many objective robust decision making for environmental ... · 8/31/2016 · many objective robust...

Kasprzyk Seminar | Slide 1

Many Objective Robust Decision Making for Environmental and Water

Systems Under Uncertainty

Water SeminarAugust 31, 2016

Joseph R. KasprzykAssistant Professor

Civil Environmental and Architectural Engineering DepartmentUniversity of Colorado Boulder


[Wall Street Journal, 8/11/16, http://www.wsj.com/articles/iea-sees-oil-demand-easing-as-economic-outlook-dims-1470902402; Fischer et al (2008) Resources for the Future Discussion Paper, http://www.rff.org/files/sharepoint/WorkImages/Download/RFF-DP-07-54.pdf]

Challenge: Scenarios that are used to inform decision-making might be wrong!

Proposal: New “bottom-up” decision making methods can help explore

alternate scenarios

http://www.wsj.com/articles/iea-sees-oil-demand-easing-as-economic-outlook-dims-1470902402

http://www.rff.org/files/sharepoint/WorkImages/Download/RFF-DP-07-54.pdf


source: http://www.cjwalsh.ie/tag/greater-dublin-strategic-drainage-study

source: http://www.swc.nd.gov

Challenge: Creating planning alternatives by hand ignores the potential to improve system performance

Proposal: Coupled simulation-optimization technique to generate new planning alternatives

[Example courtesy Rebecca Smith]


MultiobjectiveEvolutionary

Algorithm (MOEA) Search

f1 f2 f3

Objectivesmeasurements of

system performance

Simulation Model

Constraintslimits of acceptable

performance

f1 < n1 , f2 < n2 , f3 > n3

Input Scenarios

x1 x2 x3

Decision Levers management and

infrastructure options

Model Outputs

Legend:Typical model runAutomation using MOEA

Generates new alternatives

Use performance to evaluate alternatives


Some of these challenges still exist today:

• At that time, optimization was not often done for real-world planning projects

• Alternatives are not generated by a computer, though they are systematically evaluated

• Too little focus on existing systems

• Models are not easily understood

Proposal: Better communication, more transparent modeling (e.g. CADSWES), and studies that engage stakeholders

[Rogers and Fiering (1986), Water Resources Research, http://dx.doi.org/10.1029/WR022i09Sp0146S]

http://dx.doi.org/10.1029/WR022i09Sp0146S


My group’s research seeks to develop tradeoffs for problems

with conflicting objectives to increase understanding;

creating innovative solutions to these problems; and

characterizing uncertainty.


Example Problem

How to fairly balance allocations of water?Decisions: Reservoir balancing rules, conservation triggers, allocationsObjectives: Reliability of meeting supply targets, minimum flow requirements, minimizing pumpingConstraints: Regulatory limits, preserving mass balance

[Maier et al., In-Prep, Env. Mod. Soft]


Key Points

• Bottom-up decision making inverts the steps of the traditional decision aiding process, and can be helpful for problems that exhibit deep uncertainty

• Many Objective Robust Decision Making can be used for the generation of robust planning alternatives for water and environmental systems

• These approaches can be used to help solve multiple problem types such as providing drinking water treatment under climate extremes


“Bottom-up” decision frameworks

(Decision Scaling, Info-Gap, RDM, MORDM, Dynamic Adaptive Policy Pathways …)

[Courtesy Jon Herman, UC Davis]


Evaluate alternatives in multiple states of the world…

Quantify robustness measures and determine sensitive uncertainties

What do robustness analyses have in common?

[Courtesy Jon Herman, UC Davis]

Kasprzyk Seminar | Slide 11[Herman et al, 2015, Journal of Water Resources Planning and Management, http://dx.doi.org/10.1061/(ASCE)WR.1943-5452.0000509]

http://dx.doi.org/10.1061/(ASCE)WR.1943-5452.0000509


Many-Objective Robust Decision Making for Complex Environmental Systems Undergoing ChangeJoseph R Kasprzyk, Shanti Nataraj, Patrick M Reed, Robert J Lempert

Environmental Modelling and Software


Lower Rio Grande Valley (LRGV) faces rising demands with variable supply.

185

210

235

260

285

310

335

360

385

An

nu

al U

sa

ge

(x 1

0 m

)

63

1965 1969 1973 1977 1981 1985 1989 1993 1997

Year

185

210

235

260

285

310

335

360

385

An

nu

al U

sa

ge

(x 1

0 m

)

63

1965 1969 1973 1977 1981 1985 1989 1993 1997

Year

Supply

Demand

When demand >

supply, possible

shortfall!

[Example data courtesy G. Characklis]

• Rapid population growth and high irrigation water use

• Existing water market with transfers from ag to urban

• How can a single city use a water market to increase the reliability of its water supply?


What is the outcome of adding reservoir rights to meet supply?

• Each point: a volume of reservoir rights

• Non-domination (i.e., highest reliability at each cost level)

• Shows increasing cost of providing reliability

1.000 0.995 0.990 0.985 0.9809

10

11

12

13

Co

st

Reliability

Preferred

Region

Dominated

Region

Can the market help lower costs?

What other objectives are

important for planning?[Kasprzyk et al., 2009, WRR]


A many-objective approach to the LRGV helps answer these questions.• Portfolio of 3 instruments

—Permanent rights: non-market supply, % of reservoir inflows

—Spot leases: immediate transfers of water, variable price

—Adaptive options contract: reduces lease-price volatility

• Monte Carlo simulation model considers natural variability

—Sampling of historical data for hydrology, demands, lease pricing

• Use a Multiobjective Evolutionary Algorithm —Up to 6 objectives calculated using expected values under

10-year planning horizon

[Kasprzyk et al., 2009, WRR]

Problem

Formulation

Generating

Alternatives


Visualize rights (color), leases (orientation), options (size)

Two distinct groups of solutions:

1. rights-dominated

2. market use

Over-reliance on traditional water supply raised costs and surplus water volumes!

Many-Objective Results

1

2

[Kasprzyk et al., 2012, Env. Mod. Soft.]


All objectives used a single distribution of input data to calculate expectation

Issue: Is our choice of solution biased by assumptions of input data?

Challenge: Deep Uncertainty, where decision makers can’t characterize full set of events or probabilities

Our selected solutions were based on expected-value objective calculations.

[Kasprzyk et al., 2012, Env. Mod. Soft.]


Many-Objective Robust Decision Making (MORDM)

Which solutions do well under a large number of deeply uncertain trajectories?

How do we characterize values of the uncertainties that cause vulnerabilities for those

robust solutions?


Scaling factors modify input data.

• Baseline historical data

• Values exceeding highest/lowest 25% of data scaled N times likelier

• Scaling factors unique to each variable, sampled as point values How wrong do we have to be to

cause performance failures?


Uncertainty ensemble

• State of the world (SOW): a value for each of these dimensions

• A SOW controls howinput data is sampled within the Monte Carlo simulation

Parameter Lower Bound Upper Bound

Initial Rights 0.0 0.4

Demand Growth Rate

1.1% 2.3%

Initial Reservoir Volume

987 mill. m3 2714 mill. m3

Table 2: Scalar Model Parameters

Parameter Lower Bound

Upper Bound

Low Inflows 1 10

High Losses 1 10

High Demands 1 10

High Lease Prices 1 10

Losses in Storage 1 10

Table 1: Scaling Factors


Evaluating robustness

• Apply ensemble of 10,000 LHS samples of uncertainties (SOWs) to each solution

• Sort values and calculate:—10th percentile (for measures to be

maximized)

—90th percentile (for measures to be minimized)

• Percent deviation :

Cost Distribution for Solution N:

36

37

39

40

44

45

48

50

51

55

Cost in

baseline SOW

Cost in 90th

percentile SOW

c90- cbase

cbasex100 =

51-37

37x100 = 37.8%

Solution X costs 37.8% more in the 90th percentile of the ensemble than it did under the baseline SOW.

We now visualize “percent deviation” across all solutions and measures.


Percent deviation shows us robustness of the tradeoff space.

LegendSize: Critical ReliabilityOrientation: Dropped TransfersAxes: Measures in baseline SOWColor: % Deviation

Solution 4 exhibits low deviation in critical reliability and cost.

It comes from a different tradeoff region than Solution 1-3.


Scenario Discovery

• Patient Rule Induction Method (PRIM) is an interactive algorithm for discovering scenarios

—Instead of specifying scenario values a priori, the discovered ranges are clearly linked to policy vulnerabilities.


Incorporating Uncertain Factors Into the Many-Objective Optimization ProcessAbby Watson, Joseph Kasprzyk

Environmental Modelling and Software (In Revision)


Review of robust optimization approaches

Method Benefits Limitations

Optimize under baseline conditions, robustness assessment afterwards (prior MORDM study)

Solutions do well in baseline conditions

No robustness information in optimization itself

Include robustness objective in objective set [e.g. Hamarat et al 2013, http://dx.doi.org/10.1016/j.techfore.2012.10.004/]

Can focus on “maximizing” robustness and see tradeoffs between robustness and other measures

Difficult to calculate robustness for multiple objectives

Optimize under baseline conditions and on scenarios discovered through MORDM process (this study)

Can focus on how optimal strategieschange under new scenarios

Limited in the number of scenarios that you can explore

http://dx.doi.org/10.1016/j.techfore.2012.10.004/


Selecting scenarios from MORDM results

• Select values of scaling factors as a new scenario for optimization

• Scenario factors perturb the Monte Carlo simulations

• Compare the tradeoff sets


Scenarios

Scenario Scaling Factors Description

1: Baseline scenario All scaling factors = 1 Baseline conditions

2: Moderate scenario All scaling factors = 2Moderate increase in

extremes

3: Cost scenarioLow inflow scaling = 4High losses scaling = 2High demands scaling = 4

High cost of robust

solution

4: Reliability scenarioLow inflow scaling = 7High losses scaling = 2High demands scaling = 4

Low reliability of

robust solution

5: Market use

scenario

Low inflow scaling = 8High losses scaling = 4High demands scaling = 7

High market use of

robust solution


Higher values of decision levers in red and green sets indicates a more conservative portfolio under severe conditions

Relationship between planning objectives is transformed depending on the scenario used.


Ongoing Projects

Kasprzyk Seminar | Slide 31[Smith et al., In-Prep]

Participatory Tool Evaluation: working with a group of practitioners to investigate the relevance of an emerging research tool through a collaborative testbed

WRSO: Water Resource Systems Optimization

Kasprzyk Seminar | Slide 32[Smith et al., In-Prep]

Participatory Tool Evaluation: working with a group of practitioners to investigate the relevance of an emerging research tool through a collaborative testbed

WRSO: Water Resource Systems Optimization


A workshop with 6 water utilities in the Front Range of Colorado showed the most important planning concerns.

[Smith et al., In-Prep]


Coupling of natural and human systems related to Potable Water Systems (PWS).

Human Systems

Natural Systems

Source Watersheds

Water Treatment

Plants

Temperature,Precipitation

Land SurfaceCharacteristics

Perturbations• Floods• Droughts• Wildfires

PWS Operators

Advocacy Coalitions• Customers/Ratepayers• Stakeholder GroupsRegulatory Context

Feedback: Risk Perceptions

Feedback: Watershed Management

Adaptive decisions about infrastructure and watershed management

Improved PWS resilience to changes in water quality

Expected Outcomes

Improved scientific understanding of climate and disturbance risks to source water quality

DrinkingWater

Quality

Source Waters

Related project funded by EPA in collaboration with Water Research Foundation. With Balaji, Livneh, Rosario-Ortiz, Summers


Recommendations:• Better represent needs of Water Treatment Plants (WTPs)

including focus on existing systems• Explicitly handle uncertainty• Incorporate nonstationarity, especially related to extreme

weather events and climate change• Represent tradeoffs between competing objectives• Use standardized terminology to improve dissemination

A critical review of decision support systems for water treatment: Making the case for incorporating climate change and climate extremes

In revision at Environmental Science: Water Research and Technology.

By: William Raseman, Joseph Kasprzyk, Fernando-Rosario-Ortiz, Jenna Stewart, Ben Livneh


Setback distance regulations integrate concerns about ‘regional effects’ of oil and gas development into a coherent policy.

• Our work: provide information about potential impacts of setback regulations

• Approach—Create candidate alternatives for

setback limitation

—Use air quality modeling to estimate effects

—Tradeoff analysis of policy outcomes

[Alongi et al, AGU 2016]


Initial tradeoff analysis: Case study location

The red points are approximate locations of the homes and blue points are randomly generated well locations.


Initial Tradeoff Analysis: Illustrative Results

• How will regulations affect the number of potential drilling sites?

• Relative effect of setback distance or ‘density’ rules?• What is the relationship between the magnitude of pollution

and the frequency of pollution violations?


Key Points

• Bottom-up decision making inverts the steps of the traditional decision aiding process, and can be helpful for problems that exhibit deep uncertainty

• Many Objective Robust Decision Making can be used for the generation of robust planning alternatives for water and environmental systems

• These approaches can be used to help solve multiple problem types such as providing drinking water treatment under climate extremes


About your seminar speaker: Joseph Kasprzyk

• Pronounciation: Kas – pry – zick

• Office in SEEC C244• Graduated from Penn State

University BS (2007) MS (2009) PhD (2013)

• Assistant Professor at University of Colorado since Fall 2013


Courses

• Engineering Hydrology (Fall, Undergraduate, CVEN 4333)

—Hydrologic processes—Unit hydrograph, routing, flood frequency

analysis

• Water Resources Engineering (Fall, Undergrad./Grad., CVEN 4323/5423)

—Rainfall runoff modeling, basic open channel flow modeling

—Flood control—Stormwater—Extensive use of Python programming and HEC

computer models

• Water Resources Systems Analysis(Spring, Grad. CVEN 5393)

—Water resources planning and management—Reservoir operations and modeling in RiverWare

(with Prof. Zagona and CADSWES)—Simulation modeling and optimization analysis

CVEN 5393 is open to students in other majors

(e.g., geography). Come and see me if you’re interested!


Acknowledgements

• Current group members: PhD students: Rebecca Smith, Billy Raseman, Melissa Estep; MS student: Matt Alongi

• Former group members: Amy Piscopo (PhD, now at EPA research and development), Elizabeth Houle (MS, now at Lynker Consulting), Abby Watson (MS, now at Riverside Technology), Tim Clarkin (MS, now at Army Corps of Engineers)

• NOAA SARP Project—Rebecca Smith; Edie Zagona (CADSWES); Lisa Dilling,

Kristen Averyt (CU Western Water Assessment)

• Oil and Gas—Matt Alongi, Joe Ryan, Jana Milford

• Climate Change and Water Quality—Fernando Rosario-Ortiz, Katie Dickinson, Ben Livneh,

Rajagopalan Balaji, R Scott Summers, Deserai Crow, Angie Bielefeldt (CU); Elizabeth Albright (Duke); William Becker (Columbia, Hazen and Sawyer); Kenan Ozekin (Water Research Foundation)

Thanks! Questions?

Funding:Abby Watson was funded by the Integrated Teaching and Learning Program under NSF GK-12 grant DGE-0338326;Matt Alongi was funded under NSF grant CBET-1240584; Billy Raseman was funded under EPA grant R835865; Rebecca Smith was funded under NOAA grant NA14OAR4310251.

However, these contents do not necessarily represent the policies of the NSF, NOAA, or the EPA and you should not assume endorsement by the US federal government.

many objective robust decision making for environmental ... · 8/31/2016 · many objective robust...

Documents