INCREASING GRID RESILIENCE THROUGH DATA-DRIVEN MODELING FOR STORM - - PowerPoint PPT Presentation

INCREASING GRID RESILIENCE THROUGH DATA-DRIVEN MODELING FOR STORM OUTAGE PREDICTION AND LONG-TERM PLANNING

Steven Quiring, Texas A&M University Seth Guikema, Johns Hopkins University U.S. DOE State Energy Risk Assessment Workshop

Our Overarching Goals

¨ Provide useful information on weather-related

impacts to aid

¨ (1) pre-storm preparation and ¨ (2) long-term hardening and mitigation by:

¤ Electric power utilities ¤ Other utilities and organizations dependent on electric

power

¤ Local, state, and federal agencies

Power Outage Prediction

¨ Past work on hurricane power outage forecasting ¨ Ongoing work for DOE’s Energy Infrastructure

Modeling and Analysis Division

¨ Demonstration of power outage prediction tool ¨ Gaps and future research

A “standard” utility response cycle

Hurricane Progress Utility Activities & Decisions Time

T=0 T~-5 days T~-2-3 days T = 0-1 day

Off-coast,

• uter bands

begin contact Rain begins for large storms, wind increases

T > 1 day

Strongest winds, heaviest rain, highest surge Wind, rain, surge gradually diminish Monitor track, begin estimating crew needs, coordination call, begin crew requests Finalize crew requests, position materials and crews Begin repairs when safe Conduct repairs according to prioritization plan

Timescales of Focus

Storm response planning Seasonal resource planning Long-term asset management

• How many hurricanes?
• How intense?
• How big?
• Which basins?
• How will climate

change influence hurricane hazards?

• Will risk of power
• utages change?
• Are different asset

management strategies needed?

ctpost.com Ameren America

Primary Research team: Roshi Nateghi, Seth Guikema, Allison Reilly, Andrea Staid, Michael Gao (JHU), Steven Quiring (TAMU)

Power Outage Prediction: Past Work

Past Work: Goals & Data

¨ Goal: Accurately estimate power outages 4-6 days

before landfall and update every 6 hours

¨ Unit of Analysis:

¤ Utility-specific model: 12,000 ft. by 8,000 ft. grid cells ¤ Spatially general model: census tracts

¨ Data:

¤ Hurricane wind speeds & duration (wind field model) ¤ Geographic data: LU/LC, soil type, topography, DEM, etc. ¤ Climatological: soil moisture, drought indices, long-term

precipitation levels

¤ Utility-specific: system inventory, tree-trimming

Selected Prior Work

¨ Outages

¤ Liu et al. (2005): a first model – GLM ¤ Han et al. (2009a, 2009b): improved accuracy, usability ¤ Guikema & Quiring (2012): further improved accuracy with a new

statistical method

¤ Nateghi et al. (2013): substantial increase in accuracy with a Random

Forest

¤ Guikema et al. (2014): model useable anywhere along the U.S.

coastline with only publicly-available data needed as input

¨ Outage Duration

¤ Liu et al. (2007): a first model ¤ Nateghi et al. (2011): improved predictive accuracy ¤ Nateghi et al. (2014): simplified model w/o losing predictive accuracy

¨ Damage

¤ Guikema et al. (2010): Comparison of statistical models ¤ Han et al. (2014): Bayesian updating of FORM-based fragility curves

Model Development Process

10 hurricanes, 4-state area In-house Willougby-based model Have run model for 12(+) real storms Use scenario storms to (a) answer policy-related queries and (b) gain further confidence in the model

Prediction Process

Hurricane track & intensity forecast (ensemble or single) Hurricane Wind Field Model Statistical Outage Forecasting Model Iterative updating at every 6 hour hurricane forecast update

Example Predictions: Early models, Katrina

Actual Number

• f Outages

GLM Predictions (Han et al. 2009a) GAM Predictions (Han et al. 2009b)

GLM (2009a) and GAM (2009b) with approximately 100 covariates.

A “simpler”, accurate model

Random forest model, reduced to 6 covariates (Nateghi et al., Risk Analysis 2013)

A Key Challenge:

All of the above models were specific to a utility service area and required privately held data.

Spatial Generalization

Can a model be developed that can be used for entire coast using only publicly available data while still maintaining accuracy? Approach

Train & validate models in service area using

• nly public data

Validate models across hurricanes Validate models across states Predict for an approaching hurricane

Learn from storms and refine models

Experience with Hurricane Sandy

First Model Run: Oct 26, 5pm

Our Forecast: 10 million without power First Press Release by JHU

Progression of Hurricane Sandy Runs During the Event

So How Did We Do?

¨ Important distinction: We predict cumulative outages,

utilities generally report peak outages

¨ Data Gap: Outage data at the scale at which we are

making predictions

¨ Results:

¤ DOE estimated 8.5 million customers were out at peak ¤ Our final estimate as the storm transited the mid-Atlantic

was 8-10 million out

¤ We were within 8% of DOE’s estimates for NY, PA, MA, RI,

VA

¤ We overestimated outages for MD and DE ¤ We underestimated outages for CT

“What If” Scenario Analysis: Historic Storms

Ran two sets of scenarios assuming current population distribution:

¤ Historic storms (Isabel,

Rita, Ivan, Camille, Ike, Katrina, Irene, Andrew, Galveston, Sandy)

¤ Haiyan on tracks of

Andrew, Katrina, Ike, and Irene

Storm& Predicted&Population& Without&Power&Relative& to&Hurricane&Irene& Isabel' 0.33' Rita' 0.58' Ivan' 0.64' Camille' 0.68' Ike' 0.77' Katrina' 1.00' Irene' 1.00' Andrew' 1.18' Galveston' 1.26' HaiyanCAndrew'track' 1.61' HaiyanCKatrina'track' 1.86' HaiyanCIke'track' 2.03' Sandy' 2.90' HaiyanCIrene'track' 6.92' '

Results

What If Analysis Results: Historic Storms

What If Runs: Haiyan

Long-Term Risk to Power Systems in a Changing Climate

Research Questions

¨ How would potential changes in hurricane hazards –

intensity, frequency, location – influence power system risk?

¨ Which areas of U.S. coastline are most sensitive to

changes in hurricane hazards?

¨ Can the possible changes be simulated in a way

that will help support long-term utility hardening decision-making?

Existing Hurricane Climatology

Influence of Changes in Frequency

Metropolitan Area Impacts: New York City vs. Washington, DC

Ongoing Work for DOE

Year 1 Research Tasks

¨ Power outage forecasting for weather events

¤ Uncertainty quantification ¤ Assessing the importance of soil moisture and tree cover

data for improving predictive accuracy

¤ Developing a secure web portal

¨ Scalable risk and uncertainty modeling to support long-

term planning

¤ Incorporate storm forecast uncertainty into the multi-scale

model

¤ Make the model applicable for multi-utility and ISO

coordinating areas

¨ Engaging and educating stakeholders

Model forecast tracks for Hurricane Gustav on Aug 25th (left) and Aug 29th (right), 2008. Plot provided courtesy of Jonathan Vigh, Colorado State University. For more information about this graphic see http://euler.atmos.colostate.edu/~vigh/guidance/

GPCE = is the average distance between the model-predicted tropical cyclone tracks from the Geophysical Fluid Dynamics Laboratory (GFDI) Hurricane Prediction System, Global Forecast System (AVNI), Navy Operational Global Atmospheric Prediction System (NGPI), Met Office Global Model (UKMI) and the GFDL model (GFNI) run at the Fleet Numerical Meteorology and Oceanography Center

Uncertainty Quantification

Monte Carlo Wind Speed Probability (MCWSP) Model

Left - the first 10 of 1000 track realizations for a Hurricane Ike forecast starting at 1200 UTC 7 Sep 2008. The white line is the National Hurricane Center official track forecast. Right – Hurricane force wind speed probabilities from the National Hurricane Center operational product for the same time.

MCWSP generates 1,000 forecast realizations by sampling from track and intensity forecast errors from the last 5 years and determines the wind radii of each realization using a simple climatology and persistence scheme

Uncertainty quantification: Hurricane Katrina

Five best-case scenarios (fewest predicted outages) and five worst-case scenarios (most predicted outages) for Hurricane Katrina on 28 August 2005 at 7 a.m. CDT, which is approximately 24 hours prior to landfall.

Best Case Scenarios Worst Case Scenarios

TD TS Cat1 Cat 2 Cat 3 Cat 4 Cat 5

Hurricane Katrina

Quiring, S. M., Schumacher, A. B. and S. D. Guikema (2014) Incorporating hurricane forecast uncertainty information into decision support applications. Bulletin of the American Meteorological Society, 95: 47-58

Demonstration of web portal

Web ¡portal ¡func/onality: ¡

• 1. Takes ¡as ¡input ¡a ¡track ¡and ¡intensity ¡from ¡manual ¡

upload, ¡text ¡entry, ¡or ¡download ¡directly ¡from ¡a-­‑ deck ¡file ¡at ¡UCAR ¡

• 2. Runs ¡the ¡wind ¡model ¡
• 3. Runs ¡the ¡power ¡outage ¡forecast ¡model ¡
• 4. Displays ¡the ¡map ¡of ¡forecast ¡outage ¡together ¡

with ¡outage ¡total ¡ ¡

Front Page

Track & Intensity Entry (Manual)

Displaying the Results – Ike

Displaying the Results – strong storm

Summary of Needs and Gaps

¨ Utility specific data from past storms (train and

validate our models in different areas)

¨ Reliable multi-utility data from past storms ¨ Enhance the outage modeling for other types of

weather events: thunderstorms, straight-line winds, ice storms, etc.

¨ Multi-scale approaches that provide risk

assessments at the scale of interest for different groups of stakeholders

Acknowledgements

Research Team

¨ Laiyin Zhu ¨ Tak Igusa ¨ Jared Beekman ¨ Michael Gao ¨ Lucas Henneman ¨ Roshi Nateghi ¨ Steven Quiring ¨ Allison Reilly ¨ Andrea Staid ¨ Gina Tonn ¨ Brent McRoberts

Funding Sources

¨ A large investor-owned

utility in the Central Gulf Coast region

¨ DOE – BER (Integrated

Assessment) and Energy Infrastructure Modeling and Analysis Office

¨ NSF – CMMI (IMEE, CIS,

and SEES)