Part 2 was about finding data I could defend. Part 3 was about getting the pipeline to behave itself.
Now comes the dangerous bit: looking at the results and resisting the urge to make them more exciting than they are.
The question from Part 1 was simple enough: when severe storms hit, do outages get longer or more frequent?
I ran it across every generator unit with a fuel label rather than disappearing straight down the offshore-wind rabbit hole.
That turned out to matter quite a lot.
What I measured
The outcome is unplanned unavailability reported through Elexon’s REMIT messages. These are the messages where a unit comes out of service without advance notice.
The storm exposure comes from a catalogue of 44 named storms since June 2017, each with a single date and the highest peak gust reported across the affected region. Each storm is treated as exposing the three days from its catalogue date.
Wind curtailment from NESO’s balancing data goes in as a control. It only covers wind, so it’s a weak control for everything else.
Across all assets, storms don’t clearly bring more outages
Storm-window days averaged 30.3 unplanned starts per day, against 26.7 on baseline days. That’s a difference of 3.6, and it isn’t statistically meaningful (permutation p = 0.20). Restricting to 2022 onward, where reporting is denser, gives a similar answer (p = 0.39).
A regression with wind curtailment, month and year effects gives a storm coefficient of +1.5 with a standard error of 3.6.
In technical terms: not much to get excited about.
The chart did, however, make me stop and stare at it for a bit. Unplanned events are almost absent before 2022 and then cluster heavily afterwards. That looks much more like a change in how outages get reported than a change in how often they happen. I’ve partly controlled for it, but I haven’t fully resolved it.
Durations look the same for most fuel types
Across all assets, the median outage lasts 2.0 hours in storm windows and 2.0 hours on baseline days. There’s no difference.
Splitting duration by fuel type is where things start looking interesting.
Which is also exactly where I need to become more suspicious, not less.
Wind outages that start in a storm window have a median of 9.0 hours, against 4.0 hours on baseline days. That’s significant (p = 0.007), but it rests on 65 events.
Coal shows a small but significant rise in counts during storm windows (+1.4 events per day, p = 0.004).
Unlabelled units show the longest storm-window outages of any group, with a median of 13.4 hours against 6.5 hours on baseline days (p < 0.001). Most of the fleet has no fuel label, so this group is hard to interpret.
Biomass shows the largest count difference of any group, +5.8 events per day during storm windows (p < 0.001). Almost all of it comes from two stations, Lynemouth and Drax. The effect survives controls for month and year, so season doesn’t explain it. But the REMIT cause text for those stations during storm windows reads the same as on normal days: boiler and fuel supply, plant-available notices, milling. None of it mentions weather. I can’t yet say whether this is a storm effect, a change in how those stations report, or something else.
Gas, nuclear, pumped storage and most of the rest show nothing distinguishable from baseline.
Where I refuse to get carried away
Across generation assets, storms don’t show a consistent increase in how often units go unplanned-unavailable, or in how long those outages last.
The wind duration result is probably the number most likely to make a nice headline.
Naturally, it is also the one I’d be most nervous about leaning on. It’s 65 events, from a sample that changes depending on which years you include.
Several limits matter here:
- Nine fuel groups and several tests each mean some results will pass p < 0.05 by chance. I haven’t applied a correction yet.
- Most generator units have no fuel label in the reference data, so their outages sit in an unknown group.
- The storm catalogue is national: one date and one peak gust per storm, not a location-specific hazard.
- Counts are per unit, not per station, so a station with several units gets counted several times.
- REMIT does not state cause, so an association with a storm is not evidence the storm caused the outage.
So, did storms cause more outages?
Across the generation fleet, I can’t show a clear link between named storms and either more unplanned outages or longer ones. Most storm-window days look frustratingly similar to normal days.
That isn’t a failure. It is the answer this version of the data gives me.
There are three pockets where the data does differ, and none of them can yet be tied to weather:
- Wind and unlabelled units have longer outages in storm windows, but the wind sample is only 65 events.
- Two biomass stations have more outage starts in storm windows, with no weather wording in the cause text.
- Coal shows a small rise in starts, which may be chance given how many groups I tested.
So my working conclusion is deliberately boring: in this dataset, the fleet-wide relationship is weak or absent.
The odd pockets are leads to investigate, not discoveries to announce. REMIT can show me that an outage and a storm happened at roughly the same time. It cannot tell me the storm caused the unit to trip.
And I don’t want to turn correlation into causation just because it makes the project sound better.
Where this goes next
The bit I would chase next is Lynemouth and Drax.
Something is happening around those storm windows. The useful question now is whether operator notes or planning records explain it as weather, normal plant behaviour, or simply a reporting pattern that happens to line up with storms.
After that, the upgrades I’d make are storm severity from ERA5 at actual site locations, a Poisson or negative-binomial model for the counts, and a multiple-comparison correction across fuel groups.
Gareth Winterman