18 September 2026  ·  🔭 Astronomy 🖥️ Technology

Is tonight worth it? My Astro Weather page – and where every number in it comes from

The same question comes up before every night: is it worth starting the rig? The weather apps on my phone do not answer it. They tell me whether I need an umbrella – not whether there will be a gap in the high cloud during astronomical darkness between 10 pm and 4 am, whether the jet stream will smear every star, and whether the moon will be in the way at exactly that hour.

What finally tipped it was a discovery: Open-Meteo offers a genuinely good free API – clean data from the official models, with no account, no key and no registration. And an astro weather display that I can read at a glance was something I had wanted to build for a long time. With a data source like that there was no reason left not to.

So I built the answer myself: Astro Weather, seven days at a glance, made for that one question. It runs entirely in the browser, with no account and no app. This article shows what it displays – and above all where every single number comes from: what is model output, what is my own estimate, and where your IP address goes along the way.

The four locations you can look up

The drop-down is short, and deliberately so – the reason is further down. These are the places where I observe, or have someone observe for me:

• Starfront, Rockwood (Texas) – my remote rig, and the place the screenshots in this article come from. HRRR runs there at 3 km for the first 30 hours, then the Canadian GEM at 10 km for the rest of the week.
• Public Observatory Hannover – my observatory on my doorstep. Here the DWD’s ICON-D2 covers the first days at 2.2 km, then the Danish weather service’s HARMONIE at 2 km, then ICON-EU.
• Sankt Andreasberg Observatory – the dark sky of the Upper Harz, two hours by car. The same chain of ICON-D2 and HARMONIE.
• Finca Olivar, Gaucín (Spain) – Andalusia, much further south and drier. Spain lies outside the D2 domain, so ICON-EU runs there at 7 km.

On top of those there is a custom coordinates entry for anywhere in the world – the only one I had to put behind a password. More on that below, too.

Seven days at a glance

The overall row condenses each hour into a value from 0 to 100: 85 per cent and up is excellent, 65 good, 45 average, 25 poor, below that very poor. Cloud dominates almost everything in it – half covered means 25 per cent at best, and that is deliberate: under solid cloud the finest seeing buys you nothing.

Below it each ingredient gets its own row, so you can see why a night scores badly: cloud, the two comparison models’ cloud, seeing, transparency, wind with its direction, temperature and dew point. When temperature and dew point close in on each other, dew forms on the optics – which is why the spread turns orange below 4 °C and red below 2 °C.

And the answer stands as a sentence at the very top, so that nobody has to read the chart to know where they are: “8.0 h of good sky in a row from 20:58, 3.0 h of it moonless.” The same figures sit in the table’s best window column – the longest stretch whose rating stays above 65 per cent, and how much of it the moon leaves alone. The night average alone hides both: six good hours and two miserable ones average out to something middling, although the night is perfectly usable.

And one night in detail

A click on a night in the table opens it up: noon to noon, hour by hour.

This is also where the line I care about most sits: astronomically dark 21:02–06:02 (9.0 h) – moon down 6.4 h. Nine hours of darkness is lovely, and the moon takes only 2.6 of them – at 44 per cent lit that is a good night for narrowband and a decent one for RGB.

Where the data comes from

All the weather data is fetched by the browser from Open-Meteo, a Swiss service that republishes the weather services’ open data. The tool calls three endpoints and works out the rest itself:

Open-Meteo’s free tier comes with two conditions, both of which are met: the use must be non-commercial, and the source must be credited – which is why Weather data by Open-Meteo.com and its CC BY 4.0 licence sit under the chart.

Which model a night came from is printed above the night detail – HRRR (NOAA) at 3 km, then GEM (CMC) at 10 km in the screenshot above. The ICON box runs from 29.5 to 70.5° north and from 23.5° west to 62.5° east: inside that box ICON, outside it NOAA.

Why Texas does not run on ICON

ICON exists as a global model too, so it could have covered everywhere. Before splitting the models by region, I compared them: 30 nights around Starfront, model forecast against the cloud reports of the surrounding airports. GFS and HRRR were clearly closer to what was actually in the sky there; global ICON was the weakest of the three. Since then it has been the German model inside the ICON box and the American ones outside it.

Later I extended that measurement to 60 nights and broke it down by lead time – that is, keeping forecasts for tonight apart from forecasts for the Friday six days out. The result was clear enough to rebuild the chain: beyond the first day, the Canadian GEM sat closer to the airport reports at Starfront than NAM, NBM, ECMWF and GFS at every lead time. So since 20 September, Texas runs HRRR at the start and GEM for the rest of the week – and since 21 September HRRR gets only 30 hours rather than two days. Why, is in the follow-up: on the second day HRRR is worse than GEM, by a factor of two. That has a pleasant side effect: three models with two seams before, where a step in the curve could be nothing more than a change of model – two models with one seam now. And because GEM, unlike NBM, publishes cloud by height as well, the low, mid and high rows there no longer come from the coarser GFS.

NBM has not disappeared for all that: it still supplies the visibility and, alone among these models, a native precipitation probability, and in North America it now stands in the third cloud row – where GEM stands in Europe.

The two further cloud rows, ECMWF and GEM, come from the same mistrust. They feed into no rating; they simply sit there and show whether three independent forecast centres expect the same thing. When they agree, you can plan; when they contradict each other, that plainly means the forecast is uncertain, so look again tomorrow. In Europe the third row is GEM – a deliberate choice, it being the model behind the Clear Sky Charts I otherwise compare against – and in North America NBM, since GEM already fills the first row there.

Three rows are my estimate – and they are labelled as such

This is the part I most want to spell out, because at first glance it looks as official as everything else: no weather service delivers seeing, transparency or the overall rating. The tool works those out itself from the model data.

• Seeing comes from the wind profile: the jet stream at around 10 km, the shear between roughly 1.5 and 10 km, and the wind at the ground. It is the same principle the well-known seeing forecasts use – but it is a derivation, not a measurement of the turbulence.
• Transparency comes from the CAMS aerosol optical depth, damped by high humidity and by high column water vapour. After four to five days CAMS ends, and the row simply stays grey instead of inventing something.
• Overall is a weighting of cloud, seeing and transparency in which cloud dominates quadratically. If one of the two estimates is missing, it hands its weight to the others.

Those weights are not calibrated against observations. They point in the right direction – I claim no more than that, and the page says so itself. Anyone planning a night on them should take the cloud row seriously and read the other two as a hint.

What has been checked

For everything the tool computes itself, I checked the result against independent sources:

• Sun and moon against the tables of the US Naval Observatory: of 48 printed rise, set and twilight times, 44 were identical and the rest one minute apart.
• The values in the tooltips against the raw response of the same request – so that nothing slips between model and display.
• Which hour a value belongs to. At Open-Meteo, gusts, precipitation and precipitation probability refer to the preceding hour, everything else to the moment. Confuse the two and you move the rain by an hour.
• Local time. Nights run from noon to noon and can have 23 or 25 hours, so the UTC offset is worked out per moment rather than once for the whole week. That is exactly where the first version came apart on the night the clocks changed.
• Which model an hour came from is detected, not guessed – from the values themselves, not from the lead time.

What the page gives away about you

A weather tool is a location lookup, and that is not something to hand over casually. So:

• Nothing happens before the click. The page loads no forecast by itself. Only Load forecast sends the request to Open-Meteo – and with it, as with any request on the internet, your IP address.
• The weather map is a two-click embed. meteoblue sets its own cookies and pulls in an analytics service, so it loads only after an explicit click – and asks again after a page reload.
• The tool itself stores nothing about you. No account, no cookies of its own, nothing stored on a server: the location and °C/°F live in the address bar. A link to a night is therefore just a URL with the coordinates in it – shareable without a profile forming anywhere.
• The locations are limited to the four named above, and custom coordinates sit behind a password. The reason is the quota: free use of Open-Meteo is capped – 10,000 calls a day, and every forecast loaded is three of them. With open coordinates the page would be a weather API for the whole world, and the quota would no longer be mine. The password is no real wall: the check runs in the browser, and anyone determined gets past it. It is about stopping other people from spending my quota, not about keeping a secret.

What it cannot do

Seven days is the reach of the models, not my choice – and the last two days are a trend, not a promise. Transparency ends with CAMS after four to five days. And seeing remains a derivation from the wind: an evening when the air right above the telescope sits still while the jet stream rages is invisible to it.

What the tool deliberately does not do is the other half of the planning: where the moon stands, which objects climb high enough at all, what culminates when. That lives next door in the Observing Planner, which knows the same locations – and which now also says how many usable dark hours an object spends above 30°.

Conclusion

The question before the night is answered, and answered with a number you can open up and follow back to its source: the overall rating, behind it the cloud, behind that ICON or GFS, behind that a weather service with a name. What is computed is shown as computed; what is estimated is shown as estimated.

In a tool I built myself, that matters more to me than another row in the chart: you ought to be able to see how certain a statement is before you trust a whole night to it.