Thinking systemically about safetey

Accidents involve much more than the reliability of parts. Safety emerges from the systemic interactions of devices, people and organizations. Nancy Leveson’s Engineering a Safer World (free pdf currently at the MIT press link, lower left) picks up many of the threads in Perrow’s classic Normal Accidents, plus much more, and weaves them into a formal theory of systems safety. It comes to life with many interesting examples and prescriptions for best practice.

So far, I’ve only had time to read this the way I read the New Yorker (cartoons first), but a few pictures give a sense of the richness of systems perspectives that are brought to bear on the problems of safety:

Leveson - Pharma safety
Leveson - Safety as control
Leveson - Aviation information flow
The contrast between the figure above and the one that follows in the book, showing links that were actually in place, is striking. (I won’t spoil the surprise – you’ll have to go look for yourself.)

Leveson - Columbia disaster

Nuclear systems thinking roundup

Mengers & Sirelli call for systems thinking in the nuclear industry in IEEE Xplore:

Need for Change Towards Systems Thinking in the U.S. Nuclear Industry

Until recently, nuclear has been largely considered as an established power source with no need for new developments in its generation and the management of its power plants. However, this idea is rapidly changing due to reasons discussed in this study. Many U.S. nuclear power plants are receiving life extensions decades beyond their originally planned lives, which requires the consideration of new risks and uncertainties. This research first investigates those potential risks and sheds light on how nuclear utilities perceive and plan for these risks. After that, it examines the need for systems thinking for extended operation of nuclear reactors in the U.S. Finally, it concludes that U.S. nuclear power plants are good examples of systems in need of change from a traditional managerial view to a systems approach.

In this talk from the MIT SDM conference, NRC commissioner George Apostolakis is already there:

Systems Issues in Nuclear Reactor Safety

This presentation will address the important role system modeling has played in meeting the Nuclear Regulatory Commission’s expectation that the risks from nuclear power plants should not be a significant addition to other societal risks. Nuclear power plants are designed to be fundamentally safe due to diverse and redundant barriers to prevent radiation exposure to the public and the environment. A summary of the evolution of probabilistic risk assessment of commercial nuclear power systems will be presented. The summary will begin with the landmark Reactor Safety Study performed in 1975 and continue up to the risk-informed Reactor Oversight Process. Topics will include risk-informed decision making, risk assessment limitations, the philosophy of defense-in-depth, importance measures, regulatory approaches to handling procedural and human errors, and the influence of safety culture as the next level of nuclear power safety performance improvement.

The presentation is interesting, in that it’s about 20% engineering and 80% human factors. Figuring out how people interact with a really complicated control system is a big challenge.

This thesis looks like an example of what Apostolakis is talking about:

Perfect plant operation with high safety and economic performance is based on both good physical design and successful organization. However, in comparison with the affection that has been paid to technology research, the effort that has been exerted to enhance NPP management and organization, namely human performance, seems pale and insufficient. There is a need to identify and assess aspects of human performance that are predictive of plant safety and performance and to develop models and measures of these performance aspects that can be used for operation policy evaluation, problem diagnosis, and risk-informed regulation. The challenge of this research is that: an NPP is a system that is comprised of human and physics subsystems. Every human department includes different functional workers, supervisors, and managers; while every physical component can be in normal status, failure status, or a being-repaired status. Thus, an NPP’s situation can be expressed as a time-dependent function of the interactions among a large number of system elements. The interactions between these components are often non-linear and coupled, sometime there are direct or indirect, negative or positive feedbacks, and hence a small interference input either can be suppressed or can be amplified and may result in a severe accident finally. This research expanded ORSIM (Nuclear Power Plant Operations and Risk Simulator) model, which is a quantitative computer model built by system dynamics methodology, on human reliability aspect and used it to predict the dynamic behavior of NPP human performance, analyze the contribution of a single operation activity to the plant performance under different circumstances, diagnose and prevent fault triggers from the operational point of view, and identify good experience and policies in the operation of NPPs.

The cool thing about this, from my perspective, is that it’s a blend of plant control with classic SD maintenance project management. It looks at the plant as a bunch of backlogs to be managed, and defines instability as a circumstance in which the rate of creation of new work exceeds the capacity to perform tasks. This is made operational through explicit work and personnel stocks, right down to the matter of who’s in charge of the control room. Advisor Michael Golay has written previously about SD in the nuclear industry.

Others in the SD community have looked at some of the “outer loops” operating around the plant, using group model building. Not surprisingly, this yields multiple perspectives and some counterintuitive insights – for example:

Regulatory oversight was initially and logically believed by the group to be independent of the organization and its activities. It was therefore identified as a policy variable.

However in constructing the very first model at the workshop it became apparent that for the event and system under investigation the degree of oversight was influenced by the number of event reports (notifications to the regulator of abnormal occurrences or substandard conditions) the organization was producing. …

The top loop demonstrates the reinforcing effect of a good safety culture, as it encourages compliance, decreases the normalisation of unauthorised changes, therefore increasing vigilance for any outlining unauthorised deviations from approved actions and behaviours, strengthening the safety culture. Or if the opposite is the case an erosion of the safety culture results in unauthorised changes becoming accepted as the norm, this normalisation disguises the inherent danger in deviating from the approved process. Vigilance to these unauthorised deviations and the associated potential risks decreases, reinforcing the decline of the safety culture by reducing the means by which it is thought to increase. This is however balanced by the paradoxical notion set up by the feedback loop involving oversight. As safety improves, the number of reportable events, and therefore reported events can decrease. The paradoxical behaviour is induced if the regulator perceives this lack of event reports as an indication that the system is safe, and reduces the degree of oversight it provides.

Tsuchiya et al. reinforce the idea that change management can be part of the problem as well as part of the solution,

Markus Salge provides a nice retrospective on the Chernobyl accident, best summarized in pictures:

Salge Chernobyl

Key feedback structure of a graphite-moderated reactor like Chernobyl

Salge Flirting With Disaster

“Flirting with Disaster” dynamics

Others are looking at the nuclear fuel cycle and the role of nuclear power in energy systems.

How to be confused about nuclear safety

There’s been a long running debate about nuclear safety, which boils down to, what’s the probability of significant radiation exposure? That in turn has much to do with the probability of core meltdowns and other consequential events that could release radioactive material.

I asked my kids about an analogy to the problem: determining whether a die was fair. They concluded that it ought to be possible to simply roll the die enough times to observe whether the outcome was fair. Then I asked them how that would work for rare events – a thousand-sided die, for example. No one wanted to roll the dice that much, but they quickly hit on the alternative: use a computer. But then, they wondered, how do you know if the computer model is any good?

Those are basically the choices for nuclear safety estimation: observe real plants (slow, expensive), or use models of plants.

If you go the model route, you introduce an additional layer of uncertainty, because you have to validate the model, which in itself is difficult. It’s easy to misjudge reactor safety by doing five things:

  • Ignore the dynamics of the problem. For example, use a statistical model that doesn’t capture feedback. Presumably there have been a number of reinforcing feedbacks operating at the Fukushima site, causing spillovers from one system to another, or one plant to another:
    • Collateral damage (catastrophic failure of part A damages part B)
    • Contamination (radiation spewed from one reactor makes it unsafe to work on others)
    • Exhaustion of common resources (operators, boron)
  • Ignore the covariance matrix. This can arise in part from ignoring the dynamics above. But there are other possibilities as well: common design elements, or colocation of reactors, that render failure events non-independent.
  • Model an idealized design, not a real plant: ignore components that don’t perform to spec, nonlinearities in responses to extreme conditions, and operator error.
  • Draw a narrow boundary around the problem. Over the last week, many commentators have noted that reactor containment structures are very robust, and explicitly designed to prevent a major radiation release from a worst-case core meltdown. However, that ignores spent fuel stored outside of containment, which is apparently a big part of the Fukushima hazard now.
  • Ignore the passage of time. This can both help and hurt: newer reactor designs should benefit from learning about problems with older ones; newer designs might introduce new problems; life extension of old reactors introduces its own set of engineering issues (like neutron embrittlement of materials).
  • Ignore the unknown unknowns (easy to say, hard to avoid).

I haven’t read much of the safety literature, so I can’t say to what extent the above issues apply to existing risk analyses based on statistical models or detailed plant simulation codes. However, I do see a bit of a disconnect between actual performance and risk numbers that are often bandied about from such studies: the canonical risk of 1 meltdown per 10,000 reactor years, and other even smaller probabilities on the order of 1 per 100,000 or 1,000,000 reactor years.

I built myself a little model to assess the data, using WNA data to estimate reactor-years of operation and a wiki list of accidents. One could argue at length which accidents should be included. Only light water reactors? Only modern designs? I tend to favor a liberal policy for including accidents. As soon as you start coming up with excuses to exclude things, you’re headed toward an idealized world view, where operators are always faithful, plants are always shiny and new, or at least retired on schedule, etc. Still, I was a bit conservative: I counted 7 partial or total meltdown accidents in commercial or at least quasi-commercial reactors, including Santa Susana, Fermi, TMI, Chernobyl, and Fukushima (I think I missed Chapelcross). Then I looked at maximum likelihood estimates of meltdown frequency over various intervals. Using all the data, assuming Poisson arrivals of meltdowns, you get .6 failures per thousand reactor-years (95% confidence interval .3 to 1). That’s up from .4 [.1,.8] before Fukushima. Even if you exclude the early incidents and Fukushima, you’re looking at .2 [.04,.6] meltdowns per thousand reactor years – twice the 1-per-10,000 target. For the different subsets of the data, the estimates translate to an expected meltdown frequency of about once to thrice per decade, assuming continuing operations of about 450 reactors. That seems pretty bad.

In other words, the actual experience of rolling the dice seems to be yielding a riskier outcome than risk models suggest. One could argue that most of the failing reactors were old, built long ago, or poorly designed. Maybe so, but will we ever have a fleet of young rectors, designed and operated by demigods? That’s not likely, but surely things will get somewhat better with the march of technology. So, the question is, how much better? Areva’s 10x improvement seems inadequate if it’s measured against the performance of existing plants, at least if we plan to grow the plant fleet by much more than a factor of 10 to replace fossil fuels. There are newer designs around, but they depart from the evolutionary path of light water reactors, which means that “past performance is no indication of future returns” applies – will greater passive safety outweigh the effects of jumping to a new, less mature safety learning curve?

It seems to me that we need models of plant safety that square with the actual operational history of plants, to reconcile projected risk with real-world risk experience. If engineers promote analysis that appears unjustifiably optimistic, the public will do what it always does: discount the results of formal models, in favor of mental models that may be informed by superstition and visions of mushroom clouds.