# Biased Risk Analysis

```Biased Risk
Analysis
Larry George
January 8, 1987
5/27/2015
PST
1
Odds Against Nuclear Accident at
Chernobyl ‘One in 10,000 Years’
5/27/2015
PST
2
Our Situation

Observed core melt frequency
 Surprise,


[until Fukushima Daiichi]
Discrepancy between observed and computed
core melt frequencies
How to correct it
 “I
just don’t believe nuclear power can be made safe,
intelligently and rationally, unless we can estimate
probabilities of accidents accurately.”
5/27/2015
PST
3
Closed Due to Uncertainties in
Risk Analyses
5/27/2015
PST
4
Computed core melt probabilities
WASH 1400 (1975 Rasmussen report)
P[CM/Reactor year] = 5X106
 SSMRP (1983 LLNL) P[CM AND
Eq/Reactor year] = 3X105
 Frequency from internal events (1987
Sandia) P[CM/Reactor year] = 1.5X104

5/27/2015
PST
5
Observed Frequency [ca. 1987]
Accumulating reactor years at the rate or
434 per year, assuming 50% availability
 E[Core Melt frequency] = 0.5*434*P[CM] =
0.585/year
 If you exclude chalk river, Windscale, and
Fermi, then E[Core Melt frequency] =
0.2/year

5/27/2015
PST
6
Observed Frequency of
Coremelt

P[CM/reactor year] = 0.0027 = 5/4300 reactor
years
 Chalk
river, 1952
 Windscale, 1956
 Fermi, 1966
 Lucano, 1969
 TMI, 1979
 Chernobyl, 1986

Not counting EBR-1, SL1, Brown’s Ferry, Cap la
R., Rocky Flats,…
5/27/2015
PST
7
Core Melt frequency as time
goes on
5/27/2015
PST
8
Does Uncertainty Explain
Discrepancy?

Engineering hypothesis
 “Why
should I do an accurate analysis, the
data is full of garbage?”

Morgenstern Hypothesis
 “The
more rudimentary the theory of the user,
the less precision is required of the data”

FALSE!
5/27/2015
PST
9
Simulations and error bands
5/27/2015
PST
10
Bias Explains Discrepancy

Bias toward smaller probabilities
 Probabilities:
human error, mechanical accident,
natural causes
 Modeling: omissions and commissions to cover
uncertainty

Bias in computations
 Analytically
convenient assumptions (multivariate
lognormal)
 Culling events for computational feasibility
 Invalid bounds
5/27/2015
PST
11
Arguments:


What happened isn’t going to happen in the
future. This may be true of early accidents, but
did Chernobyl operators learn? Humans haven’t
changed since TMI
Use Bayes law to account for computed
information as prior and failures as observations.
 NRC
used only US data with precursors deemed to
have small probability of core melt
 This will lead to continuing as if prior is biased.
5/27/2015
PST
12
Requirements for Risk Analyses

Observe and record frequencies of events,
especially human errors
 Simulated


and actual. LERs are too little
Test equipment. Don’t use expert opinions if
there is a choice
Eliminate biased computations and compute
legitimate bounds
 No
5/27/2015
point to the last if you don’t do the other two
PST
13
Remove Bias




Use legitimate and simple bounds on
P[CM/Reactor year]
Use unbiased, least squares estimates of
subjective distributions, if expert opinions must
be used; e.g., seismic hazard
simulation. Jackknife if you must simulate
5/27/2015
PST
14
Surprise! Risk per year is less
than observed

Jussi Vaurio observed that core melt
accidents were in first year(s) of
operations!
 EBR-1,
SL1, TMI, Chernobyl
 Since US construction halted in 1988, no first
year(s)
 Until 2012 and
 Fukushima Daiichi
5/27/2015
PST
15
Interesting!
Weapons
 Reprocessing
 Nuclear power
 Confirms
development
problems

5/27/2015
PST
16
```