Romania Warmed 1.2 °C in 123 Years: The Data Behind Our Solar-Powered OCR API
We run OCRskill’s inference on local solar power and certified green-hydro imports. People occasionally ask whether that is worth the trouble, and the argument usually collapses into anecdotes: the winter someone remembers being colder, the summer that felt worse than usual.
So we stopped trading anecdotes and loaded the actual thermometer record for the country our hardware sits in. Every chart below is live: hover it, tap it, search it, look up your own town.
Every year in Romania since 1901, coloured by how far it sat from normal
Blue is colder than the 1901-1930 average, red is warmer. Hover, tap or use the arrow keys.
That is one stripe per year from 1901 to 2023. Nobody drew it by hand and no model produced it. It is 156 weather stations reporting every single month for 123 years.
Where these numbers come from
The dataset is RoCliHom, published by the Romanian National Meteorological Administration and peer-reviewed in Scientific Data in 2025. It contains mean, maximum and minimum air temperature plus precipitation for 156 stations, monthly, from 1901 through 2023. That is 921,024 individual measurements, with no gaps.
“Homogenized” is the important word. Stations get moved, thermometers get replaced, cities grow around instruments, observation practices change. Each of those creates a step in the record that has nothing to do with weather. Homogenization detects those breakpoints and corrects them, which is why this archive is a better witness than any single station’s raw logbook.
The dataset is honest about its own provenance, and so are we: every value carries a flag saying whether it is an original reading, a gap-filled estimate, or a homogenized correction. In 1901, roughly 1% of monthly temperature values are original readings. In 2023, 97% are. The deep past is genuinely reconstructed, which is exactly why none of the conclusions below depend on it - the acceleration everyone argues about happens in the decades where the data is almost entirely raw observation.
Download it and check us: zenodo.org/records/14880417.
The trend, with nowhere to hide
How far each year sat from the baseline
Bars are single years, the line is a trailing ten-year mean. Baseline 1901-1930 = -.
Across the whole 123 years the national mean rose 0.12 °C per decade. But that single number hides the shape of it. Compare each decade against the 1901-1930 average:
| Decade | Mean | vs 1901-1930 |
|---|---|---|
| 1901-1910 | 8.32 °C | -0.14 |
| 1911-1920 | 8.43 °C | -0.02 |
| 1921-1930 | 8.62 °C | +0.16 |
| 1931-1940 | 8.46 °C | +0.00 |
| 1941-1950 | 8.47 °C | +0.02 |
| 1951-1960 | 8.65 °C | +0.20 |
| 1961-1970 | 8.58 °C | +0.12 |
| 1971-1980 | 8.36 °C | -0.09 |
| 1981-1990 | 8.55 °C | +0.09 |
| 1991-2000 | 8.84 °C | +0.39 |
| 2001-2010 | 9.40 °C | +0.95 |
| 2011-2020 | 10.15 °C | +1.69 |
| 2021-2023 | 10.46 °C | +2.00 |
For nine decades the country wobbles inside a band of about a fifth of a degree. Then it leaves the band and does not come back.
A few facts from the same table’s underlying data, each of which is a conversation-ender on its own:
- 2023 was the warmest year in the record, at 11.22 °C against a 1901-1930 average of 8.46 °C.
- 17 of the 20 warmest years happened after 2000.
- The last year that fell below the 1901-1930 average was 1997 - 26 consecutive years above it, and counting.
- The warmest year of the entire period 1901-1990 was 1951, at 9.67 °C. Seventeen years since 2000 have beaten it.
- The coldest year of this century, 2005, would have landed in the warmer half of the twentieth century.
“But my town is different”
Fair. It might be. National averages are exactly the kind of abstraction that makes people suspicious, so here is the raw material instead - all 156 stations, one at a time.
Look up your own town
Every one of the 156 stations in the archive, on its own.
Search for the place you grew up in. Switch between the annual mean, summer, winter and rainfall. The dashed line is that station’s own least-squares trend across its own 123 years.
You will not find a station that cooled. The smallest change in the country is +0.88 °C at Paltinis. The largest is +1.39 °C at Huedin. Every single one of the 156 moved in the same direction.
“It’s just cities and asphalt”
This is the strongest objection in the whole debate, and this dataset answers it directly. If the warming were an artifact of cities growing around thermometers, the fastest-warming stations would be the urban ones.
They are not. Flip the explorer above to “fastest” and read the list: Huedin, Sebes, Stanca Stefanesti, Fundulea, Suceava. Small towns and rural sites - Fundulea is an agricultural research field in Calarasi county. Meanwhile Filaret, Bucharest’s historic observatory sitting inside a capital that grew several times over during this period, warmed +1.29 °C and ranks 28th of 156. Twenty-seven mostly rural stations warmed faster than Bucharest.
Now flip to “slowest”: Paltinis, Tarcu, Penteleu, Fundata, Balea Lac. Every one of them is a mountain station, most above 1,400 m, none of them near a growing city. The ridge tops warmed least. Urban heat islands do not form at 2,000 m on a bare ridge, and asphalt does not explain a research field warming faster than the middle of Bucharest.
“It’s just a natural cycle”
Then it should show up in the other things a shifting climate would shift. It mostly doesn’t, and that is the point.
Switch the chart above to rain. Annual precipitation went from 649 mm to 676 mm, about +4%, buried inside year-to-year swings several times that size. There is no comparable trend. Whatever changed in Romania over 123 years, it did not change how much water falls out of the sky - it changed the temperature.
Which season moved the most
Average temperature in 1901-1930 against 1994-2023.
Two fingerprints in that figure are hard to explain with cycles.
First, the warming is lopsided across seasons: summer +1.79 °C, winter +1.22 °C, spring +1.00 °C, autumn +0.82 °C. Summer moved more than twice as far as autumn. Summer daytime highs went from 24.91 °C to 26.39 °C nationally, and in 2023 they hit 27.41 °C.
Second, and more telling: nights warmed faster than days. Minimum temperatures rose +1.51 °C while maximum temperatures rose +1.31 °C. A brighter or more energetic sun would do the reverse - it would push the daytime peaks up hardest. Heat being held in overnight instead of radiating away is the signature of a more insulating atmosphere, not a stronger sun.
Why an OCR company cares
Because this is the grid we plug into.
Our inference hardware runs on 84% local solar generation and certified green-hydro imports for the remainder. The summers now shown in that seasonal chart are the same summers our panels harvest in and our GPUs shed heat into. The +1.79 °C summer shift is not an abstraction on our side of the meter: it is more solar yield in June and a harder cooling problem in July, in the same year.
Every OCR request has an energy cost. AI vision workloads are among the most power-hungry things you can put on a server, and the industry’s standard answer is to buy certificates from a distant market and call it neutral. We generate ours on site instead, and we publish the monthly mix rather than a slogan.
This article is the same instinct pointed at a different dataset. We would rather hand you 921,024 measurements and a search box than ask you to trust a claim.
Building something that reads images at scale? Get a free OCR API key and run it on sunlight and river water.
Frequently Asked Questions (FAQ)
Where can I download this data and check the charts myself?
RoCliHom is on Zenodo at zenodo.org/records/14880417 under a CC BY 4.0 licence, and the method is documented in Scientific Data. Every figure on this page is computed in your browser from that archive, condensed into a single JSON file. Nothing is smoothed, cherry-picked or pre-baked beyond averaging months into years.
What exactly does “national mean” mean here?
The unweighted average across all 156 stations of each station’s own annual mean. No area weighting, no interpolation, no altitude adjustment - so mountain stations count the same as lowland ones. It is the most boring possible method, which is the point. Any station you pick individually tells the same story.
Isn’t 1.2 °C small?
It is the difference between two 30-year averages for an entire country, which is a very heavy object to move. Locally it shows up as the frost-free season stretching, 2023 arriving 2.76 °C above the old normal, and every year since 1997 sitting above a line that used to be the middle.
Does homogenization mean the numbers were adjusted to show warming?
Homogenization removes non-climatic jumps - a station moving, a screen being replaced, an observation hour changing - by comparing each station against its neighbours. It is applied blind to whether the correction increases or decreases the trend, and the flags in the dataset let you see which values were touched. If you distrust it, use the recent decades where 90%+ of the values are original readings: that is where the steepest warming sits.
Why is winter defined as December to February of the same year?
Meteorological winter normally spans December of one year into January of the next, which leaves the first year of any record with an incomplete season. Grouping December, January and February within the same calendar year keeps all 123 years complete and comparable. The seasonal contrast is unaffected.
Does running on solar power slow down the OCR API?
No. Battery storage and grid-tie handle the gaps, and green-hydro contracts cover the winter deficit. Latency and accuracy are identical - the only difference is where the electrons came from.
Data: RoCliHom - Dumitrescu, A., Micu, D., Guijarro, J. et al. “Long-term homogenized air temperature and precipitation datasets in Romania, 1901-2023.” Scientific Data 12, 1116 (2025). doi.org/10.1038/s41597-025-05371-4. Distributed by the Romanian National Meteorological Administration under CC BY 4.0 via Zenodo.