In short: Scope growth is the turnaround planning risk that every other risk feeds into, and controlling it starts with publishing what a day of the outage costs so every scope argument uses the same number. Scope arrives from three different places and each needs its own control, so a single freeze date addresses about a third of the problem. A schedule built from expected durations finishes late for structural reasons, and better estimating does not remove that. Discovery work is unpredictable item by item and fairly predictable in aggregate, which is exactly the condition where an allowance beats a contingency.
Day nineteen of a twenty-four day turnaround. A column has been opened, the internals are worse than the last inspection suggested, and the repair needs a fabricated part with a three week lead time. The alternatives are to extend the outage, to run the unit at reduced rate until the next opportunity, or to accept a deferral that the integrity engineer will sign for with visible reluctance.
Whichever one gets chosen, the decision is being made at the worst possible moment, with the whole contractor spread on site burning day rate while it is discussed. The interesting question is what could have been done eight months earlier so that this conversation happened when options were still cheap.
What a day of the turnaround costs
Every argument about turnaround scope is really an argument about days, so the day rate needs to be a published number that everyone in the process is working from.
Take a 200,000 barrel a day refinery with a gross margin of 8 per barrel. A day of full outage forgoes 1.6 million of margin. Add the direct cost of holding the execution spread on site: 400 contractor personnel at a fully loaded 900 a day is 360,000, plus cranes, scaffold, rental equipment and supervision. Call it two million a day, and substitute your own figures.
That number changes how several conversations go.
An acceleration proposal costing 700,000 to save a day is obviously worth taking, and in most organisations it gets refused because it exceeds a budget line while the two million sits in a different one.
A scope item added late that adds half a day costs a million, which is usually far more than the item's own cost, and it should be priced that way at the point of the request rather than at the post-mortem.
And the argument about whether the turnaround should run 24 days or 27 is a six million pound argument, which justifies considerably more planning effort than most organisations put into it.
Scope arrives from three places and they need different controls
Treating scope growth as one phenomenon produces one control, usually a freeze date, which addresses about a third of it.
Deferred work carried forward. Items that could not be done in the last outage or during a run-and-maintain window accumulate on a list. This population is knowable a year out and it grows in a predictable way, so it can be forecast rather than discovered.
Discovery on opening. Work identified when equipment is opened and inspected during the outage itself. This is genuinely uncertain in its specifics and highly predictable in aggregate, which is the section below.
Late additions from outside the process. A capital project that wants a tie-in, a regulatory requirement, a reliability improvement someone has been advocating for two years. These arrive because the outage is the only window, and they are the population a freeze date actually controls.
The three want different treatments. Carried-forward work wants a live register with an owner and a challenge process running through the year. Discovery wants a quantified allowance in the schedule and the labour plan. Late additions want a hard gate with a price attached, where the sponsor of a late addition sees the day cost of their request.
What happens instead in most organisations is a single scope freeze applied to all three, which the first two ignore because they cannot comply, so the freeze loses authority and then the third category walks through it too.
Why a schedule built from expected durations finishes late
There is a structural reason turnaround schedules overrun that has nothing to do with estimating optimism, and it is worth understanding because no amount of better estimating removes it.
A turnaround is many parallel workfronts converging on a common milestone. The unit restarts when the last one finishes, so the duration is the maximum across paths rather than the average.
The expectation of a maximum exceeds the maximum of the expectations. Clark set out the analysis of the greatest of a set of random variables in Operations Research in 1961, and Fulkerson showed in the same journal in 1962 that the critical path length computed from expected activity durations is a lower bound on the expected project duration. In project management practice this is called the merge bias, and it means a schedule where every individual path is planned at its own most likely duration will finish late, systematically, even if every single estimate is unbiased.
The arithmetic is stark. Take five parallel workfronts, each independently planned so that it has a 50 percent chance of finishing on its planned date. The probability that all five finish on time is 0.5 to the fifth power, which is 3 percent. Plan each one at an 80 percent confidence level and the probability all five make it is 0.8 to the fifth, or 33 percent. With twelve workfronts each at 90 percent confidence, the chance of an on-time restart is 0.9 to the twelfth, which is 28 percent.
Two things follow. Path-level confidence has to be much higher than the confidence you want at the milestone, and the required level rises with the number of converging paths. And the protection belongs at the convergence point rather than distributed across paths, for the same variance-pooling reason described in CC7 for drilling durations: a shared buffer sized from a simulation of the total is smaller than the sum of individual pads that deliver the same milestone confidence.
The practical implementation is a Monte Carlo run over the network with duration distributions on the activities, reporting the distribution of the restart date. That is a day of work with any modern scheduling tool, and it replaces a single date that is known to be optimistic with a curve that can be committed against.
Your own history is a better estimator than your bottom-up estimate
Every turnaround is estimated bottom-up, from a work list, with hours built from standards, and every turnaround overruns the bottom-up estimate. The estimate is not improved by doing it more carefully, because the thing it omits is the work that is not on the list yet.
Flyvbjerg's argument for reference class forecasting, set out in the Project Management Journal in 2006, applies here almost without modification. Take the class of comparable past events, look at the distribution of outcomes relative to their estimates, and apply that distribution to your current estimate rather than reasoning about why this one will be different.
For a site with eight past turnarounds, the reference class is on your own shared drive. For each one, record the estimated direct labour hours at the freeze date and the actual hours executed. If the median ratio is 1.39 and the interquartile range runs from 1.22 to 1.61, then a current bottom-up estimate of 180,000 hours implies a central expectation of 250,000 and a realistic upper case near 290,000.
That is the number to resource against, and it will be resisted, because it appears to reward poor estimating. The counter-argument is that the ratio has been stable across eight events under four different turnaround managers, which makes it a property of the process rather than of any individual's estimating.
Where the site has fewer than a handful of past events, the reference class can be built across sites within the group provided the unit types and turnaround intervals are comparable. What it cannot be built from is published benchmarks whose definitions of scope, hours and boundary you cannot see.
Discovery work, quantified before it happens
Discovery feels unpredictable at the item level and is quite predictable in aggregate, which is exactly the condition under which an allowance beats a contingency.
Build it from the inspection list. Suppose the turnaround includes 1,200 inspection items. From past events, 22 percent of items generate additional work, and the additional work averages 140 hours. Expected discovery is 1,200 times 0.22 times 140, which is about 37,000 hours.
Now convert that to days. With 400 people working 10 hour shifts, the site executes about 4,000 hours a day, so 37,000 hours of discovery is 9.2 days of work. If the schedule has three days of float and no explicit discovery allowance, the outage is going to overrun by roughly six days and everybody will describe it afterwards as an unusually bad turnaround.
The variance matters as much as the mean. The number of discovery items is close to binomial across 1,200 trials at a 22 percent rate, giving a standard deviation of about 14 items, and the hours per item have their own spread with a long right tail. Simulating that gives a distribution of discovery hours, and the resourcing decision is which quantile to staff against.
Staffing to the mean guarantees an overrun half the time. Staffing to the ninetieth percentile costs idle labour in the cases where discovery is light, and idle labour at 900 a day is cheap against a day of outage at two million. The asymmetry pushes hard toward over-resourcing, and the reason it rarely happens is that idle contractors are visible on site while the counterfactual delay is not.
Sequencing inspection so the surprises arrive early
The last controllable variable is when discovery happens, and it gets almost no attention.
Discovery on day two is manageable. The same discovery on day nineteen is the scene at the top of this piece, because the repair lead time now exceeds the remaining outage. So the inspection sequence should be ordered to bring forward the items most likely to generate work with long remedy lead times.
An index for that is easy to construct: for each inspection item, multiply the probability that it generates work by the lead time of the likely remedy, and inspect in descending order of that product, subject to whatever physical access constraints exist. The logic is the same family as Smith's weighted shortest processing time rule from Naval Research Logistics Quarterly in 1956, where ordering by a ratio of weight to duration minimises weighted completion time; here the weight is the cost of learning late.
In practice the ordering is usually driven by access and by scaffold sequencing, which is a genuine constraint, and there is normally more freedom than the plan uses. Opening the two vessels with the worst inspection history on the first available day, ahead of items that are opened first out of habit, is a change that costs nothing.
Pre-turnaround inspection is the stronger version of the same idea. Anything that can be assessed on a running unit, through non-intrusive inspection or the condition monitoring covered in N11, moves discovery out of the outage entirely. Each item moved is a scope item that can be planned, materialled and resourced at normal cost rather than at outage cost.
The limit
The estimation techniques above depend on a comparable reference class and an inspection history recorded consistently. Many sites have neither, because the turnaround team assembles for each event and disperses afterwards, and the records that survive are the cost report rather than the estimate-to-actual by work package. Building the reference class is a retrospective data exercise that has to be done once, from documents that are progressively harder to find the further back you go.
There is also a limit on how far duration can be compressed regardless of planning quality. Critical path work in a turnaround includes activities with fixed physical durations: cooling, purging, gas freeing, catalyst change, hydrotesting, drying and heating back up. Those consume days that no amount of labour buys back, and a schedule that has been compressed onto them is at the floor. Knowing where that floor is, for your units, prevents a lot of pointless acceleration spending.
And a caution about the discovery allowance. Publishing an expected discovery figure creates room for work to expand into it, because a scope item that would have been challenged now fits inside an approved allowance. The allowance needs the same challenge process as the planned scope, with the discovery register reviewed daily during the outage rather than absorbed silently.
Pull the estimated and actual labour hours from your last five turnarounds, compute the ratio for each, and take the median into the next estimate as a multiplier that nobody is allowed to argue away.