (NTU) Time Step

Number of Data Points Source Settled CFE

Boulder City, NV

Alfred Merritt Smith Water Treatment Facility (AMSWTF)

Lake Mead 0.00 - 8.37

Direct Filtration

0.01 - 0.58

Daily

3596

River Mountains Water Treatment Facility (RMWTF)

0.01 - 0.10 3596

Houston, TX East Water Purification Plant #3 Trinity River 4.00 -

87.00 1.02 - 8.60 0.02 - 0.49 Daily 2192

Knoxville, TN Mark B. Whitaker Water Treatment Plant

Tennessee River

0.00 - 100.00

0.00 - 10.00

0.00 - 1.00 15 minutes 318481

Sacramento, CA

EA Fairbairn Water Treatment Plant (EAFWTP)

American River

0.59 - 22.90 0.15 - 2.58 0.03 -

0.26 Daily 2442

Sacramento River Water Treatment Plant (SRWTP)

Sacramento River

1.65 - 240.00 0.18 - 9.04 0.03 -

0.20 Daily 2470

Westminster, CO

Semper Treatment Plant Standley Lake 0.57 -

59.70 0.35 - 2.39 0.02 - 0.05 Daily 1827

4.3.3 Probability of Target Exceedances 4.3.3.1 Method Turbidity targets were related to prior process point turbidities to determine the probability of a prior process point resulting in a turbidity above the target on a per utility basis. The Partnership for Safe Water CFE target 0.1 NTU was related to raw water turbidity and settled water turbidity levels, respectfully. The settled water turbidity targets of 1 and 2 NTUs were related to raw water levels.

The probability of the CFE target being exceeded was assessed in terms of the raw water turbidity level using: 𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑜𝑓 𝐶𝐹𝐸 > 0.1 𝑁𝑇𝑈 = 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑜𝑖𝑛𝑡𝑠 𝑤𝑖𝑡ℎ 𝐶𝐹𝐸 > 0.1 𝑁𝑇𝑈 𝑓𝑜𝑟 𝑎 𝑟𝑎𝑤 𝑡𝑢𝑟𝑏𝑖𝑑𝑖𝑡𝑦 < 𝑋 𝑁𝑇𝑈𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑑𝑎𝑡𝑎 𝑝𝑜𝑖𝑛𝑡𝑠 𝑓𝑜𝑟 𝑟𝑎𝑤 𝑡𝑢𝑟𝑏𝑖𝑑𝑖𝑡𝑦 < 𝑋 𝑁𝑇𝑈

where the raw turbidity level, X NTU, was determined on a per utility basis. Likewise, the probability of an exceeding the CFE target from a settled water turbidity level of Y NTU was calculated using:


𝑃𝑟𝑜𝑏𝑎𝑏𝑖𝑙𝑖𝑡𝑦 𝑜𝑓 𝐶𝐹𝐸 > 0.1 𝑁𝑇𝑈 = 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑝𝑜𝑖𝑛𝑡𝑠 𝑤𝑖𝑡ℎ 𝐶𝐹𝐸 > 0.1 𝑁𝑇𝑈 𝑓𝑜𝑟 𝑎 𝑠𝑒𝑡𝑡𝑙𝑒𝑑 𝑡𝑢𝑟𝑏𝑖𝑑𝑖𝑡𝑦 < 𝑌 𝑁𝑇𝑈𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑑𝑎𝑡𝑎 𝑝𝑜𝑖𝑛𝑡𝑠 𝑓𝑜𝑟 𝑠𝑒𝑡𝑡𝑙𝑒𝑑 𝑡𝑢𝑟𝑏𝑖𝑑𝑖𝑡𝑦 < 𝑌 𝑁𝑇𝑈

This equation form was also used to determine the probabilities of a settled turbidity greater than the 1 NTU and 2 NTU targets using the raw water turbidity.

Two treatment plants provided data with enough filter effluent points above 0.1 NTU to investigate occurrences of CFE values exceeding 0.1 NTU using source water turbidity values and settled water turbidity values. These were the Houston and Knoxville utilities. All utilities were assessed for probabilities of the settled turbidity levels exceeding the settling targets based on raw turbidity levels.

All probabilities for targets being exceeded were made using at least 2.4% of the total dataset, with an exception of probabilities constructed for the CFE given a settled water turbidity level of less than 2 NTU. 0.3% of the Houston dataset and 0.1% of the Knoxville dataset had data within the settled less than 2 NTU and CFE greater than 0.1 NTU grouping. Although the probabilities of exceeding the target are built using a small fraction of these datasets, these results were kept as means to analyze the utility of the settling target of 2 NTU.

4.3.3.2 Results A depiction of the targets analyzed using prior process point turbidity levels is shown in Figure 4-8 for the Houston water utility. Plots of settled versus raw, filter effluent versus settled, and filter effluent versus raw turbidities were used as guides to choose groupings of prior process turbidity levels for analysis of subsequent target exceedances.


Figure 4-8. Plots of Turbidity Data at the Houston, TX Water Facility.

Red dashed lines are the target turbidities. For the settled turbidity versus raw turbidity plot, the thinner red dashed line is the settling target turbidity of 2 NTU, whereas the thicker red dashed line is the target of 1 NTU. The

blue line shown on each plot is a smoother function showing the overall trend of the dependent variable in relation to the independent variable.

Probabilities of target exceedances for all plants with sufficient data above the target levels were calculated and reported in Table 4-4. Probability for exceeding the CFE target using raw water turbidity levels for the Knoxville facility were not reported because there were too few data points in raw turbidity groupings. Similarly, the Sacramento plants and the Westminster utility had too few data points above the settling target of 2 NTU to determine probabilities of target exceedance using raw water turbidity levels. Houston, on the other hand, only has a probability for the settling target of 2 NTU due to all of the settling turbidity data points being greater than 1 NTU, resulting in a 100% probability of 1 NTU being exceeded for all raw turbidity values. Knoxville had enough settling data points clustered above the 2 NTU limit for one raw water turbidity level to warrant the calculation of an exceedance probability.


Table 4-4. Probability of Target Turbidities Given the Turbidity at a Prior Process Point Less than a Specified Value.

Probability of Settled Turbidity Greater than 1 NTU Based on Source Water Turbidity

EAFWTP, Sacramento, CA Raw (NTU) < 3 5 10

Probability (%) 2.8 5.1 7.5

SRWTP, Sacramento, CA Raw (NTU) < 10 20 50

Probability (%) 3.5 6.3 9.6

Westminster, CO Raw (NTU)< 3.5 7 10.5 14

Probability (%) 8.8 12.2 16.0 18.5

Knoxville, TN Raw (NTU) < 6 9 12 15 22

Probability (%) 25.2 27.3 30.7 32.8 34.9

Probability of Settled Turbidity Greater than 2 NTU Based on Source Water Turbidity

Houston, TX Raw (NTU) < 11 12 15 25 40

Probability (%) 32.3 47.2 59.9 65.9 72.5

Knoxville, TN Raw (NTU) < 13

Probability (%) 5.5

Probability of CFE Turbidity Greater than 0.1 NTU Based on Source Water Turbidity

Houston , TX Raw (NTU) < 20 30 40 60

Probability (%) 8.5 12.2 15.0 18.3

Probability of CFE Turbidity Greater than 0.1 NTU Based on Settled Water Turbidity

Houston , TX Settled (NTU) < 2 3 4 9

Probability (%) 1.4 10.1 16.5 18.4

Knoxville, TN Settled (NTU) < 2

Probability (%) 0.10

4.3.3.3 Conclusions Results from this activity show substantial increases in the probability of target level turbidities being exceeded given increases in prior process turbidity levels. For the Houston utility, for example, a raw turbidity level less than 11 NTU has a 32.3% chance of resulting in a settled water turbidity above the 2 NTU target. Increasing the raw turbidity range by one NTU increases the probability of exceeded the settled water target to 47.2%. Similarly, the CFE target of 0.1 NTU has an 8.5% chance of being exceeded when raw water turbidity levels are less than 20 NTU. When the range of raw water turbidity is increased to levels less than 30 NTU, the probability of the Houston facility exceeding the CFE target rises to 12.2%.


This research also supports the settled water turbidity target of 2 NTU. The probability of exceeding the CFE target from a settled turbidity level of less than 2 for Houston was only 1.4%. However, increasing the settled level turbidity range to less than 3 NTU resulted in a 7 fold increase in the target exceedance probability, resulting in a 10% change of the CFE target being exceeded. For Knoxville, our results indicate having a settled level of less than 2 NTU results in a very low probability of CFE target exceedance (i.e., a 0.1% chance).

These probabilities of target exceedances provide utilities with an opportunity to perform a risk based assessment of their utility operations. These probabilities can be utilized with results from Activity 3, Section 4.4, where raw turbidity values are modelled and predicted for surface water systems. Predicting raw water levels of turbidity, in conjunction with these established probabilities for ranges of raw water levels resulting in settled or CFE target exceedances, enables utilities to assess the future of their treatment facility.

4.3.4 Turbidity Spikes throughout the Treatment Train 4.3.4.1 Method Data for raw turbidity, settled turbidity, and filter effluent turbidity were normalized to allow for a comparison of relative changes from turbidity at prior process points to subsequent process points. Normalizing rescales the data from 0 to 1, alleviating issues of comparing variability in raw, settled, and effluent turbidities due to each turbidity process point having significantly different ranges in turbidity levels. Turbidity values were normalized for the raw, settled, and CFE turbidity datasets collected from utilities listed in Table 4-3 using: 𝑋 , = 𝑋 − 𝑋𝑋 − 𝑋

where 𝑋 , is each normalized turbidity data point, 𝑋 (NTU) is the original data point being normalized, 𝑋 (NTU) is the minimum turbidity level for the data set, and 𝑋 (NTU) is the maximum turbidity level for the data set. For each dataset, the average turbidity was also calculated using the normalization equation to determine the normalized average raw, settled, and CFE turbidity levels.

Normalizing turbidity allows for an analysis of relative changes in prior process turbidity levels to subsequent turbidity levels. For example, normalized raw turbidity values can be related to normalized raw settled turbidity to determine when above average raw turbidity levels were associated with above average, average, or below average settled water turbidity levels. This enables an analysis of how spikes in the raw water turbidity can result in spikes in the settled water turbidity. This method of determining propagating spiked turbidity levels throughout the treatment plant was applied to analyze relationships between raw and settled water, settled water and CFE, and raw and CFE turbidities. Relating relative changes from the raw turbidity to settled water turbidity provides information on how the conventional treatment processes of coagulation, flocculation, and sedimentation handle variability in raw turbidity levels. Similarly, relating settled water turbidity to CFE levels provides information on the removal of variability in settled water turbidity via filtration. Lastly, analyzing how changes in the raw turbidity relate to relative changes in the CFE provides information on how the treatment plant as whole handles variability in raw water levels. The following ratios were used to assess removal of turbidity from prior process points to subsequent process points: 𝑅𝑎𝑡𝑖𝑜 1 = 𝑆𝑒𝑡𝑡𝑙𝑒𝑑 𝑊𝑎𝑡𝑒𝑟 𝑇𝑢𝑟𝑏𝑖𝑑𝑖𝑡𝑦 𝐿𝑒𝑣𝑒𝑙 (𝑁𝑇𝑈)𝑅𝑎𝑤 𝑊𝑎𝑡𝑒𝑟 𝑇𝑢𝑟𝑏𝑖𝑑𝑖𝑡𝑦 𝐿𝑒𝑣𝑒𝑙 (𝑁𝑇𝑈)


𝑅𝑎𝑡𝑖𝑜 2 = 𝐹𝑖𝑙𝑡𝑒𝑟 𝐸𝑓𝑓𝑙𝑢𝑒𝑛𝑡 𝑇𝑢𝑟𝑏𝑖𝑑𝑖𝑡𝑦 𝐿𝑒𝑣𝑒𝑙 (𝑁𝑇𝑈)𝑆𝑒𝑡𝑡𝑙𝑒𝑑 𝑊𝑎𝑡𝑒𝑟 𝑇𝑢𝑟𝑏𝑖𝑑𝑖𝑡𝑦 𝐿𝑒𝑣𝑒𝑙 (𝑁𝑇𝑈)

𝑅𝑎𝑡𝑖𝑜 3 = 𝐹𝑖𝑙𝑡𝑒𝑟 𝐸𝑓𝑓𝑙𝑢𝑒𝑛𝑡 𝑊𝑎𝑡𝑒𝑟 𝑇𝑢𝑟𝑏𝑖𝑑𝑖𝑡𝑦 𝐿𝑒𝑣𝑒𝑙 (𝑁𝑇𝑈)𝑅𝑎𝑤 𝑊𝑎𝑡𝑒𝑟 𝑇𝑢𝑟𝑏𝑖𝑑𝑖𝑡𝑦 𝐿𝑒𝑣𝑒𝑙 (𝑁𝑇𝑈)

Ratio 1 captures the removal efficiency of coagulation, flocculation, and sedimentation, whereas Ratio 2 captures the removal efficiency of filtration and Ratio 3 captures the removal efficiency of conventional surface water treatment: coagulation, flocculation, sedimentation, and filtration.

A framework was developed in which normalized turbidity data for subsequent process points were plotted against normalized data from prior process points to analyze relationships of relative changes in turbidity levels from one process point to the next in the treatment facility. An illustration of this framework plot for normalized settled versus normalized turbidity is shown in Figure 4-9. Lines representing the normalized average settled and raw turbidity levels, as well as a 1:1 line, were overlaid onto the plot to create 6 zones. Each of these zones relates how data points in that zone compare to the average normalized values of the settled and raw turbidities, allowing relationships to emerge that explain how raw turbidity variability influences settled water turbidity. Settled water turbidity data points that are greater than the average settled water turbidity are data points above the blue line in Figure 4-9, i.e., data that can be in Zones 1, 2, or 3. Normalized raw water turbidity data points greater than the average raw turbidity are to the right of the red line, meaning they can be located in Zones 2, 3, 4 and 6. If a peak was observed in the normalized raw turbidity and propagated through the treatment as a relative peak in the normalized settled water turbidity, then the data for that event would be located in either Zones 2 or 3.

Zones 2 and 3, as well as Zones 4 and 6, are separated by the 1:1 line. This 1:1 line indicates how well a relative level of turbidity in the raw was treated to produce the relative level of turbidity in the settled water. For a peak in the raw turbidity that results in the same relative peak in the settled water, the data point would be exactly on the 1:1 line. For peaks in the raw turbidity that are dampened via treatment resulting in less of a relative peak in the settled water, the data point would be beneath the 1:1 line in Zone 3 or 4. However, if peaks in the raw turbidity resulted in a larger relative peak in the settled water turbidity, then the treatment removal efficiency had declined below the average removal rate and the data would be above the green line in Zone 2. That is, raw and settled water levels that had a ratio 1 greater than the average ratio 1 would be located below the green line, whereas raw and settled levels resulting in ratio 1 less than the average ratio 1 would be above the green line in Figure 4-9. Data in Zone 6 indicate that while the peak in the normalized settled water is higher than that in the raw water, that the settled water peak is still lower than the average normalized settled water turbidity.


Figure 4-9. Example Zoning of a Normalized Settled Turbidity versus Normalized Raw Turbidity Plot.

Dashed red lines are the normalized average raw turbidity, dotted blue lines the normalized average settled turbidity, and solid green lines the 1:1 line for the data. The figure on the right demonstrates an example when

Zone 6 occurred.

A summary of the zone definitions for example Figure 4-9 is provided in Table 4-5; the definitions in the table hold true for plots of normalized filter effluent versus normalized settled turbidity and normalized filter effluent versus normalized raw turbidity.

To further illustrate the concept of zoning, the example below provides zones for hypothetical data points. Table 4-6 presents the data and Figure 4-10 depicts each individual data point in its respective zone.

Table 4-5. Meaning of Zones Shown in Figure 4-9 in Terms of the Normalized Data in Each Zone Compared to the Normalized Average Value of the Data.

Zone Normalized settled turbidity compared to the normalized average settled turbidity

Normalized raw turbidity compared to the normalized average raw turbidity

Ratio 1 compared to the average Ratio 1

Zone 1 Greater than Less than -

Zone 2 Greater than Greater than Less than

Zone 3 Greater than Greater than Greater than

Zone 4 Less than Greater than Greater than

Zone 5 Less than Less than -

Zone 6 Less than Greater than Less than


Table 4-6. Summary Statistics, Data, and Normalized Data for an Example Illustrating Zones for Individual Data Points.

Raw Turbidity (NTU) Normalized Raw Turbidity Settled Turbidity (NTU) Normalized Settled

Turbidity Zone Data Point

10 0.02 2.00 0.66 Zone 1 1 100 0.44 3.00 1.00 Zone 2 2 92 0.40 1.80 0.59 Zone 2 3 220 1.00 0.12 0.01 Zone 4 4 5 0.00 0.10 0.00 Zone 5 5 60 0.26 1.00 0.31 Zone 5 6 20 0.07 1.20 0.38 Zone 5 7 10 0.02 1.20 0.38 Zone 5 8 30 0.12 2.00 0.66 Zone 1 9 100 0.44 1.00 0.31 Zone 4 10 80 0.35 1.30 0.41 Zone 6 11 180 0.81 2.00 0.66 Zone 3 12

Raw Turbidity Statistics Settled Turbidity Statistics

Average (NTU) 76 Average (NTU) 1.39

Max (NTU) 220 Max (NTU) 3.00

Minimum (NTU) 5.0 Minimum (NTU) 0.10

Normalized Average 0.33 Normalized Average 0.45

Figure 4-10. Example Data Points from Table 4-6 Plotted into Zones.

Data from each treatment facility listed in Table 4-3 was zoned to provide insight on how signals in turbidity from the raw water affects settled water turbidity and effluent turbidity levels, as well as how signals in the settled water relate to signals in the filter effluent turbidity.


4.3.4.2 Results Data from each utility listed in Table 4-3 were normalized and zones relating prior process turbidity points to subsequent process turbidity points were constructed. An example of the zones generated for the SRWTP, Sacramento, CA data is provided in Figure 4-11. The percentage of data points per zone and utility are reported in Table 4-7.

Figure 4-11. Figures Showing Zoning of the SRWTP, Sacramento, CA Turbidity Data.

Red lines are the normalized average independent variable, blue lines are the normalized average dependent variable, and green lines are the 1:1 lines.

Table 4-7. Percentage of Data Points per Each of the 6 Zones for Each Plot of Normalized Data.

Water Utility Settled versus Raw Effluent versus Settled 1 2 3 4 5 6 1 2 3 4 5 6

Houston, TX 16.0 6.3 21.0 9.9 46.8 0.0 14.6 4.2 23.1 16.0 42.2 0.0

Knoxville, TN 18.4 8.2 7.5 14.8 51.0 0.0 33.2 0.0 10.0 24.1 32.6 0.0

EAFWTP, Sacramento, CA 15.0 19.8 0.6 1.4 59.9 3.4 46.3 0.2 21.4 13.7 18.3 0.0

SRWTP, Sacramento, CA 20.0 8.1 7.7 3.7 59.9 0.6 18.0 11.9 3.6 18.0 46.2 2.4

Westminster, CO 20.4 23.2 0.2 0.4 42.9 12.9 27.6 22.0 2.2 13.5 28.6 6.1

Water Utility Effluent versus Raw 1 2 3 4 5 6

Houston, TX 20.2 1.5 20.2 15.6 42.6 0.0

Knoxville, TN 30.7 0.0 12.5 18.0 38.7 0.0

EAFWTP, Sacramento, CA 51.4 9.3 7.3 8.6 23.5 0.0

SRWTP, Sacramento, CA 24.5 5.5 3.5 7.8 55.4 3.3

Westminster, CO 35.9 15.8 0.1 1.0 27.5 19.8

AMSWTF, Boulder City, NV 19.7 0.7 20.6 13.7 45.3 0.0

RMWTF, Boulder City, NV 21.8 17.2 0.1 1.1 43.2 16.6


When comparing normalized settled water turbidity to that of the raw water, the majority of the data for all five plants fell into Zone 5 (average of 52%) followed by Zone 1 (18%). When comparing normalized effluent water turbidity to that of the settled water, the data fraction in Zone 5 was highest in four of the five cases (34%) followed by closely by data in Zone 1 (28%). When comparing normalized effluent water turbidity to raw water turbidity, the Zone 5 was highest in five of the seven cases (40%) gain followed by data in Zone 1 (29%). The Zone 5 data dominated in these cases and represent treated water that is lower than average treated water turbidity. However, data in Zone 1, with the next highest frequency, represents treated water turbidity that is higher than average treated water turbidity, for process influent data that is below average influent water turbidity, i.e., low turbidity influent water (relative to the average influent turbidity) yields high turbidity treated water relative average treated water turbidity in the process influent.

4.3.4.3 Conclusions Relationships between raw water turbidity levels to settled and filter effluent turbidities, and settled water turbidity levels to CFE, are uniquely described via the novel zoning methodology presented in this activity. Each zone provides insight to how signals of turbidity propagate throughout the treatment train.

Data that falls into Zone 1 is a result of operations at the facility. The data at a prior location was less than its average, but yet there is an above average level of turbidity at the subsequent location. Surprisingly, a large percentage of data was placed in Zone 1 for all relationships explored across all utilities.

Zone 2 and Zone 3 represent cases where higher turbidity levels at a prior location in the treatment facility resulted in higher turbidity levels at a subsequent location. These zones indicate that the treatment system between the prior and subsequent turbidity points is not effectively removing spikes from the prior turbidity point. Data that falls into Zones 2 and 3 is concerning when considering increasing variability in surface water turbidity due to future climate change. Especially concerning is data that falls into Zone 2, for the spikes in the prior turbidity point resulted in greater turbidity at the subsequent turbidity point and reduced the removal efficiency of the treatment method between those two points compared to its average removal efficiency. The time series of the raw and settled water turbidity data for the treatment plants with a high fraction of the data in Zone 2 and 3 validated this concern: it was evident that spikes in the raw turbidity were resulting in spikes to the settled water turbidity. Figure 4-12 illustrates this for the EAFWTP, Sacramento, CA utility.


Figure 4-12. Time Series of Raw and Settled Turbidity for the EAFWTP, Sacramento, CA Utility.

Gaps in the time series are from missing data points due to plant shutdown for routine maintenance.

Zone 4 represents above average turbidity levels at the prior process point being attenuated, resulting in below average turbidity levels at the subsequent point. Zone 4 has removal ratios that are greater than the average, meaning that a spike in the prior process point was well handled by the treatment system. Data that falls into zone 6, in contrast, represents cases where the prior process turbidity level was attenuated, but the removal efficiency of the treatment system between the points was reduced compared to its average removal efficiency.

Lastly, zone 5 indicates below average turbidity levels in both the subsequent and prior locations.

Results from this activity allow utilities to understand how turbidity levels in the raw are affecting turbidity levels in the settled and CFE at their utility, as well as how turbidity levels in the settled relate to CFE levels. These results characterize how spikes in turbidity propagate through the treatment process and how frequently they result in spikes at subsequent locations at the treatment facility.

4.4 Modelling Water Quality Using Climate and Land Cover Predictors 4.4.1 Overview Temporal variability of surface water quality can negatively affect the ability of drinking water utilities to produce safe, reliable drinking water and meet regulations. Variations in water quality can be seasonal or long-term as a result of climate and/or land use change. Processes employed in conventional water treatment are sensitive to source water quality, reducing treatment plant performance when utilities are not proactive in planning for changes to source water supplies. Historically, water treatment plants were built with additional capacity to enable flexibility in operations according to the influent water quality. This method of treatment design is problematic: oversized plants require excess energy, costs, and land. In urban areas where land areas are constrained and expensive, this can be an especially difficult method of managing variable influent water quality. Mixing source water supplies is a common method to handle this dilemma, where the utility will blend multiple water sources, yielding a consistent


water quality mixture to be treated. This approach, however, is impossible for utilities constrained to only one source water supply. There is a need to produce methods to assess and plan for variability in water quality. Modelling and predicting variability in source water quality in the short term and long term on a per-utility basis will facilitate cost effective, efficient decision making regarding treatment modifications to produce reliable, potable water, for both current day and the future.

Modeling surface water quality is traditionally done using catchment specific data. Methods to model and forecast TOC, for example, generally require streamflow data. Likewise, methods to model surface water turbidity levels most commonly use streamflow in power law equations with empirical coefficients (Mather, 2015). Streamflow data can be difficult to obtain, may not represent changing stream conditions due to human perturbations on natural streams (irrigation diversions, dam controls), or may lead to models that are not applicable under changing climate (Samson et al., 2016).

To address the issue of catchment specific data being required to model surface water quality, our group recently developed a statistical model of surface water TOC concentration using climate and land surface variables. The predictor data set consisted of only accessible data with coverage for the US, such as temperature, precipitation, normalized difference vegetation index (NDVI), and Palmer Drought Severity Index (PDSI). The motivation to use climate and land surface variables comes from previous studies. Increases in dissolved organic matter concentration (DOM) have been found to be consistent with increases in air temperatures (Delpla et al., 2009; Freeman et al., 2001; Worrall et al., 2003). Air temperature drives DOM dissolution and microbial activity, which are key factors in mobilizing TOC. Precipitation characteristics have also been shown to influence surface water TOC concentration. Heavy precipitation increases surface runoff, resulting in increased TOC (Delpla et al., 2009). Long dry spells have been shown to increase TOC (Hope et al., 1994) and short periods of heavy rainfall followed by longer dry spells have shown to increase DOM mobilization (Schindler et al., 1997). In the last case, the rainfall period leads to vegetation growth and the subsequent dry period leads to its demise and decay, resulting in organic matter that is ready to be mobilized in the following rainy period. Interplay between temperature and precipitation is also important. Köhler et al. (2008) observed increasing TOC concentration during warm summers in wet years, but not in dry years. This rich literature clearly indicates that surface water TOC concentrations are strongly linked to climate events (e.g., temperature, precipitation, snowmelt, etc.) and their interactions with land surface (e.g., vegetation, OM in soil, etc.) leading to seasonal concentration variations. In Samson et al. (2016), a local polynomial regression framework using climate and land surface predictors was applied for three utilities with differing surface water sources in differing geographic regions. This work demonstrated that TOC variability in surface water could be modelling used climate and land surface variables, negating the need for catchment-specific data collection.

Modelling surface water TOC using climate and land cover predictors is advantageous compared to traditional methods, aside from the applicability of the modelling approach in regions lacking ground-based observations. Because the predictor data set mainly consists of climate variables, it can be manipulated to represent projected future climate, allowing for analysis of climate change impacts on water quality. The objective of this study was to expand on these findings in several ways. New modelling methods that are easier to communicate and use by decision makers were explored, and new methods to assess predictive skill of the models were developed. A new water quality variable was also assessed alongside TOC: turbidity.


4.4.2 Methods 4.4.2.1 Data Collection Surface water quality data was collected by water utilities and aggregated to a monthly basis. Data sets where interpolated to fill gaps of missing monthly data points and omitted from the analysis if they consisted of fewer than 60 data points. For points recorded as below their detection limit, the value of the data point was set to half of the detection limit.

Average, monthly independent variables used in the modelling were normalized difference vegetation index (NDVI), Palmer Drought Severity Index (PDSI), maximum temperature, and minimum temperature. Summed, monthly predictors used were the number of dry days in a month and precipitation. All variables were lagged to represent climatic and land surface conditions one, two, and three months prior.

Climate and land surface data for was collected using three websites with open access to the data. The NOAA Climate Data Online Search tool for Global Historical Climatology Network data (NOAA, n.d.) was used to collect daily summaries of precipitation and temperature at a station gauge nearest to the intake of the water utility. Gridded NDVI data from the National Oceanic and Atmospheric Administration (NOAA) on a 16 km resolution was collected via the IRI Data Library (IRI 2019). The HUC 8 watershed of the surface water body delineated at the inlet to the water utility was used to find an average NDVI value per watershed and time. Lastly, NOAA climate division data (NOAA 2019) was used to find PDSI for the climate division where the source water was located.

4.4.2.2 Model Types Investigated Three regression models and two tree models were investigated per water utility to determine their performance with modeling source water quality using monthly climate and land surface data.

The regression models developed were a generalized linear regression (GLM) with gamma family, standard linear regression, and local polynomial regression with gamma family. The GLM is a regression framework wherein the response variable is assumed to be a realization from the exponential family, as opposed to standard linear regression where it is assumed to be from a normal distribution (McCullagh and Nelder, 1989). Depending on the distribution of the response variable, a function of the distribution parameters is modeled linearly with selected covariates. This framework offers flexibility in modeling a variety of response variables, such as variables that are positive, binary, categorical, threshold exceedances for extremes; it has been applied widely to environmental and water applications (Samson et al., 2016; Suchetana et al., 2019; Weirich et al., 2015, 2011). Local polynomial regression uses a fraction of nearby data points to the data point being modelled, in contrast to the standard linear regression and GLM which use the full set of data to perform the regression. Similarly to the GLM, various distributions of the response variable can be incorporated within the local polynomial regression. For both the GLM and local polynomial regressions, the gamma distribution was selected. The gamma family has the constraint that the modeled variable will be positive, which is not the case in the normal distribution. For this analysis, this constraint is reasonable because it is impossible to observe negative water quality concentrations. Linear models were assessed using model diagnostic tests to ensure model validity. Diagnostic tests checked linear model assumptions of normal, identically distributed, and independent residuals. Predictor data sets were chosen for the linear and GLM models using Bayesian information criterion (BIC), whereas generalized cross validation criterion (GCV) was used for the local polynomial model predictor selection. Significance of the chosen predictor sets were assessed using t tests for the linear and GLM models and F tests for the local polynomial model.

Tree models investigated were classification and regression trees (CART) and random forests. CART uses


binary partitioning in the data to repeatedly split the dependent variable data by a single independent variable. The result is a tree-like structure where the first split denotes the most influential predictor on determining the modeled variable, the successive splits denote branches, and finally the modeled variable is represented at the end nodes, otherwise known as leaves. The tree maps the pathways of dependency between the modeled variable and the suite of independent variables, allowing for an analysis on the dominant pathways responsible for explaining variance in the modeled variable. Trees are pruned by minimizing cross validation. Random forests (Breiman, 2001) build upon CART by growing several decorrelated trees, reducing the instability associated with the CART method. Each tree is constructed using a different bootstrap sample from the original data. At each node, a random subset of the variables is used to determine the splitting decision. Variable importance in random forests models is determined using the IncNodePurity, i.e., the total decrease in node impurity that results from splits over each individual variable averaged over all the trees. 500 trees were built for each random forest model. For a detailed and easy exposition of the tree-based methods see the book by Hastie et al. (2009). These methods are being employed widely in a variety of applications for their simplicity in interpretation and for decision-making. Some recent applications include - for efficient building systems (May-ostendorp et al., 2013), in construction safety management (Tixier et al., 2016), and for wastewater treatment plant performance modeling (Atanasova and Kompare, 2002; Suchetana et al., 2017).

4.4.2.3 Model Selection A procedure for determining the best model per location was developed using a balance between model fit and prediction power. The model selection goal was to choose a model that had the relatively best model fit, i.e., a high R squared with the predicted water quality variable versus the actual, and the best prediction power of all the modelling methods investigated. Prediction power was assessed using a drop 10% cross validation method, which entails dropping 10% of the observed data and then predicting that 10%. The modelled 10% and actual 10% of the dropped data were analyzed using R squared after performing the drop 10% cross validation 500 times to see how well the model predicted the dropped data. R squared from the 500 cross validations was plotted as a boxplot, with the R squared from the model fitting overlaid onto the boxplot. Models were determined to have good predictive skill that was consistent with how well the model fit to the observations when the boxplots of the R squared from the drop 10% cross validation had a relatively small spread and a median that aligned well with the R squared from fitting.

4.4.2.4 Case Study Descriptions This report will focus on models built using water quality data from two utilities: Houston, TX and Fort Collins, CO. These two utilities use surface waters that are unique in their geographic location and watershed characteristics.

The intake to the Fort Collins utility is the Cache la Poudre (CLP) River, located in northern Colorado. The CLP originates near the continental divide and is dominantly fed by snowmelt run off events. The CLP watershed, depicted in Figure 4-13, experiences a Köppen climate classified as cold-semi arid. The watershed is mainly mountainous and forested, with little perturbation by human settlements. The Houston utility, in contrast, is fed by the Trinity River. The Trinity River originates north of Dallas, TX and flows southeast toward Houston. The Trinity River watershed is shown in Figure 4-13 and has a Köppen climate of humid subtropical. Within the Trinity River watershed is the cities of Dallas and Fort Worth. Dallas and Fort Worth discharge treated wastewater into the Trinity River, contributing a large fraction of the flow of the Trinity. The Trinity flows south of Dallas and Fort Worth until it deposits into Lake Livingston, a main reservoir used by the Houston water utility; it is estimated that half of the annual


supply to the Houston utility is derived from treated wastewater effluent (National Research Council, 2012).

Figure 4-13. HUC 8 Watersheds for Fort Collins (left) and Houston (Right).

Source: USGS 2019.

Data Description for Fort Collins, CO Case Study TOC concentrations and turbidity levels in the CLP were collected for the Fort Collins case study. Two data sets were explored: one consisted of only turbidity data, whereas the second included turbidity and TOC data. The data sets are described in Table 4-8. Data set 1 included observations for each month of the year, whereas data set 2 only has observations from April to November. Data set 2 focuses on only April to November because this is a period of interest due to the spring snowmelt dominating the water quality in the CLP.

Table 4-8. Description of Data Sets Used to Model Water Quality in the CLP at Fort Collins, CO. Data Description Data Set 1 Data Set 2

Date Range 1/2007 – 12/2016 4/2008 – 11/2015

Number of Monthly Observations 120 64

Original Time Step of Data Data point every 15 minutes 1 to two data points per month

Turbidity Range 0.54 – 30.94 NTU 0.25 – 8.81 NTU

TOC Range - 1.64 – 9.38 mg/L

Data Description for Houston, TX Case Study Water quality data analyzed in the Trinity River case study for Houston, TX is presented in Table 4-9.

Table 4-9. Description of Water Quality Data Set for the Trinity River at Houston, TX. Data Description Data set

Date Range 11/2011 – 12/2016

Number of Monthly Observations 62

Original Time Step of Data Monthly Averages

Turbidity Range 9.3 – 52.3 NTU

TOC Range 4.6 – 9.8 mg/L


4.4.3 Results 4.4.3.1 Fort Collins Case Study Four models were built for the CLP River, two relying on data set 1 and two relying on data set 2. An overview of the inputs to the four models is provided in Table 4-10. Each data set and predictor set combination described in Table 4-10 was evaluated using the three regression modelling methods and two tree modelling methods. Each model presented in this section was the superior model in terms of balancing model fit and predictive performance per predictor set and data set.

Table 4-10. Models Built for the CLP, Fort Collins Case Study. Model Description Model 1 Model 2 Model 3 Model 4

Modelled Variable Turbidity Turbidity Turbidity TOC

Predictor Set Evaluated

Climate and land cover

Climate, land cover, and lagged turbidity

Climate, land cover, and TOC

Climate and land cover

Data Set Used 1 1 2 2

Fort Collins Case Study – Models 1 and 2 Results Model 1 and Model 2 both modelled turbidity in the CLP using data set 2. Data set 2 included monthly turbidity values for each month of the year and had a greater number of turbidity data points than data set 1, making data set 2 advantageous for capturing annual variability in turbidity. Model 1 is a GLM with gamma family, while Model 2 is a linear model. Both models 1 and 2 resulted in the modelled data fitting well to the observed data, shown as the high R squared value of 0.62 for Model 1 and 0.71 for Model 2 in Figure 4-14. The figure illustrates that both models performed well at capturing variability in the observations. Model 2 shows an improvement in model fit over Model 1 due to incorporating lagged turbidity as a predictor variable.

Figure 4-14. Scatterplots of Observed versus Modelled Turbidity for Models 1 and 2.

Under each scatterplot is the list of predictors used in the model.


Prediction skill for the models was assessed using the drop 10% cross validation method. R squared was calculated to measure how close the predicted data points were to the actual dropped, data points for each of the 500 times 10% of the data was dropped and predicted. Both models performed well in prediction mode, as illustrated by Figure 4-15. This is depicted as boxplots with a relatively small spread in the box and the average from the model fitting with the entire dataset (i.e., the red dot) overlaid close to the median from the drop 10% cross validation data points. Model 2 was slightly more consistent in prediction mode as it was in fitting mode compared to Model 1, as shown by the R squared from the Model 2 fit being on the median of the boxplot.

Figure 4-15. Box Plots of R Squared from Performing the Drop 10% Cross Validation Method 500 Times. Left:

Results from Model 1. Right: Results from Model 2. The red dot indicates the R squared from assessing the modelled turbidity versus the observed turbidity using the

entire data set (from Figure 4-14).

Fort Collins Case Study – Models 3 and 4 Results Models 3 and 4 utilized data set 2. Data set 2 allowed for model development on TOC and for an analysis between TOC and turbidity levels in the CLP. Model 3 was on turbidity using TOC as a predictor, as TOC in the predictor dataset was found to improve model fitting to turbidity observations. Model 4 is on TOC in the CLP. For both Models 3 and 4, the best model type of all the 5 modelling methods explored was the CART model.

Model 3 is a CART model on turbidity where the dominant predictors were TOC and PDSI one month prior. 4 possible levels of turbidity in the CLP were identified based on TOC and PDSI 1 month prior: 3.20 NTU, 5.72 NTU, 0.79 NTU, and 1.90 NTU. Figure 4-16 depicts the CART Model 3, model fit between the observed and modelled turbidity, and the boxplot of R squared from prediction mode. Model 3 performed well at fitting and predicting. R squared from assessing the modelled versus observed turbidity was high at 0.78, which was consistent with the model in prediction mode as evidenced by the median of the drop 10% cross validation boxplot.

Figure 4-16. Model 3 on Turbidity in the CLP.

Left to right: (1) the CART model, (2) scatter plot demonstating model fit between the observed and modelled data, (3) boxplot of R squared from the drop 10% cross validation, with the red dot indicating the R squared from

fitting the model to the entire dataset.


Model 4 is the only model on TOC for the CLP river. The CART model is depicted in Figure 4-17, as well as the model fit and predictive skill. Three variables are shown in Model 4 as necessary in modelling TOC: the NDVI 2 months prior, minimum temperature of the month, and NDVI 1 month prior. Relating these predictors results in four possible TOC concentrations for the CLP: 3.00 mg/L, 3.78 mg/L, 5.74 mg/L, and 8.02 mg/L. Model 4 out performed the three models on turbidty in the CLP in terms of model fit and predictive skill. As shown in Figure 4-17, model fit had an R squared of 0.83. The model in prediction mode resulted in similar R squared values. The model demonstrated stability in prediction mode, as depicted via the small range of the boxplot shown in Figure 4-17.

Figure 4-17. Model 4 on TOC in the CLP River.

Left to right: (1) the CART model, (2) scatter plot demonstating model fit between the observed and modelled data, (3) boxplot of R squared from the drop 10% cross validation, with the red dot indicating the R squared from

fitting the model to the entire dataset.

4.4.3.2 Houston Case Study Two models were built for water quality on the Trinity River in Houston, one on turbidity and on TOC. Model 5 is a linear model on turbidity and Model 6 is a GLM on TOC.

Model 5 had excellent fit to the turbidity data with an R squared of 0.81. This is in large part to incorporating lagged turbidity in the predictor set, as the R squared from model fitting without lagged turbidity was 0.45. Model 6 also shows a good fit to the TOC observations, where the R squared was 0.58. Predictors and model fits for both models are provided in Figure 4-18.


Figure 4-18. Scatterplots of Observed versus Modelled Turbidity for Models 5 and 6.

Under each scatterplot is the list of predictors used in the model.

Models 5 and 6 also performed well in prediction mode, as assessed via the drop 10% cross validation method. Boxplots of the cross validation R squared results are presented in Figure 4-19. For both models, the R squared from fitting aligned closely with the median of the R squared from the 500 cross validations. Model 2 has a larger spread in the boxplot than desired, but overall still performed well in prediction mode with the median of the boxplot being close to the R squared from fitting.

Figure 4-19. Box Plots of R Squared from Performing the Drop 10% Cross Validation Method 500 Times.

The red dot indicates the R squared from assessing the modelled turbidity versus the observed turbidity using the entire data set (i.e., from Figure 4-14). Left: results from Model 1. Right: results from Model 2.

4.4.4 Conclusions Results from this activity show that water quality can be modelled using climate and land cover data. Results from several models were presented: 3 models on turbidity in the CLP, 1 model on TOC in the CLP, 1 model on turbidity in the Trinity River, and one model on TOC in the Trinity River. All models were assessed to determine how well the model fit to the data and the skill of the model in prediction mode.

Model results for the CLP in Fort Collins compared to the Trinity River in Houston offer interesting information regarding key climate and land cover actors per location and water quality variable.


TOC models for both locations relied on NDVI as a predictor. This finding that NDVI is important for surface water TOC levels is likely due to much of the organic content concentrations being derived from vegetation. For the CLP, the best model incorporated NDVI and temperature related predictors, whereas NDVI and PDSI related predictors were chosen in the best model identified for the Trinity River on TOC. Because a large fraction of the Trinity River is derived from wastewater effluent that contains concentrations of TOC, PDSI may have emerged as a dominant predictor for the Trinity River due to its ability to capture episodes of the river being mainly wastewater derived (i.e., during droughts or low flow conditions in hot summer months).

The models on turbidity for both the CLP and Trinity River built using only climate and land cover predictors showed that predictors related to precipitation, temperature, and NDVI were all important. This indicates that vegetation may be a dominating source of turbidity in surface waters, mobilized by precipitation events. Temperature may be indicative of the state of vegetation and its likelihood to be mobilized into surface waters, for example autumn temperatures correlating with large quantities of leaves falling into streams. Lagging turbidity proved advantageous in capturing variability in turbidity for both cases studies, showing persistence in riverine turbidity levels. The main difference for the predictor sets of the Trinity River versus the CLP was the Trinity River turbidity model also included PDSI as a predictor. It is unclear exactly why PDSI is influential in modelling turbidity for the Trinity River, but it may be related to the ability of PDSI to capture dilution events of turbidity via wet conditions or to differentiate levels of turbidity from low, mainly wastewater derived flows versus flow events related to wetter conditions.

For turbidity modelling in the CLP River, the additional model incorporating TOC as a predictor in the dataset improved the model fit and also resulted in PDSI being an influential predictor. This finding alludes to the possibility of turbidity being correlated with TOC.

The results from this activity did not determine one statistical modelling method of the five investigated as superior over the others. That is, the results showed that the best model per constituent and source water was uniquely determined for each scenario. However, the novel method of balancing predictive skill with model fitting did lead to the conclusion that local polynomial models would likely never be the best model for a location and water quality constituent. For the water quality datasets analyzed, applying the local polynomial method always resulted in very poor predictive skill. This likely is due to the local polynomial method using a small subset of the data to predict the next data point, removing the ability for longer term trends in the water quality to be well captured and resulting in the polynomial order heavily influencing the predicted data points.


CHAPTER 5

Activity 4: Decision Support Tool for Adapting to Variable Water Quality and Competing Objectives The purpose of Activity 4 in this project is to evaluate a suite of adaptation strategies (e.g., watershed management, wildfire mitigation, treatment plant modifications) using multi-objective optimization techniques. Current practices in water treatment decision making rely heavily on engineering experience (Dudley et al., 2008). Computer-based decision making methods, known as decision support tools, can provide alternatives to traditional engineering practices and can increase the efficiency of treatment and reliability of water quality (Hamouda et al., 2009; Zhang et al., 2014). For a review of the current state of the art in water treatment decision support, see Raseman et al. (2017).

In this chapter, we describe the development of a decision support tool that leverages multi-objective optimization to produce innovative treatment solutions. Utilities can apply and adapt this tool to a variety of scenarios, such as changing source water conditions related to climate change, to improve the reliability and efficiency of their systems. The development of this tool involves five steps (Figure 5-1):

1. Stakeholder input: identify uncertainties, performance metrics, decisions, and constraints for the decision making problem

2. Iterative problem formulation: refine the problem based on available models, data, and additional feedback

3. Water quality scenarios: generate realistic influent water quality scenarios that a utility may experience

4. Simulation-optimization: combine simulations of water treatment with optimization to discover the best performing alternatives

5. Interactive visualization: use dynamic visualization techniques to select the alternatives preferred by the decision maker


Figure 5-1. Overview of Water Treatment Decision Support Tool for Choosing Treatment Alternatives That Are Resilient to Water Quality Variability: 1) Gather Stakeholder Input to Inform Problem Formulation, 2) Refine Problem Based on Additional Information, 3) Generate Influent Water Quality Scenarios, 4) Use Simulation-

Optimization Techniques to Generate Alternatives, and 5) Select among Alternatives Using Interactive Visualization Techniques.

The remainder of this chapter describes each component of this decision support tool and implications of its use.

5.1 Stakeholder Input The collaboration among stakeholders and analysts (i.e., modelers, optimization specialists) is essential for the development of useful optimization tools (Smith et al., 2017; Wu et al., 2016). To solicit feedback from stakeholders about their decision making needs and priorities, we organized a Water Research Foundation workshop in December 2016 with water utilities. Specifically, the goals of the workshop were to collect influent water quality data, survey the utilities’ use of simulation models, and inform the optimization problem. Participants of the workshop included water utilities, consultants, federal agencies, and research institutions, but focused heavily on the utilities’ views. These utilities represented eight different states in the United States (Figure 5-2) and served populations from as low as 60,000 to over 2,000,000.


Figure 5-2. Map of Utilities Participating in December 2016 Workshop and the Influent Water Quality Data

Collection Effort.

5.1.1 XLRM Framework To facilitate discussion at the workshop, we broke up participants into small groups and asked them to identify the core elements of their water treatment issues using the XLRM Framework (Lempert, 2003):

1. Uncertainties (X): factors beyond the decision maker's control 2. Decision Levers (L): actions decision makers can take to modify their system 3. Relationships (R): a tool which maps decision maker actions to outcomes, typically using a simulation

model 4. Measures (M): metrics used to gauge success

Using this framework, participants were equipped with a common vocabulary to discuss treatment problems and those discussions could then be translated into components of the decision support tool.

To ensure that these discussions representative of workshop participants, small groups were composed of utilities from different regions of the United States and had a similar balance of utilities, researchers, and consultants. Two facilitators ran each small group: a primary facilitator, who led the discussion, and a secondary facilitator, who took notes and synthesized findings to workshop participants. Facilitators were given concrete objectives and outputs for each discussion topic, so they could most effectively guide discussion. In small groups, workshop participants discussed current concerns for drinking water quality as well as future concerns related to climate change and extreme weather events. The conclusions of these XLRM discussions are documented in Table 5-1.

The small group discussions revealed that utilities had an array of questions and concerns. What are best practices for monitoring and preventing algal blooms? What steps had other utilities taken to prepare for wildfire? There was no shortage of topics to explore about water quality uncertainty or best practices for emergency management and source water protection. However, there was no clear consensus about which issues were the most pressing. Each utility had unique concerns based on the size of their utility, source water type and quality, water supply forecasts, geographic region, which is reflected in the XLRM framework ideas in Table 5-1. To transform these disparate ideas into a practical tool, we must first determine what simulation model and data are available. That information will shape the a revised XLRM framework, and ultimately, the nature of the decision support tool.


Table 5-1. XLRM Framework Ideas Generated at the December 2016 Workshop. Uncertainties (X) Decision Levers (L)

- Long-range water supply availability - Algal blooms - Extreme events - Snow-to-rain ratio - Rules, guidelines, and regulations

- Emergency protocol - Watershed protection - Alternative source blending - Pretreatment - Sensors/monitoring - Treatment plant upgrade

Relationships (R) Performance Measures (M)

- Water treatment modeling - Source water quality modeling - Hydrologic modeling - Water supply and demand forecasts - Pilot- and bench-scale testing

- Finished water standards - Taste and odor - Raw water quality

5.2 Iterative Problem Formulation During the workshop, it became clear that choosing an appropriate model would determine what other XLRM components could be considered for the tool. For instance, source water protection strategies were widely discussed; however, their impacts are difficult to quantify. In contrast, water treatment models—although not widely used in industry—provide much more accurate water quality and cost-related predictions. Because modeling seemed to be the driving force for the XLRM framework, we outlined the following steps for its refined formulation:

1. Relationships (R): select an appropriate treatment model 2. Uncertainties (X): quantify uncertainty in model inputs 3. Decision levers (L): identify relevant decisions based on model inputs and data availability 4. Performance measures (M): identify relevant metrics based on model outputs

5.2.1 Relationships: Select an Appropriate Treatment Model Several utilities and consultants discussed the use of distribution system models (e.g., EPANET, see Grayman et al., 2012; Rossman et al., 1994) and water supply/demand models for decision making. However, these models address issues of water quantity instead of water quality. Because the WRF 4636 project focuses on water quality impacts related to climate and extreme weather events, we chose to focus on water quality concerns. To do so, we needed to select and develop models to simulate source water quality and water treatment.

With regard to water treatment, the following models have been developed for a range of research and real-world applications (Dudley et al., 2008): OTTER (Head et al., 2002), Stimela (Helm and Rietveld, 2002), Metrex, Water Treatment Plant (WTP) Model (Harrington et al., 1992; Solarik et al., 2000), and WatPro. These models and other empirical approaches have been used to aid in operational decision making because they provide a detailed representation of system performance (e.g., Baxter et al., 1999; Maier et al., 2004; Rietveld et al., 2010; Worm et al., 2010). Among water treatment simulation models, the WTP Model created by the US Environmental Protection Agency (EPA) best suits the needs of this project because the code is publicly available and because it was developed specifically to estimate the performance of US water treatment plants.


5.2.2 Uncertainties: Quantify Uncertainty in Model Inputs The WTP Model inputs include source water quality parameters and the design and operation of the treatment plant. Among these WTP Model inputs, water quality parameters are uncertain due to seasonal variability and long-term changes. Therefore, we identified six critical water quality parameters to analyze—alkalinity, pH, temperature, total organic carbon (TOC), UV254 (ultraviolet absorbance at 254 nm), and bromide. To understand the uncertainty and variability of these parameters in source waters for average conditions and extreme events, we requested data from our partner utilities. Based on feedback from utilities about data sampling frequency and availability, we asked each of them to provide eight years of monthly data. Among the utilities, several provided complete or nearly complete datasets that we could analyze.

Figure 5-3. A Representative Example of Source Water Temperature Based on Data from Denver Water in

Colorado. Because the monthly pattern repeats itself from year to year, temperature is considered a strongly seasonal

variable.


Figure 5-4. A Representative Example of Total Organic Carbon Based on Data from Houston, TX.

There is some seasonality in the data but also considerable variability between years.

Based on analyses of the complete datasets, we observed that certain variables, like water temperature, are strongly seasonal—with a predictable monthly pattern from year to year (Figure 5-3). Other variables, like TOC, showed weak seasonality in most sources (Figure 5-4). These observations suggest that seasonality is an important aspect of water quality uncertainty. However, we found that it was difficult to quantify seasonality and other water quality characteristics across sources using a generalized water quality model. To address this difficultly, we developed a data-driven, site-specific model of water quality to capture realistic bounds of uncertainty. This novel method is described in the Water quality scenarios section.

5.2.3 Decision Levers: Identify Relevant Decisions Based on Model Inputs and Data Availability Decision levers are actions that decision makers can take to modify their system. Both infrastructural (i.e., adding/modifying unit processes) and operational (i.e., altering chemical doses or blending different sources waters) changes can be considered with the WTP Model. Bench- and pilot-scale treatability data can be used to calibrate the WTP Model to improve model estimates. The availability of these data may determine which decision levers can be considered. Because there is generally more data available for existing treatment configurations, operational changes are easiest to model reliably.

5.2.4 Performance Measures: Identify Relevant Metrics Based on Model Outputs The WTP Model outputs include data on water quality throughout the system, chemical usage, and solids handling production. Specifically, the WTP Model was developed to predict disinfection byproducts (DBPs) for surface water treatment plants across the US and was originally created to inform DBP regulations known as the Stage 1 and 2 Disinfection Byproduct Rules (EPA, 2010, p. 1). Therefore, performance measures related to disinfection and DBP risk follow logically from model outputs.


Although the WTP Model does not provide direct estimates of cost, cost measures can be calculated based on chemical usage and solids handling output by the model.

By combining the information from each of the previous four sections, Table 5-2 describes the final version of the XLRM framework developed for the decision support tool.

Table 5-2. A Revised XLRM Framework Based on Availability of Models and Data. Uncertainties (X) Decision Levers (L)

- Seasonal variability in influent water quality - Impact of extreme events on influent quality

- Operational changes, such as chemical dosing strategies

- Infrastructural modifications

Relationships (R) Performance Measures (M)

- Water treatment modeling - Source water quality modeling - Bench- and pilot-scale testing

- Disinfection byproduct risk - Operations and maintenance costs

5.3 Decision Support Tool Case Study To provide a use case for the decision support tool, we developed an example based on influent water quality and bench-scale treatability data for the Cache la Poudre (CLP) River in Fort Collins, Colorado. This background information on the CLP watershed are described in Hohner et al. (2016) and Writer et al. (2014). The data collection effort was supported by the City of Fort Collins Utility; however, please note that this case study is not intended to represent the City of Fort Collins system. Our case study has significant differences from their system regarding source waters considered and the design and operation of the treatment plant and distribution system.

We selected the CLP River water quality dataset for several reasons: it was longest and most complete dataset we received from partner utilities, the record included multiple extreme events (i.e., drought, wildfire, and flooding), and there was large seasonal variability in water quality. The seasonal variability and water quality impact of extreme events highlights the connection between climate and water quality in source waters, which made it a good candidate for this project. In the following sections, we describe parts 2-4 of the decision support tool process (Figure 5-1): water quality scenarios, simulation-optimization, and interactive visualization.

5.4 Water Quality Scenarios To quantify the uncertainty in the CLP source water quality, we developed a statistical method for generating water quality scenarios based on the historical record. Since the problem formulation outlined in Table 5-2 involves DBP risk, these water quality scenarios must capture the time-dependent nature of DBP regulation—which is based on a running annual average that is calculated quarterly—time series data are necessary. However, due to the limited length (i.e., a decade or less) and multivariate nature of water quality datasets commonly available, developing realistic scenarios is not trivial.

There are a wide variety of statistical methods to generate these scenarios, including non-parametric and parametric approaches. Non-parametric methods are useful when there is limited data to justify such assumptions. However, since these models resample the observed data (e.g., Lall and Sharma 1996), they cannot generate new values in their simulations. Therefore, resampling methods cannot simulate unobserved extremes that may realistically occur. In contrast, highly parameterized


approaches, such as multivariate autoregressive models, have the freedom to generate unobserved values. Such models can explicitly capture the joint correlation among water quality variables and temporal correlations. However, the amount of data necessary for adequate parameter estimation for these models is often impractically large (Bras and Rodríguez-Iturbe, 1985)—especially with the short datasets typical for water quality.

To address these challenges, we developed a modified k-Nearest Neighbor (k-NN) Bootstrap Resampling method (Raseman et al., in review). Unlike other non-parametric approaches, we have added variability to the resampled data to capture realistic extremes not seen on the observed record. Summarizing the results of Raseman et al. (In review), the authors found that despite the limited water quality data available, the modified k-NN was able to produce scenarios which captured the important statistical properties of the historical datasets. In other words, this technique was able to generate realistic water quality scenarios even with relatively short water quality records. Furthermore, they found that this technique was able to simulate water quality extreme events, which is a critical element of this project. To demonstrate this for the CLP River dataset, first we will discuss the case in which no random variation has been added to the simulations.

The boxplots in Figure 5-5 show how the k-NN simulations of the CLP River can reproduce the observed statistics of the historical record. The red and white boxplots describe the observed and simulated data, respectively. The black line inside the box represents the median value for a given quarter and the extents of the boxplots represent the 25th and 75th percentiles (i.e., the interquartile range). The lines—or whiskers—represent data that is no farther than 1.5 times the interquartile range beyond the box. All data beyond the whiskers are represented as points. By comparing the observed and simulated box plots, there are few differences between them. This means that the simulated data has similar statistics as the observed data. In other words, the simulated data captures the nature of the historical record with respect the statistics represented by the box plots. In addition to these statistics, known as “distributional statistics,” metrics such as autocorrelation and joint correlation are significant as well.

Joint correlation describes the relationship between water quality parameters. For instance, alkalinity and pH are interconnected parameters. The presence of alkalinity allows water to buffer from changes in pH . Capturing this relationship—i.e., the joint correlation between water quality parameters—is necessary for generating realistic scenarios. The autocorrelation is a measure of the “persistence” or “memory” of a given source water. For instance, if organic carbon is high in one month, it may be likely to stay high in the next month. If this is the case, that persistence in water quality should be captured in the simulations. The results show that the distributional statistics, autocorrelation, and joint correlation are largely preserved for short, water quality datasets like the CLP data. Moreover, this is still true when variation is added to these data, which allows for the inclusion of realistic extremes in the scenarios.


Figure 5-5. Quarterly Boxplots of the Observed Record (nyears = 11) and Simulated Data (nsimulations = 500 Each of nyears = 11) Simulations for the Fort Collins Cache la Poudre Dataset (2007-2017).

5.5 Simulation-Optimization The CLP water quality scenarios generated by the modified k-NN algorithm can serve as water quality inputs to the WTP Model. Using these scenarios, the WTP Model can estimate the finished water quality for a given treatment plant design and operating policy. This approach allows stakeholders to consider any number of plausible situations that they may experience and evaluate which decisions lead to the best performance. However, evaluating hand-picked alternatives is inefficient and does not guarantee the optimal set of decisions can be identified. The search for optimal solutions can be automated by coupling a water treatment simulation model with an optimization algorithm. This coupled type of optimization is known as simulation-optimization.

In this work, we consider using a multi-objective evolutionary algorithm (MOEA) to search for optimal operating decisions. MOEAs optimize for multiple objectives simultaneously, resulting in a tradeoff set that can help identify the relationship between competing objectives. In this case study, we are considering the tradeoffs between DBP risk and operational costs. Pareto optimality is used to define the tradeoffs; a solution is Pareto optimal if its performance in any objective is not exceeded by any other feasible solution in any objective (Reed et al., 2013). For this research, we used the Borg MOEA, a state-of-the-art adaptive MOEA (Hadka and Reed, 2013), which we coupled with the WTP Model to perform the optimization.

As shown in Figure 5-6, the XLRM framework outlined in Table 5-2 informs each component of the simulation-optimization workflow. The water quality scenarios quantify the uncertainty (X) of influent quality, the decision levers (L) are based on WTP Model inputs for treatment operations, the relationships (R) are provided by the WTP Model, and the model outputs are translated into performance measures (M) which are divided into objectives and constraints. In the following sections, we describe the details of the XLRM framework for this case study.


Figure 5-6. A Conceptual Diagram of the Optimization Using a Multi-Objective Evolutionary Algorithm (MOEA) and the WTP Model to Evaluate Operational Decisions for Treating Cache la Poudre River Water.

Because we discussed the water quality scenarios (i.e., representation of uncertainty) previously, we will address the other XLRM components below. For the WTP Model, a conventional treatment train was simulated (Figure 5-7). Conventional treatment is representative of typical surface water treatment plants that treat CLP water and is a common surface water treatment plant design in the US. Due to low alkalinity in the CLP River (see Figure 5-5), pH adjustment using lime and carbon dioxide is included at the front of the plant. To ensure that disinfection regulations are met and that corrosion control measures are taken, automatic chemical dosing logic was added to the WTP Model to maintain a chlorine residual in the distribution system and set the effluent pH of the treatment plant.

Figure 5-7. Conventional Treatment Train Simulated Using the WTP Model. Chemical dosing points related to the decision levers are highlighted in blue.

The DBP models within the WTP Model were calibrated using the results of bench-scale treatability tests described in Hohner et al. (2016). Because conventional treatment was the only treatment train with bench-scale data for calibration, no other treatment designs were considered for the simulation-optimization case study. This optimization reflects a case in which decision makers are searching for the best way to operate an existing plant. Specifically, the decision levers chosen for operations optimization are water quality setpoints and a so-called DBP safety factor. A value for each decision lever is chosen for each quarter of the year to adapt to seasonal changes in water quality.


5.5.1 Decision Levers The decision levers include water quality setpoints for pH and alkalinity. These setpoints are both located at the front of the treatment plant before the rapid mix basin. These decision levers were chosen to account for the low alkalinity of the CLP and high organic carbon concentrations, especially in the second quarter of the year. Since organic carbon is the primary driver of DBP formation, organic carbon removal is essential to minimize DBP risk. In conventional treatment, coagulation/flocculation is an important carbon removal mechanism. However, the addition of coagulants like aluminum sulfate (i.e., alum) can lower the pH of treated water (Edzwald and Benschoten, 1990). Without enough alkalinity to buffer pH change, organic carbon removal due to coagulation can be limited. By altering the pH and alkalinity of the influent water, additional organic carbon can be removed via coagulation.

The disinfection byproduct safety factor decision lever controls two different chemical dosing units in the treatment plant: the coagulant and disinfectant dose. The dose is adjusted based on a safety factor, defined with respect to the maximum contaminant limit (MCL) for the regulated groups of DBPs: total trihalomethanes (TTHM) and five species of haloacetic acids (HAA5). A safety factor of 0.25 means that, if possible, the decision maker would like to stay at least 25% below the MCL for that quarter. Imagine that a treatment plant can achieve a DBP safety factor of 0.25 for each quarter except during quarter two due to high organic carbon concentrations during that time of year (Figure 5-5). Because DBPs are regulated on a running annual average basis and not quarterly, the utility can go above the MCL in quarter two as long as the annual average is below the MCL. In general, utilities would prefer not to have DBP concentrations above the MCL in any quarter, if possible. This preference is reflected in the objectives chosen for this case study.

5.5.2 Performance Measures: Objectives The objectives are a subset of the performance measures (M) and reflect a desire to minimize or maximize some aspect of performance. All objectives are calculated with respect to treatment performance across all water quality scenarios that were considered. To reflect a utility’s risk aversion to concentrations of DBPs above the MCL in any quarter, two of the objectives are to minimize the worst-case frequency of quarterly 1) TTHM and 2) HAA5 concentrations being greater than the MCL. The term worst-case is defined as the worst value (i.e., the highest frequency of DBP concentrations > MCL) for any of the water quality scenarios. Therefore, these objectives reflect performance for a single water quality scenario. The worst-case metric represents aversion toward DBP risk and increased risk of violating regulations. The other three objectives are related operational costs due to chemical dosing and solids handling.

Unlike DBP risk, we assume that utilities are risk neutral to decisions about operational costs; meaning that the utility wants to lower costs on average and could cope with high operating costs in the unlikely event that high costs were to occur. For this reason, cost objectives are calculated as the average performance, or expected value, across all water quality scenarios. Specifically, the two chemical dosing objectives are to minimize the expected 1) lime dose and 2) CO2 dose. The third cost objective is to minimize the expected solids production. To formulate these objectives in a more risk-averse manner, one could calculate them more similarly to the DBP risk objectives. For instance, one could consider the worst-case or the 75th or 95th percentiles of poor performance.

5.5.3 Performance Measures: Constraints In addition to objectives, the other subset of performance measures is constraints. In this case study, there is a single regulatory constraint: maintaining a locational running annual average for both TTHM and HAA5 that is less than the MCL. The running annual average is the average concentration of DBPs


over the past four quarters of the year. The “locational” running annual average means that the sampling point with the highest average DBP concentration determines whether the utility is in compliance. Other regulations, such as disinfection credits, maintenance of a disinfection residual during distribution, and organic carbon removal, are enforced within the automatic chemical dosing logic described earlier. In contrast, constraints are thresholds that are enforced by the optimization algorithm.

The above discussion of decision levers, objectives, and constraints is summarized in Table 5-3.

Table 5-3. Summary of Decision Levers and Performance Measures. The objectives are calculated based on aggregate treatment performance for all water quality scenarios

considered. The term expected refers to the average value across all scenarios. The term worst-case is value associated with the worst outcome of any scenario.

Decision Levers Objectives Constraints

pH setpoint Minimize expected lime dose

Locational running annual average of regulated DBPs < MCL

Minimize expected CO2 dose

Alkalinity setpoint

Minimize expected solids production

Minimize worst-case frequency of quarters with TTHM > MCL

DBP safety factor Minimize worst-case frequency of quarters with HAA5 > MCL

5.5.4 Pareto Optimal Solutions Putting each component of the XLRM framework together, the MOEA generates Pareto optimal solutions, such as those shown in Figure 5-8. As a reminder, a solution is Pareto optimal if its performance in any objective is not exceeded by any other feasible solution in any objective. Here, we use the term “feasible” to mean solutions that meet the constraints. The resulting set of Pareto optimal solutions reflects the tradeoffs between conflicting objectives in the problem. To determine which solutions are “best,” the decision maker needs to explore the data and use their judgement to select the solutions that they prefer. Due to the large number of Pareto optimal solutions, interactive visualization techniques are used to facilitate this process. These techniques are described in the next section: Interactive visualization.

One way to visualize these solutions is using a parallel coordinates plot (Inselberg, 2009). In Figure 5-8, for example, the “polylines” represent the performance of each Pareto optimal operating policy. The polylines highlighted in red, blue, and green illustrate this concept most clearly. Each axis corresponds to one of the five objectives, and for each objective, the preferred direction is down. Therefore, the ideal solution would be a straight polyline across the bottom of each axis. However, because there are tradeoffs between objectives, this ideal solution is not possible to achieve. Crossing lines between axes represent tradeoffs between objectives, such as between the expected solids and expected lime dose axes. Whereas straight lines represent no conflict, such as between the expected lime and CO2 dose axes.

Each of the polylines represent a Pareto optimal operating policy; however, we will limit our discussion to the policies highlighted in red, blue, and green. Each of these policies represents a unique set of decision maker preferences and priorities. The red polyline represents an operating policy that has the


least possible DBP risk—zero for both TTHM and HAA5—and moderate costs related to solids handling and chemical doses. The person that would choose the red line is extremely risk averse to DBP concentrations above the MCL in any quarter. The blue operating policy has higher TTHM risk (concentrations were above the MCL in 10% of quarters in the worst-case scenario), lower solids production, and higher chemical dosing costs. The green policy has moderate TTHM risk (5% of quarters with high concentrations in the worst-case) but among the lowest chemical dosing costs of any of the MOEA-generated solutions. These three operating policies represent different risk and cost preferences, but it is important to note that any one of the MOEA-generated policies would satisfy drinking water regulations.

Figure 5-8. Parallel Coordinates Plot of Pareto Optimal Operating Policies Produced by the Simulation-

Optimization Outlined in Figure 5-6 and Table 5-3. The different operating policies reflect different decision maker priorities. Each polyline on the plot is a Pareto

optimal policy; however, only the three highlighted solutions are examined in detail.

5.6 Interactive Visualization As evidenced in Figure 5-8, static visualizations often fail to display Pareto optimal solutions effectively. Such parallel coordinates plots can become overcrowded and difficult to interpret. For this reason, these data should be analyzed using interactive visualization techniques. Such techniques allow decision makers to sift through these large, multi-dimensional data sets. To support this process, we developed Parasol, an open-source, interactive visualization library for creating web applications (Raseman et al., 2019). These web applications are sharable and easy to use for decision makers. To demonstrate its use, we have developed a Parasol-based application for the water treatment case study described in this chapter (Figure 5-9).


Figure 5-9. A Parasol-Based Web Application for Interactive Visualization of Water Treatment Operating Policies

Generated by the Borg MOEA. The performance of each solution is represented in the “objectives” plot and the operating decisions are shown in the “decisions” plot. Both plots are linked to one another and an interactive data table. By hovering a mouse over a row in the data table the user can highlight a single operating policy, as shown in the plot. Note that this data set

represents a slightly different optimization than in Figure 5-8. Click the following link or copy the URL into a browser to use the web application: https://wraseman.github.io/parasol/demo/water-treatment.html

Parasol has several features which improve interpretability and aid in the decision making process. Using “clutter reduction techniques,” such as brushing (Becker and Cleveland, 1987), highlighting, linking, and clustering (Luo et al., 2008), Parasol makes it easier for users to differentiate between solutions and select the best ones. Brushing, for instance, allows the user to filter the data dynamically, removing any unwanted solutions. Highlighting, as shown in Figure 5-9, enables users to highlight a single solution of interest. Clustering, a statistical method used to identify similar groups of data, is used in Parasol apps to sort optimal solutions into categories. By clustering the solutions—clusters are shown in orange, blue, and green in Figure 5-9—the user can examine a few groups of solutions rather than analyzing hundreds of individual solutions. Linking is another critical feature of Parasol. By linking plots and interactive data tables together, the user can examine the performance of various solutions and the associated decision levers with relative ease. This information is shown in the “Objectives” and “Decisions” plots of Figure 5-9 for the CLP water treatment case study. The interested reader can follow the link in the figure caption to use the web application.

The interactive visualization step is the last component of the decision support tool. It brings each of the other components—stakeholder feedback, the iterative problem formulation, and simulation-


optimization—into an actionable tool with which decision maker can improve the operations of their treatment plant.

5.7 Summary and Conclusions The decision support tool developed in this chapter consists of five components: stakeholder input, iterative problem formulation, water quality scenarios, simulation-optimization, and interactive visualization. As shown by the Cache la Poudre (CLP) case study, this tool can generate resilient drinking water solutions for decision makers faced with variable water quality and competing treatment objectives. The tool achieves this by testing solutions against a range of plausible water quality scenarios, taking into account the inherent variability in source water quality. For a historical record like the CLP River, such scenarios represent a range of average conditions and extreme events, such as wildfire, drought, and flooding. Using multi-objective optimization, the decision makers can explore the relationship between conflicting objectives such as risk and cost. By assessing the tradeoffs between these objectives, decision makers can use interactive visualizations to explore optimal solutions and implement the engineering solution that aligns with their values and preferences.

This tool and other decision support techniques allow researchers to combine stakeholder knowledge with simulation to better design and operate our drinking water infrastructure considering changing climate and extremes. Decision support gives decision makers the power to ask “what if” before an extreme event occurs and allows managers to discover proactive solutions.


CHAPTER 6

Conclusion This project has contributed several scientific advancements that constitute an integrated framework for addressing water quality hazards and supporting decisions for water treatment plants. The four activities required new experimentation, modeling, and method development that has been detailed in this report. Therefore, one of the main deliverables are the attendant methods that were explained and can be applied to water systems by managers and engineers to deal with future external stressors.

In particular, Chapter 4 contributed a modeling framework for water quality variables that uses land and climate variables as predictors, using threshold exceedance and tree-based models. The modeling framework allows a user to generate future predictions of water quality based on likely scenarios of predictors (including variables such as temperature and drought indices). The approach has suggested how utilities can develop models for their own respective regions and create future scenarios. Given the uncertainty associated with climate model projections and socioeconomic scenarios, the specific application of climate change information within these models is an area of future research. However, the authors feel that the models presented in this work can be applied to climate change research in a straightforward fashion, given this information. Similarly, the decision support in Chapter 5 can be tailored to utilities’ needs; since the water quality timeseries can be trained on any statistics, the methods can be adapted for climate change research.


References

Akaike, H. 1998. “A New Look at the Statistical Model Identification.” In Parzen, E., Tanabe, K., Kitagawa, G. (Eds.), Selected Papers of Hirotugu Akaike. New York: Springer. https://doi.org/10.1007/978-1-4612-1694-0_16.

Albalasmeh, A.A., Berli, M., Shafer, D.S., and Ghezzehei, T.A. 2013. “Degradation of Moist Soil Aggregates by Rapid Temperature Rise under Low Intensity Fire.” Plant Soil, 362: 335–344. https://doi.org/10.1007/s11104-012-1408-z.

Arnold, J.G., Srinivasan, R., Muttiah, R.S., and Williams, J.R. 1998. “Large Area Hydrologic Modeling and Assessment Part I: Model Development.” J. Am. Water Resour. Assoc., 34: 73–89. https://doi.org/10.1111/j.1752-1688.1998.tb05961.x.

Atanasova, N., and Kompare, B. 2002. Modelling of Wastewater Treatment Plant with Decision and Regression Trees. Third Workshop on Binding Environmental Sciences and Artificial Intelligence.

Baldock, J., and Skjemstad, J. 2000. “Role of the Soil Matrix and Minerals in Protecting Natural Organic Materials Against Biological Attack.” Org. Geochem., 31: 697–710. https://doi.org/10.1016/S0146-6380(00)00049-8.

Baxter, C.W., Stanley, S.J., and Zhang, Q. 1999. "Development of a Full-Scale Artificial Neural Network Model for the Removal of Natural Organic Matter by Enhanced Coagulation.” J. Water Serv. Res. Technol.-Aqua, 48: 129–136.

Becker, R.A., and Cleveland, W.S. 1987. “Brushing Scatterplots.” Technometrics, 29: 127–142. https://doi.org/10.1080/00401706.1987.10488204.

Bicknell, B.R., Imhoff, J.C., Kittle Jr, J.L., Donigian Jr, A.S., and Johanson, R.C. 1996. Hydrological Simulation Program-FORTRAN. User’s Manual for Release 11. EPA.

Bodí, M.B., Martin, D.A., Balfour, V.N., Santín, C., Doerr, S.H., Pereira, P., Cerdà, A., and Mataix-Solera, J. 2014. “Wildland Fire Ash: Production, Composition and Eco-Hydro-Geomorphic Effects.” Earth-Sci. Rev., 130: 103–127. https://doi.org/10.1016/j.earscirev.2013.12.007.

Box, G.E.P., Jenkins, G.M., Reinsel, G.C., and Ljung, G.M. 2015. Time Series Analysis: Forecasting and Control. John Wiley & Sons.

Bras, R.L., and Rodríguez-Iturbe, I. 1985. Random Functions and Hydrology. Courier Corporation.

Breiman, L. 2001. “Random Forests.” Mach. Lear., 45: 5–32. https://doi.org/10.1023/A:1010933404324.

Burton, A., and Bathurst, J.C. 1998. “Physically Based Modelling of Shallow Landslide Sediment Yield at a Catchment Scale.” Environ. Geol., 35: 89–99. https://doi.org/10.1007/s002540050296.

Cantor, K.P., Villanueva, C.M., Silverman, D.T., Figueroa, J.D., Real, F.X., Garcia-Closas, M., Malats, N., Chanock, S., Yeager, M., Tardon, A., Garcia-Closas, R., Serra, C., Carrato, A., Castano-Vinyals, G., Samanic, C., Rothman, N., and Kogevinas, M. 2010. “Polymorphisms in GSTT1, GSTZ1, and CYP2E1, Disinfection By-products, and Risk of Bladder Cancer in Spain.” Environ. Health Perspect., 118: 1545–1550. https://doi.org/10.1289/ehp.1002206.


Cawley, K.M., Hohner, A.K., Podgorski, D.C., Cooper, W.T., Korak, J.A., and Rosario-Ortiz, F.L. 2017. “Molecular and Spectroscopic Characterization of Water Extractable Organic Matter from Thermally Altered Soils Reveal Insight into Disinfection Byproduct Precursors.” Environ. Sci. Technol., 51: 771–779. https://doi.org/10.1021/acs.est.6b05126.

Center for Drinking Water Optimization. 2001. Water Treatment Plant Model, Version 2.0, User’s Manual. Boulder, CO: Center for Drinking Water Optimization.

Clark, R.M., Adams, J.Q., and Lykins, B.W. 1994. “DBP Control in Drinking Water: Cost and Performance.” J. Environ. Eng., 120: 759–782. https://doi.org/10.1061/(ASCE)0733-9372(1994)120:4(759).

Coles, S. 2001. An Introduction to Statistical Modeling of Extreme Values. London: Springer-Verlag.

Cowman, G.A., and Singer, P.C. 1996. “Effect of Bromide Ion on Haloacetic Acid Speciation Resulting from Chlorination and Chloramination of Aquatic Humic Substances.” Environ. Sci. Technol., 30: 16–24. https://doi.org/10.1021/es9406905.

DeBano, L.F., Neary, D.G., and Ffolliott, P.F. 1998. Fire Effects on Ecosystems, 1 ed. New York: Wiley.

Delpla, I., Jung, A.-V., Baures, E., Clement, M., and Thomas, O. 2009. “Impacts of Climate Change on Surface Water Quality in Relation to Drinking Water Production.” Environ. Int., 35: 1225–1233. https://doi.org/10.1016/j.envint.2009.07.001.

Dudley, J., Dillon, G., and Rietveld, L.C. 2008. “Water Treatment Simulators.” J. Water Supply Res. Technol. – Aqua, 57: 13–21. https://doi.org/10.2166/aqua.2008.096.

Dunn, P.H., Barro, S.C., and Poth, M. 1985. “Soil Moisture Affects Survival of Microorganisms in Heated Chaparral Soil.” Soil Biol. Biochem., 17: 143–148. https://doi.org/10.1016/0038-0717(85)90105-1.

Edwards, M. 1997. “Predicting DOC Removal during Enhanced Coagulation.” J. Am. Water Works Assoc., 89: 78–89. https://doi.org/10.1002/j.1551-8833.1997.tb08229.x.

Edzwald, J.K., and Benschoten, J.E.V. 1990. "Aluminum Coagulation of Natural Organic Matter.” In Chemical Water and Wastewater Treatment. Berlin, Heidelberg: Springer. https://doi.org/10.1007/978-3-642-76093-8_22.

Eidenshink, J., Schwind, B., Brewer, K., Zhu, Z.-L., Quayle, B., and Howard, S. 2007. "A Project for Monitoring Trends in Burn Severity.” Fire Ecol., 3: 3–21. https://doi.org/10.4996/fireecology.0301003.

EPA (U.S. Environmental Protection Agency). 1998. National Primary Drinking Water Regulations: Stage 1 Disinfectants and Disinfection Byproducts Rule. Federal Register, National Archives and Records Administration.

EPA (U.S. Environmental Protection Agency). 2010. “Comprehensive Disinfectants and Disinfection Byproducts Rules (Stage 1 and Stage 2): Quick Reference Guide.” Accessed 4.7.17. https://nepis.epa.gov/Exe/ZyPDF.cgi?Dockey=P100C8XW.txt.

Falcone, J.A. 2011. GAGES-II: Geospatial Attributes of Gages for Evaluating Streamflow. Reston, VA: U.S. Geological Survey.


Freeman, C., Evans, C.D., Monteith, D.T., Reynolds, B., and Fenner, N. 2001. “Export of Organic Carbon from Peat Soils.” Nature, 412: 785. https://doi.org/10.1038/35090628.

Fried, J.S., Torn, M.S., and Mills, E. 2004. “The Impact of Climate Change on Wildfire Severity: A Regional Forecast for Northern California.” Clim. Change, 64: 169–191. https://doi.org/10.1023/B:CLIM.0000024667.89579.ed.

Gray, J.R., and Simoes, F.J.M. 2008. “Estimating Sediment Discharge: Appendix D.” In Sedimentation Engineering: Processes, Measurements, Modeling, and Practice. American Society of Civil Engineers.

Grayman, W.M., Kshirsagar, S., Rivera-Sustache, M., and Ginsberg, M. 2012. “An Improved Water Distribution System Chlorine Decay Model Using EPANET MSX.” J. Water Manag. Model. https://doi.org/10.14796/JWMM.R245-21.

Haagenson, E., Rajagopalan, B., Summers, R.S., and Roberson, J.A. 2013. “Projecting Demand Extremes under Climate Change Using Extreme Value Analysis.” J. Am. Water Works Assoc., 105: E40–E50. https://doi.org/10.5942/jawwa.2013.105.0013.

Hadka, D., and Reed, P. 2013. “Borg: An Auto-Adaptive Many-Objective Evolutionary Computing Framework.” Evol. Comput., 21: 231–259. https://doi.org/10.1162/EVCO_a_00075.

Hamouda, M.A., Anderson, W.B., and Huck, P.M. 2009. “Decision Support Systems in Water and Wastewater Treatment Process Selection and Design: A Review.” Water Sci. Technol., 60: 1757. https://doi.org/10.2166/wst.2009.538.

Harden, J.W., Mack, M., Veldhuis, H., and Gower, S.T. 2002. “Fire Dynamics and Implications for Nitrogen Cycling in Boreal Forests.” J. Geophys. Res. Atmospheres, 107: WFX 4-1-WFX 4-8. https://doi.org/10.1029/2001JD000494.

Harrington, G.W., Chowdhury, Z.K., and Owen, D.M. 1992. “Developing a Computer Model to Simulate DBP Formation During Water Treatment.” J. Am. Water Works Assoc., 84: 78–87.

Hastie, T., Tibshirani, R., and Friedman, J. 2009. The Elements of Statistical Learning: Data Mining, Inference, and Prediction, 2nd ed. New York: Springer-Verlag.

Head, R., Shepherd, D., Butt, G., and Buck, G. 2002. “OTTER Mathematical Process Simulation of Potable Water Treatment.” Water Sci. Technol. Water Supply, 2: 95–101.

Heath, J., and Oropeza, J. 2016. 2015 Annual Report of the Upper Cache la Poudre Watershed Collaborative Water Quality Monitoring Program. Fort Collins, CO: City of Fort Collins Utilities.

Heller-Grossman, L., Manka, J., Limoni-Relis, B., and Rebhun, M. 1993. “Formation and Distribution of Haloacetic Acids, THM and TOX in Chlorination of Bromide-Rich Lake Water.” Water Res., 27: 1323–1331. https://doi.org/10.1016/0043-1354(93)90219-8.

Helm, A.W.C. van der, and Rietveld, L.C. 2002. “Modelling of Drinking Water Treatment Processes within the Stimela Environment. Water Sci. Technol.” Water Supply, 2: 87–93.

Hohner, A.K., Cawley, K., Oropeza, J., Summers, R.S., and Rosario-Ortiz, F.L. 2016. “Drinking Water Treatment Response Following a Colorado Wildfire.” Water Res., 105: 187–198. https://doi.org/10.1016/j.watres.2016.08.034.


Hope, D., Billett, M.F. and Cresser, M.S. 1994. “A Review of the Export of Carbon in River Water: Fluxes and Processes.” Environ. Pollut., 84: 301–324. https://doi.org/10.1016/0269-7491(94)90142-2.

Inselberg, A. 2009. Parallel Coordinates: Visual Multidimensional Geometry and Its Applications. New York: Springer-Verlag.

IRI (International Research Institute for Climate and Society). 2019. “IRI/LDEO Climate Data Library.” http://iridl.ldeo.columbia.edu/.

Jansson, M.B. 1988. "A Global Survey of Sediment Yield.” Geogr. Ann. Ser. Phys. Geogr., 70: 81–98. https://doi.org/10.2307/521127.

Johanson, R.C., and Davis, H.H. 1980. Users Manual for Hydrological Simulation Program-Fortran (HSPF). Environmental Protection Agency.

Kampf, S.K., Brogan, D.J., Schmeer, S., MacDonald, L.H., and Nelson, P.A. 2016. “How Do Geomorphic Effects of Rainfall Vary with Storm Type and Spatial Scale in a Post-Fire Landscape?” Geomorphology, 273: 39–51. https://doi.org/10.1016/j.geomorph.2016.08.001.

Karanfil, T., Erdogan, I., and Schlautman, M.A. 2003. "Selecting Filter Membranes for measuring DOC and UV254.” J. Am. Water Works Assoc., 95: 86–100. https://doi.org/10.1002/j.1551-8833.2003.tb10317.x.

Katz, R.W., Parlange, M.B., and Naveau, P. 2002. "Statistics of Extremes in Hydrology.” Adv. Water Resour. 25: 1287–1304. https://doi.org/10.1016/S0309-1708(02)00056-8.

Keeley, J.E. 2009. “Fire Intensity, Fire Severity and Burn Severity: A Brief Review and Suggested Usage.” Int. J. Wildland Fire, 18: 116–126. https://doi.org/10.1071/WF07049.

Knapen, A., Poesen, J., Govers, G., Gyssels, G., and Nachtergaele, J. 2007. “Resistance of Soils to Concentrated Flow Erosion: A Review.” Earth-Sci. Rev., 80: 75–109. https://doi.org/10.1016/j.earscirev.2006.08.001.

Knicker, H. 2007. “How Does Fire Affect the Nature and Stability of Soil Organic Nitrogen and Carbon? A Review.” Biogeochemistry, 85: 91–118. https://doi.org/10.1007/s10533-007-9104-4.

Knicker, H., Almendros, G., González-Vila, F.J., Martin, F., and Lüdemann, H.-D. 1996. “13C- and 15N-NMR Spectroscopic Examination of the Transformation of Organic Nitrogen in Plant Biomass during Thermal Treatment.” Soil Biol. Biochem., 28: 1053–1060. https://doi.org/10.1016/0038-0717(96)00078-8.

Köhler, S.J., Buffam, I., Laudon, H., and Bishop, K.H. 2008. “Climate’s Control of Intra-Annual and Interannual Variability of Total Organic Carbon Concentration and Flux in Two Contrasting Boreal Landscape Elements.” J. Geophys. Res. Biogeosciences, 113. https://doi.org/10.1029/2007JG000629.

Kondolf, G.M., Gao, Y., Annandale, G.W., Morris, G.L., Jiang, E., Zhang, J., Cao, Y., Carling, P., Fu, K., Guo, Q., Hotchkiss, R., Peteuil, C., Sumi, T., Wang, H.-W., Wang, Z., Wei, Z., Wu, B., Wu, C., and Yang, C.T. 2014. “Sustainable Sediment Management in Reservoirs and Regulated Rivers: Experiences from Five Continents.” Earths Future, 2: 256–280. https://doi.org/10.1002/2013EF000184.


Krasner, S.W., McGuire, M.J., Jacangelo, J.G., Patania, N.L., Reagan, K.M., and Aieta, E.M. 1989. “The Occurrence of Disinfection By-products in US Drinking Water.” J. Am. Water Works Assoc., 81: 41–53. https://doi.org/10.1002/j.1551-8833.1989.tb03258.x.

Lall, U., and Sharma, A. 1996. “A Nearest Neighbor Bootstrap For Resampling Hydrologic Time Series.” Water Resour. Res., 32: 679–693. https://doi.org/10.1029/95WR02966.

Larsen, M.A.D., Christensen, J.H., Drews, M., Butts, M.B., and Refsgaard, J.C. 2016. “Local Control on Precipitation in a Fully Coupled Climate-Hydrology Model.” Sci. Rep., 6. https://doi.org/10.1038/srep22927.

Lempert, R.J. 2003. Shaping the Next One Hundred Years: New Methods for Quantitative, Long-Term Policy Analysis. Rand Corporation.

Liang, L., and Singer, P.C. 2003. “Factors Influencing the Formation and Relative Distribution of Haloacetic Acids and Trihalomethanes in Drinking Water.” Environ. Sci. Technol., 37: 2920–2928. https://doi.org/10.1021/es026230q.

Liang, X., Lettenmaier, D.P., Wood, E.F., and Burges, S.J. 1994. “A Simple Hydrologically Based Model of Land Surface Water and Energy Fluxes for General Circulation Models.” J. Geophys. Res. Atmospheres, 99: 14415–14428. https://doi.org/10.1029/94JD00483.

Livneh, B., Bohn, T.J., Pierce, D.W., Munoz-Arriola, F., Nijssen, B., Vose, R., Cayan, D.R., and Brekke, L. 2015. “A Spatially Comprehensive, Hydrometeorological Data Set for Mexico, the U.S., and Southern Canada 1950-2013.” Sci. Data, 2: 150042. https://doi.org/10.1038/sdata.2015.42.

Lohmann, D., Nolte-Holube, R., and Raschke, E. 1996. "A Large-Scale Horizontal Routing Model to Be Coupled to Land Surface Parametrization Schemes.” Tellus A, 48: 708–721.

Luo, Y., Weiskopf, D., and Kirkpatrick, A.E. 2008. “Cluster Visualization in Parallel Coordinates Using Curve Bundles. IEEE Trans.” Vis. Comput. Graph., 12.

Maier, H.R., Morgan, N., and Chow, C.W.K. 2004. “Use of Artificial Neural Networks for Predicting Optimal Alum Doses and Treated Water Quality Parameters.” Environ. Model. Softw., 19:, 485–494. https://doi.org/10.1016/S1364-8152(03)00163-4.

Mather, A.L. 2015. Stream Turbidity Modeling: A Foundation for Water Quality Forecasting. Doctoral Dissertation, Oregon Health and Science University.

May-Ostendorp, P.T., Henze, G.P., Rajagopalan, B., Charles, D., Henze, G.P., Rajagopalan, B., and Charles, D. 2013. “Extraction of Supervisory Building Control Rules from Model Predictive Control of Windows in a Mixed Mode Building in a Mixed Mode Building.” Journal of Building Performance Simulation, 6(3): 199-219. https://doi.org/10.1080/19401493.2012.665481.

McCullagh, P., and Nelder, J.A. 1989. Generalized Linear Models. London: Chapman and Hall.

McTigue, N.E., Cornwell, D.A., Graf, K., and Brown, R. 2014. “Occurrence and Consequences of Increased Bromide in Drinking Water Sources.” J. Am. Water Works Assoc., 106: E492–E508. https://doi.org/10.5942/jawwa.2014.106.0141.


Momm, H.G., Wells, R.R., and Bennett, S.J. 2018. “Disaggregating Soil Erosion Processes within an Evolving Experimental Landscape.” Earth Surf. Process. Landf., 43: 543–552. https://doi.org/10.1002/esp.4268.

Moody, J.A., Martin, D.A., and Cannon, S.H. 2008. “Post-Wildfire Erosion Response in Two Geologic Terrains in the Western USA.” Geomorphology, 95: 103–118. https://doi.org/10.1016/j.geomorph.2007.05.011.

Morgan, R.P.C. 2009. Soil Erosion and Conservation. John Wiley & Sons.

Murphy, S.F., Barber, L.B., Verplank, P.L., and Kinner, D.A. 2003. “Environmental Setting and Hydrology of the Boulder Creek Watershed.” In Murphy, S.F., Verplank, P.L., Barber, L.B. (Eds.), Comprehensive Water Quality of the Boulder Creek Watershed, Colorado, during High-Flow and Low-Flow Conditions, 2000. U.S. Geological Survey.

Murphy, S.F., McCleskey, R.B., Martin, D.A., Writer, J.H., and Ebel, B.A. 2018. “Fire, Flood, and Drought: Extreme Climate Events Alter Flow Paths and Stream Chemistry.” J. Geophys. Res. Biogeosciences, 123: 2513–2526. https://doi.org/10.1029/2017JG004349.

Murphy, S.F., Writer, J.H., McCleskey, R.B., and Martin, D.A. 2015. “The Role of Precipitation Type, Intensity, and Spatial Distribution in Source Water Quality after Wildfire.” Environ. Res. Lett., 10: 084007. https://doi.org/10.1088/1748-9326/10/8/084007.

National Interagency Figure Center. 2018. National Report of Wildland fires and Acres Burned by State. National Interagency Figure Center.

National Research Council. 2012. Water Reuse: Potential for Expanding the Nation’s Water Supply Through Reuse of Municipal Wastewater. https://doi.org/10.17226/13303.

NOAA (National Oceanic and Atmospheric Administration). n.d. “Climate Data Online Search.” https://www.ncdc.noaa.gov/cdo-web/search;jsessionid=DF8A9CF0CBFF687E6987E5FCD1CF63FD.

NOAA (National Oceanic and Atmospheric Administration). 2019. “Divisional Data.” https://www7.ncdc.noaa.gov/CDO/CDODivisionalSelect.jsp.

Partnership for Safe Water. 2014. Guidelines for Phase IV: Application for the “Excellence in Water Treatment Award.” American Water Works Association.

Peter, D., and Silvia, P. 2012. ARIMA vs. ARIMAX Which Approach Is Better to Analyze and Forecast Macroeconomic Time Series? Silesian Univ Opava, School Business Administration Karvina, Karvina.

Raseman, W.J., Jacobson, J., and Kasprzyk, J.R. 2019. “Parasol: An Open Source, Interactive Parallel Coordinates Library for Multi-Objective Decision Making.” Environ. Model. Softw., 116:, 153–163. https://doi.org/10.1016/j.envsoft.2019.03.005.

Raseman, W.J., Kasprzyk, J.R., Rosario-Ortiz, F.L., Stewart, J.R., and Livneh, B. 2017. “Emerging Investigators Series: A Critical Review of Decision Support Systems for Water Treatment: Making the Case for Incorporating Climate Change and Climate Extremes.” Environ. Sci. Water Res. Technol., 3: 18–36. https://doi.org/10.1039/C6EW00121A.


Raseman, W.J., Rajagopalan, B., Kasprzyk, J.R., Kleiber, W., In Review. Nearest Neighbor Bootstrap for Generating Influent Time Series for Water Treatment. Stoch. Environ. Res. Risk Assess.

Reed, P.M., Hadka, D., Herman, J.D., Kasprzyk, J.R., and Kollat, J.B. 2013. “Evolutionary Multiobjective Optimization in Water Resources: The Past, Present and Future.” Adv. Water Resour., 51: 438–456.

Reneau, S.L., Katzman, D., Kuyumjian, G.A., Lavine, A., and Malmon, D.V. 2007. “Sediment Delivery after a Wildfire.” Geology, 35: 4. https://doi.org/10.1130/G23288A.1.

Rhoades, C.C., Entwistle, D., and Butler, D. 2011. “The Influence of Wildfire Extent and Severity on Streamwater Chemistry, Sediment and Temperature following the Hayman Fire, Colorado.” Int. J. Wildfire Sci., 20: 430–442.

Richardson, S.D., Plewa, M.J., Wagner, E.D., Schoeny, R., and Demarini, D.M. 2007. “Occurrence, Genotoxicity, and Carcinogenicity of Regulated and Emerging Disinfection By-products in Drinking Water: A Review and Roadmap for Research.” Mutat. Res., 636: 178–242. https://doi.org/10.1016/j.mrrev.2007.09.001.

Rietveld, L.C., van der Helm, A.W.C., van Schagen, K.M., and van der Aa, L.T.J. 2010. "Good Modelling Practice in Drinking Water Treatment, Applied to Weesperkarspel Plant of Waternet.” Environ. Model. Softw., Thematic Issue on Modelling and Automation of Water and Wastewater Treatment Processes, 25: 661–669. https://doi.org/10.1016/j.envsoft.2009.05.015.

Rossman, L.A., Clark, R.M., and Grayman, W.M., 1994. “Modeling Chlorine Residuals in Drinking-Water Distribution Systems.” J. Environ. Eng., 120: 803–820. https://doi.org/10.1061/(ASCE)0733-9372(1994)120:4(803).

Runkel, R.L., Crawford, C.G., and Cohn, T.A. 2004. Load Estimator (LOADEST): A FORTRAN Program for Estimating Constituent Loads in Streams and Rivers. USGS.

Samson, C.C. 2016. Modeling Relationships between Climate, Source Water Quality, and Disinfection Byproduct Formation and Speciation in Treated Drinking Water. PhD Dissertation, University of Colorado Boulder.

Samson, C.C., Rajagopalan, B., and Summers, R.S. 2016. “Modeling Source Water TOC Using Hydroclimate Variables and Local Polynomial Regression.” Environ. Sci. Technol., 50: 4413–4421. https://doi.org/10.1021/acs.est.6b00639.

Santín, C., Doerr, S.H., Otero, X.L., and Chafer, C.J. 2015. “Quantity, Composition and Water Contamination Potential of Ash Produced under Different Wildfire Severities.” Environ. Res., 142: 297–308. https://doi.org/10.1016/j.envres.2015.06.041.

Santos, F., Russell, D., and Berhe, A.A. 2016. “Thermal Alteration of Water Extractable Organic Matter in Climosequence Soils from the Sierra Nevada, California.” J. Geophys. Res. Biogeosciences, 121: 2877–2885. https://doi.org/10.1002/2016JG003597.

Santos, F., Wymore, A.S., Jackson, B.K., Sullivan, S.M.P., McDowell, W.H., and Berhe, A.A. 2019. “Fire Severity, Time Since Fire, and Site-Level Characteristics Influence Streamwater Chemistry at Baseflow Conditions in Catchments of the Sierra Nevada, California, USA.” Fire Ecol., 15: 3. https://doi.org/10.1186/s42408-018-0022-8.


Schindler, D.W., Curtis, P.J., Bayley, S.E., Parker, B.R., Beaty, K.G., and Stainton, M.P. 1997. “Climate-Induced Changes in the Dissolved Organic Carbon Budgets of Boreal Lakes.” Biogeochemistry, 36: 9–28. https://doi.org/10.1023/A:1005792014547.

Sexton, T. 2006. “U.S. Federal Fuel Management Programs: Reducing Risk to Communities and Increasing Ecosystem Resilience and Sustainability.” In Fuels Management-How to Measure Success: Conference Proceedings. 28-30 March 2006, Portland, OR.

Smith, R., Kasprzyk, J., and Dilling, L. 2017. “Participatory Framework for Assessment and Improvement of Tools (ParFAIT): Increasing the Impact and Relevance of Water Management Decision Support Research.” Environ. Model. Softw., 95: 432–446. https://doi.org/10.1016/j.envsoft.2017.05.004.

Sobol’, I.M., Tarantola, S., Gatelli, D., Kucherenko, S.S., and Mauntz, W. 2007. “Estimating the Approximation Error when Fixing Unessential Factors in Global Sensitivity Analysis.” Reliab. Eng. Syst. Saf., 92: 957–960. https://doi.org/10.1016/j.ress.2006.07.001.

Solarik, G., Summers, R.S., Sohn, J., Swanson, W.J., Chowdhury, Z.K., and Amy, G.L. 2000. “Extensions and Verification of the Water Treatment Plant Model for Disinfection By-Product Formation.” In Natural Organic Matter and Disinfection By-Products, ACS Symposium Series. American Chemical Society. https://doi.org/10.1021/bk-2000-0761.ch004.

Stewart, J. 2017. A Multi-Algorithm Approach to Suspended Sediment Modeling in the Colorado Front Range. Civ. Eng. Grad. Theses Diss. University of Colorado at Boulder.

Stewart, J., Rajagopalan, B., Kasprzyk, J., Raseman, W., and Livneh, B. 2017a. The Use of Ensemble Modeling of Suspended Sediment to Characterize Uncertainty. World Environmental and Water Resources Congress, May 21–25, Sacramento, California.

Stewart, J.R., Livneh, B., Kasprzyk, J.R., Rajagopalan, B., Minear, J.T., and Raseman, W.J. 2017b. “A Multialgorithm Approach to Land Surface Modeling of Suspended Sediment in the Colorado Front Range: Multi-Physics Sediment Modeling.” J. Adv. Model. Earth Syst., 9: 2526–2544. https://doi.org/10.1002/2017MS001120.

Suchetana, B., Rajagopalan, B., and Silverstein, J. 2019. "Modeling Risk Attributes of Wastewater Treatment Plant Violations of Total Ammonia Nitrogen Discharge Limits in the United States.” Stoch. Environ. Res. Risk Assess., 33: 879–889. https://doi.org/10.1007/s00477-019-01654-6.

Suchetana, B., Rajagopalan, B., and Silverstein, J. 2017. “Assessment of Wastewater Treatment Facility Compliance with Decreasing Ammonia Discharge Limits Using a Regression Tree Model.” Sci. Total Environ., 598: 249–257. https://doi.org/10.1016/j.scitotenv.2017.03.236.

Summers, R.S., Benz, M.A., Shukairy, H.M., and Cummings, L. 1993. “Effect of Separation Processes on the Formation of Brominated THMs.” J. Am. Water Works Assoc., 85: 88–95. https://doi.org/10.1002/j.1551-8833.1993.tb05925.x.

Tixier, A.J.-P., Hallowell, M.R., Rajagopalan, B., and Bowman, D. 2016. “Application of Machine Learning to Construction Injury Prediction.” Autom. Constr., 69: 102–114. https://doi.org/10.1016/j.autcon.2016.05.016.


Towler, E., Rajagopalan, B., Gilleland, E., Summers, R.S., Yates, D., and Katz, R.W. 2010. "Modeling Hydrologic and Water Quality Extremes in a Changing Climate: A Statistical Approach Based on Extreme Value Theory.” Water Resour. Res., 46: W11504. https://doi.org/10.1029/2009WR008876.

Toy, T.J., Foster, G.R., and Renard, K.G. 2002. Soil Erosion: Processes, Prediction, Measurement, and Control. John Wiley & Sons.

U.S. Census Bureau. 2018. “2017 National Population Projections Tables.” https://www.census.gov/data/tables/2017/demo/popproj/2017-summary-tables.html.

USGS (U.S. Geological Survey). 2019. “Science in Your Watershed.” https://water.usgs.gov/wsc/map_index.html.

Valois, R., Vouillamoz, J.-M., Lun, S., and Arnout, L. 2017. “Assessment of Water Resources to Support the Development of Irrigation in Northwest Cambodia: A Water Budget Approach.” Hydrol. Sci. J., 62: 1840–1855. https://doi.org/10.1080/02626667.2017.1351030.

Weirich, S.R., Silverstein, J., and Rajagopalan, B. 2015. “Simulation of Effluent Biological Oxygen Demand and Ammonia for Increasingly Decentralized Networks of Wastewater Treatment Facilities.” Environ. Eng. Sci., 32: 232–239. https://doi.org/10.1089/ees.2014.0407.

Weirich, S.R., Silverstein, J., and Rajagopalan, B. 2011. "Effect of Average Flow and Capacity Utilization on Effluent Water Quality from US Municipal Wastewater Treatment Facilities.” Water Res., 45: 4279–4286.

Westerling, A.L., Hidalgo, H.G., Cayan, D.R., and Swetnam, T.W. 2006. “Warming and Earlier Spring Increase Western U.S. Forest Wildfire Activity.” Science, 313: 940–943. https://doi.org/10.1126/science.1128834.

Wicks, J.M., and Bathurst, J.C. 1996. “SHESED: A Physically Based, Distributed Erosion and Sediment Yield Component for the SHE Hydrological Modelling System.” J. Hydrol., 175: 213–238. https://doi.org/10.1016/S0022-1694(96)80012-6.

Williams, J.R. 1975. “Sediment Routing for Agricultural Watersheds.” J. Am. Water Resour. Assoc., 11: 965–974. https://doi.org/10.1111/j.1752-1688.1975.tb01817.x.

Wischmeier, W.H., and Smith, D.D. 1960. "A Universal Soil-Loss Equation to Guide Conservation Farm Planning. Trans.” 7th Int Congr Soil Sci, 1: 418–425.

Worm, G.I.M., van der Helm, A.W.C., Lapikas, T., van Schagen, K.M., and Rietveld, L.C. 2010. “Integration of Models, Data Management, Interfaces and Training Support in a Drinking Water Treatment Plant Simulator.” Environ. Model. Softw., Thematic Issue on Modelling and Automation of Water and Wastewater Treatment Processes, 25: 677–683. https://doi.org/10.1016/j.envsoft.2009.05.011.

Worrall, F., Burt, T., and Shedden, R. 2003. “Long Term Records of Riverine Dissolved Organic Matter.” Biogeochemistry, 64: 165–178. https://doi.org/10.1023/A:1024924216148.

Writer, J.H., Hohner, A., Oropeza, J., Schmidt, A., Cawley, K., and Rosario-Ortiz, F.L. 2014. “Water Treatment Implications after the High Park Wildfire, Colorado.” J. Am. Water Works Assoc., 106: E189–E199. https://doi.org/10.5942/jawwa.2014.106.0055.


Wu, W., Maier, H.R., Dandy, G.C., Leonard, R., Bellette, K., Cuddy, S., and Maheepala, S. 2016. “Including Stakeholder Input in Formulating and Solving Real-World Optimisation Problems: Generic Framework and Case Study.” Environ. Model. Softw., 79: 197–213. https://doi.org/10.1016/j.envsoft.2016.02.012.

Zhang, K., Zargar, A., Achari, G., Islam, M.S., and Sadiq, R. 2014. “Application of Decision Support Systems in Water Management.” Environ. Rev., 22: 189–205. https://doi.org/10.1139/er-2013-0034.

1199 North Fairfax Street, Suite 900Alexandria, VA 22314-1445

6666 West Quincy AvenueDenver, CO 80235-3098

www.waterrf.org | [email protected]

advancing the science of water®

Date post:	17-Jun-2020
Category:	Documents
Upload:	others
View:	5 times
Download:	2 times

An Integrated Modeling and Decision Framework to Evaluate ... · An Integrated Modeling and...

Documents