streamflow hics presentation 021610 rkm

© 2008, Aptima, Inc. 1

www.aptima.comMA ▪ DC ▪ OH ▪ FL

© 2008, Aptima, Inc.

Accounting for Human Intervention in Streamflow Forecasting

Presentation to HICs Workshop

Silver Spring, MD

Andy Chang, Bob McCormack Lawrence Wolpert

February 16, 2010

© 2008, Aptima, Inc. 2

Agenda

Who are we?What do we do?What’s the issue?What’s the problem?What’s the solution?What are the next steps?Questions – as if these are not enough!

© 2008, Aptima, Inc. 3

Who are we?Aptima, Inc.

Founded in 1995; consistent annual growth (43% CAGR)Serving government and commercial clients300+ contracts with the DoDOffices– Woburn, MA, (HQ) – Washington, DC – Dayton, OH– Ft Walton Beach, FL

Interdisciplinary Small Business100+ staff (80% graduate degrees)

© 2008, Aptima, Inc. 4

What’s the issue?Streamflow Forecasting

NWS is responsible for providing weather, hydrologic, and climate forecasts and warnings – provides web-based Advanced Hydrologic Prediction Services (AHPS)

through 13 regionally-based River Forecast Centers– AHPS uses computer models that simulate river flow, rainfall, etc. to

generate predictions and these are conveyed to users

Predictions are filled with uncertainty, which is difficult to convey– Human behavior, which directly impacts the water levels and flood

conditions, remains one of the chief sources of uncertainty in hydrology forecasts

Current models do not explicitly account for behavior of the humans in the loop (forecasters, WRMs)– River regulation complicates streamflow forecasting due to lack of

human influences

© 2008, Aptima, Inc. 5© 2008, Aptima, Inc. 5© 2008, Aptima, Inc. 5

Visual Display Design

Support EMs’ decision making regarding courses of action.

– Visualizations of river forecasts with uncertainty – Visualizations of “impact” of various flooding scenarios

Visualization of uncertainty model inputs for clear understanding of “why” predictions are what they are.

Allow for local knowledge to be incorporated into predictions

Fusion / organization of information for clear understanding of relationship to one another and enhanced orientation.

6

What’s the problem?

7

What’s the solution?

© 2008, Aptima, Inc. 8

ObjectivesUnderstand Influences and Actions ofWater Resource Managers Develop Models which Account for

Human Behavior

Link Hydrology ForecastsTo Human Models

Incorporate Human Behavior ForecastsInto Hydrology Predictions

© 2008, Aptima, Inc. 9

What do we do?Human–Centered Engineering

TechnologyCapabilities

Social &Organizational

Structures

Mission, Tasks & Work

Processes

Human Agents

Congruence or ...

Disruption

SME Interviews

Northeast RFC – Taunton, MADevelopment and Operations Hydrologists (DOH Workshop in SS)US Army Corps of Engineers – Reservoir ControlFirst Light Power Resources – New Milford, CTCalifornia-Nevada RFC – Sacramento, CA– NWS, USACE, SMUD, PG&E, SFWATER, USBR, CA Water

© 2008, Aptima, Inc. 10

© 2008, Aptima, Inc. 11

[Decision Ladders]

© 2009, Aptima, Inc. 12

Decision Ladder Structure

© 2009, Aptima, Inc. 13

Decision Ladder Details

© 2009, Aptima, Inc. 14

Decision Ladder Details

© 2008, Aptima, Inc. 15

[Model]

Model Approach

© 2009, Aptima, Inc. 16

Hydro Models

Data

Environmental

Human

Forecasts

Human ModelFinite State Automata

Bayesian Inference

Visualization

© 2008, Aptima, Inc. 17

Initial Model: Finite State Automata

Bayesian Model

© 2008, Aptima, Inc. 18

Representation of information and influencesA Bayesian Network is a directed acyclic graph with an associated set of random variables

Joint probability distribution

We use the Bayesian Network to calculate the likely actions based on the observed information

),( EVB =

))(pa,|(),,( 11

ijxXxXPxXxXP

jjii

nn

∈==∏=== K

{ }vXVv∈

© 2006, Aptima, Inc. 19

•Prune Nodes•Adjust Conditional

Probabilities•Insert Known Prior

Distributions

Finite State Automata based on Decision Ladders

Hydro Forecasts

Integrated Model

“Adversarial” Modeling

How do we model the actions of “non-cooperative” water resource managers, i.e., those individuals who don’t share their operational plans, schedules, etc.?The Bayesian model can learn the conditional probability distributions based on observed data.

© 2008, Aptima, Inc. 20

Contact information

Lawrence Wolpert (PM)lwolpert@aptima.com617.913.5399

Bob McCormack (PI)rmccormack@aptima.com781.496.2476

Andy Changachang@aptima.com781.496.2447

© 2008, Aptima, Inc. 23

[ACE]

© 2008, Aptima, Inc. 24

© 2008, Aptima, Inc. 25

© 2008, Aptima, Inc. 26

© 2008, Aptima, Inc. 27

© 2008, Aptima, Inc. 28