Transcript why risk analysis?

Risk Assessment
Vicki M. Bier
(University of WisconsinMadison)
Introduction
• Risk assessment is a means to
characterize and reduce uncertainty to
support our ability to deal with
catastrophe
• Modern risk assessment for engineered
systems began with the Reactor Safety
Study (1975):
– Applications to engineered systems and
infrastructure are common
What is Risk Assessment?
• “A systematic approach to organizing
and analyzing scientific knowledge and
information for potentially hazardous
activities or for substances that might
pose risks under specified
circumstances”
– National Research Council (NRC), 1994
Definitions of Risk
• “Both uncertainty and some kind of loss or
damage” (Kaplan and Garrick 1981)
• “The potential for realization of unwanted,
negative consequences of an event” (Rowe
1976)
• “The probability per unit time of the
occurrence of a unit cost burden” (Sage and
White 1980)
• “The likelihood that a vulnerability will be
exploited” (NRC 2002)
Paradigm for Risk Assessment
• A form of systems analysis
• Answers three questions (Kaplan and
Garrick 1981):
– “What can go wrong?”
– “How likely is it that that will happen?”
– “If it does happen, what are the consequences?”
What is Probabilistic Risk Assessment?
• An integrated model of the response of an
engineered system to disturbances during
operations
• A rigorous and systematic identification of the
levels of damage that could conceivably result
from those responses
• A probabilistic (that is, quantitative) assessment
of the frequency of such occurrences and our
uncertainty in that assessment
• A tool to help owners/operators make good
decisions about system operations
ESSENCE OF PRA
• A PRA is an assessment of how well a system responds
to a variety of situations
• It answers three basic questions:
1. What can go wrong during operation?
2. How likely is it to go wrong?
3. What are the consequences when it goes wrong?
• We answer the first question in terms of scenarios
• We answer the second by quantifying our knowledge of
the likelihood of each scenario
• We answer the third by quantifying our knowledge of the
response of the system and its operators in terms of:
- damage states
- release states and source terms
- scenario consequences
GRAPHICAL PRESENTATION OF RISK
SCENARIO
PROBABILITY
DAMAGE
s1
s2
s3
.
p1
p2
p3
.
.
.
pN-1
pN
x1
x2
x3
.
.
.
xN-1
xN
.
.
sN-1
sN
CUMULATIVE
PROBABILITY
P1=p2+p1
.
.
.
.
.
PN-1=PN+pN-1
PN=pN
P
RISK CURVE
p(>x)
X
STRUCTURE OF THE MODERN PRA MODEL
LEVEL
3
2
1
INITIATING
EVENTS
PLANT (ACTIVE
SYSTEMS) MODEL
SUPPORT
SYSTEMS MODEL
SUPPORT
SYSTEM
STATES
PLANT
DAMAGE
STATES
CONTAINMENT
STRENGTH AND
CORE DAMAGE
PROGRESSION
MODEL
FRONTLINE
SYSTEMS – EARLY
RESPONSE
MODEL
OFFSITE
RADIOACTIVE
RELEASE MATERIAL
CATEGORIES DISPERSION AND
HEALTH IMPACT
MODEL
SUBTREE
FREQUENCIES
FRONTLINE
SYSTEMS – LATE
AND
CONTAINMENT
SAFETY
FEATURES
RESPONSE
MODEL
RISK BY
HEALTH
EFFECT
TYPE
QUANTIFYING SCENARIOS
INITIATING
EVENT
x
A
B
C
D
NODE B1
f(A | I )
NODE A
f (B | I A )
NODE C3
1  f(A | I )
S I A B CD
 (S)   (I)f ( A | I)f (B) | IA )f (C | IA B)f (D | IA BC)
IA BCD  S
EVENT SEQUENCE QUANTIFICATION
 (S)   (I)f ( A | I)f (B) | IA )f (C | IA B)f (D | IA BC)
WHERE
 (S )= the frequency of scenario S
 (I)= the frequency of initiating event I
f ( A | I)= the fraction of times system A
succeeds given that I has
INITIATING
happened
EVENT
A
B
C
D
=
the
fraction
of
times
system
B
f (B ) | IA )
1
NODE B1
fails given that I has happened
and A has succeeded
f (C | IA B)= the fraction of times C succeeds
given that I has happened, that
A has succeeded, and B has failed
SIMPLIFIED EVENT TREE DIAGRAM
f (D | IA BC)= the fraction of times D fails given
SCENARIO
IA BCD  S
STAGES TO EVENT TREE LINKING
LATE
FRONTLINE
SYSTEMS
ELECTRIC
POWER
SYSTEMS
OTHER
SUPPORT
SYSTEMS
PLANT
DAMAGE
STATES
EARLY
FRONTLINE
SYSTEMS
INITATING
EVENTS
AFW
SCOPING
REQUIREMENTS
TANK
PUMP
1
PUMP
2
PUMP
3
RELATIONSHIP OF FAULT TREES TO
EVENT TREES
INITIAL
CONDITIONS
STAGE A TOP
EVENTS
DAMAGE
STATE
OK
PLS
LOC/V
AFW
PLS
LOC/V
LEGEND
= “OR” GATE
TANK
ISOLATION
VALVE 1
ISOLATION
VALVE 2
APU
MODULE
GGVM
COOLING 1
= “AND” GATE
GGVM
COOLING 2
FAULT TREES AND EVENT TREES
• Both useful
• Event trees used to display order of events
and dependent events
• Fault trees used to display combinations of
events:
– Order and dependencies are obscured
• Logically equivalent
RISK MANAGEMENT
•
•
•
•
Develop an integrated plant-specific risk model
Rank order contributors to risk by damage index
Decompose contributors into specific elements
Identify options, such as design and procedure changes,
for reducing the impact of the contributor on risk
• Make the appropriate changes in the risk model:
– And re-compute the risk for each option
• Compute the cost impacts of each system configuration,
relative to the base case:
– Including both initial and annual costs
• Present the costs, risks, and benefits for each option
RISK DECOMPOSITION
(ANATOMY OF RISK)
LEVEL OF DMAGE
TYPE OF RELEASE
P
TYPE OF PLANT DAMAGE
CDF
INITIATING EVENT
CDF
MIEs
PDS

P
P

X
EVENT SEQUENCE
IE
A
B
P

SYSTEM UNAVAILABILITY
C
FAILURE CAUSES
System B Cause Table
B

INPUT DATA
1. Initiating events
2. Components
CAUSES
LOGIC
FREQUENCIES
1
MAJOR
SYSTEM
2
3
4
n
EFFECTS
3. Maintenance
4. Human error
5. Common cause
6. Environmental
DOMINANT
SEQUENCE
DOMINANT
FAILURE
MODES
7. Other
REACTOR TRIP SYSTEM CAUSE TABLE
CONTRIBUTORS TO SYSTEM FAILURE FREQUENCY
CAUSE
FREQUENCY--FAILURES
PER 10,000 DEMANDS
 Common cause failures of
reactor trip breakers
5.1
 Multiple independent
failures of reactor trip
breakers
0.39
 Reactor trip system in test
mode and one breaker fails
0.032
TOTAL
5.5
This analysis was performed in November 1982
(occurred at Salem in
February 1983)
SUCCESSFUL RISK MANAGEMENT
A FEW EXAMPLES DUE TO PLG STUDIES
DESCRIPTION
APPROXIMATE BENEFIT
PRA identified that interaction of two buildings during
an earthquake dominated the risk of an operating plant.
Installing rubber bumpers between the buildings
eliminated the problem.
Factor of 10 reduction in core damage frequency.
PRA allowed the utility to justify installation of a nonsafety grade AFW pump sharing common lines, instead
of the usual safety grade post-TMI requirement.
Core damage frequency reduction, and millions of
dollars.
PRA identified station blackout as the major contributor
to core damage frequency. It also identified a procedure
change to direct operators to manually cross-connect
like buses from the adjacent unit.
33% reduction in core damage frequency.
The PRA identified a peculiarity in the AC power supply
in which the three so-called redundant, independent
fuel-oil transfer pumps to the emergency dieselgenerators were not independent at all. One pump
actually depended on the operation of the other two
diesels. A simple correction of the power supply logic
fixed the problem.
Factor of 50 reduction in core damage frequency.
PRA study showed that risk to population beyond two
miles did not depend on evacuation. Recommended
reduction in EPZ.
Reduction of EPZ from 10 to one or two miles
considered by NRC.
Data Analysis
• Input parameters are quantified from available data:
– Typically using expert judgment and Bayesian statistics
– Due to sparseness of directly relevant data
• Hierarchical (“two-stage”) Bayesian methods common:
– Partially relevant data used to help construct prior distributions
• Numerous areas in which improvements can be made:
– Treatment of probabilistic dependence
– Reliance on subjective prior distributions
– Treatment of model uncertainty
Dependencies
• The failure rates (or probabilities) of components can be
uncertain and dependent on each other:
– For example, learning that one component had a higher failure
rate than expected may cause one to increase one’s estimates
of the failure rates of other similar components
• Failure to take such dependence into account can result
in substantial underestimation of the uncertainty about
the overall system failure rate:
– And also the mean failure probability of the system
• Historically, dependencies among random variables
have often been either ignored:
– Or else modeled as perfect correlation
Dependencies
• The use of copulas or other multivariate distributions has
become more common:
– But tractable models still are not sufficiently general to account
for all realistic assumptions, such as E(X|D) > E(Y|D) for all D
• High-dimensional joint distributions are also challenging:
– Correlation matrices must be positive definite
– There can be numerous higher-order correlations to assess
• Cooke et al. developed a practical method for specifying
a joint distribution over n continuous random variables:
– Using only n(n1)2 assessments of conditional correlations
– (Bedford and Cooke 2001; Kurowicka and Cooke 2004)
Subjectivity
• PRA practitioners sometimes treat the
subjectivity of prior distributions cavalierly:
– Best practice for eliciting subjective priors is
difficult and costly to apply
– Especially for dozens of uncertain quantities
• The use of “robust” or “reference” priors
may minimize the reliance on judgment:
– Although this may not work with sparse data
Probability Bounds Analysis
• Specify bounds on the cumulative distribution functions of the inputs:
– Rather than specific cumulative distributions
– (Ferson and Donald 1998)
• These bounds can then be propagated through a model:
– The uncertainty propagation process can be quite efficient
– Yielding valid bounds on the cumulative distribution function for the final
result of the model (e.g., risk)
• Can take into account not only uncertainty about the probability
distributions of the model inputs:
– But also uncertainty about their correlations and dependence structure
• This is especially valuable:
– Correlations are more difficult to assess than marginal distributions
– Correlations of 1 or -1 may not yield the most extreme distributions for
the output variable of interest (Ferson and Hajagos 2006)
Exposure to Contamination
• Regan et al. (2002) compare a two-dimensional Monte
Carlo analysis of this problem to the results obtained
using probability bounds
• The qualitative conclusions of the analysis (e.g., that a
predator species was “potentially at risk” from exposure
to contamination) remained unchanged:
– Even using bounds of zero and one for some variables
• Bounding analysis can help support a particular decision:
– If results and recommendations are not sensitive to the specific
choices of probability distributions used in a simulation
Model Uncertainty
• Uncertainty about model form can be important
• Assessing a probability distribution over multiple
plausible models is frequently not reasonable:
– “All models are wrong, some models are useful” (Box)
– Models are not a collectively exhaustive set
– Some models are intentionally simple or conservative
• Bayesian model averaging avoids giving too much
weight to complex models (Raftery and Zheng 2003):
– But still relies on assigning probabilities to particular models
– Using Bayes theorem to update those probabilities given data
Joint Updating
• In general, one will be uncertain about both model inputs
and outputs
• One would like to update priors for both inputs and
outputs consistently:
– With the wider distribution being more sensitive to model results
• Raftery et al. (1995) attempted this (Bayesian synthesis):
– But that approach is subject to Borel’s paradox
– Since it can involve conditioning on a set of measure zero
• Joint updating of model inputs and outputs is largely an
unsolved problem