The winter of 2010: how my robust neighbourhood fell through
A neighbourhood sized robustly on three weather years meets its 90% promise in 8 of the 15 years between 2005 and 2023. The cold winter of 2010 is the real yardstick: 82.5% instead of 90%, and covering that year takes nearly double the seasonal store.
4 min read
Robust was my best word
Image above: AI impression, not a construction drawing.
In July I sized the neighbourhood robustly. That word had a precise meaning: the installation has to meet its self-sufficiency target in each of three real weather years, 2018 (normal), 2021 (grey) and 2022 (record sunshine), not on average. That requirement sits in the public assumptions register as A55, and the explorer on this site runs on it.
It was true, too. And still it nagged, because three is a small number, and the assumption said so itself: three weather years is an indication of spread, not a statement of reliability.
So I built the test my own word had to survive. Fifteen real weather years, 2005 to 2023, and for each of them the same question: does the neighbourhood I called robust keep its promise in a winter it has never seen?
The ranking I did not want to see
For every year I first computed the deficit trough: the deepest continuous gap between demand and solar yield, the amount of energy storage has to bridge. That number takes seconds to compute, and back in July it already predicted faultlessly which weather year drives a sizing.
The ranking was harsh. At the top sits 2010, the cold winter, with a trough of 6,073 kWh. Then 2017 with 4,974. My three sizing years rank fifth, seventh and tenth out of fifteen. Mid-table. I had sized for average winters and called it robust.
Annual sunshine predicts nothing here, by the way: across fifteen years the correlation between annual irradiation and deficit trough is minus 0.12, effectively zero. 2010 and 2017 have exactly the same annual irradiation and the two deepest troughs. "Take an average weather year" is not a safe simplification. Not roughly unsafe. Just unsafe.
Eight out of fifteen
Then the real test: every robustly sized installation, capacities fixed, run through each of the fifteen years with perfect foresight, the most generous reading there is.
The 90% neighbourhood, with over 8 MWh of seasonal storage, meets its target in 8 of the 15 years. In 2010 it gets stuck at 82.5%. The 80% neighbourhood drops to 73.2% in that same year, and the 70% neighbourhood without seasonal storage to 65.6%. The median, meanwhile, sits neatly on target: as an expected value the sizing holds up, as a guarantee it does not. That distinction is exactly the difference between a brochure and a feasibility study.
Below the winter wall the picture changes. The 50% neighbourhood makes all fifteen years, 2010 included. Promise modestly, and you can keep your promise for a decade.
What the winter of 2010 costs
Then I let the model size again, this time on the three harshest years in the series: 2010, 2017 and 2021. The answer: the seasonal store grows from 8.6 to 14.7 MWh, nearly double. The battery actually shrinks a little, because 2010 is an energy problem, not a power problem. And strikingly: a low and a high storage price produce almost the same store. The physics dictates the size; the price only sets the pain.
On one honest yardstick, that decade coverage costs 40% extra at optimistic storage prices and 62% at realistic ones, roughly 5,300 euros per home per year on top of the old sum. The robustness premium inside the old window was a few percent. The jump from window to decade is an order larger, because the window missed the tail of the distribution. The distribution has a tail, and the tail is called 2010.
All of this now sits on the explorer, as the decade test below the weather-year strip. Set the slider to 90% and watch what 2010 does. Set it to 50% and see that the promise holds there. Check it. Attack it.
What this rests on
The weakest points of this result, for anyone who wants to attack it:
- The test assumes perfect foresight of every weather year. Every missed promise above is therefore a lower bound: a real controller misses more.
- Fifteen out of nineteen years: leap years drop out due to a limitation in the demand-profile alignment. And the real extremes are not even in the sample; the winters of 1963, 1985 and 1996 were harsher than anything in this series.
- Electricity prices before 2015 are synthetic (labelled as such); all cost comparisons therefore sit on one shared yardstick, and no single cost figure from those years is used on its own.
- The old explorer claim, robust across three weather years, was always literally true. This test does not change that. It only shows how big "three" is.
This article follows the winter wall and seasonal storage breaks the winter wall. The full method and all fifteen year tables are in the public lab log and the dataset on the explorer page.