AI WX Predictions

[G4IPB]

Using AI to to get weather predictions

Firstly I think it is sensible to say that this is NOT foolproof but it certainly provides some food for thought and I have shared this out of interest - a bit like trying a different antenna rather than a recommendation.

On a Facebook group about Munros (Scottish Mountains) someone had made a complex prompt to help work out which hills might be worth researching and investigating the weather forecast. I was skeptical but on the basis of garbage in garbage out this gave so much detail to the AI model it seems to make quite a good go of it. The changes from Munro to SOTA based are not very pretty and I’m sure there is much room for improvement in the prompts but it is certainly somewhere between interesting and downright scary.

Thought I would share, and no I’m not about to follow Grok or any other model blindly up a mountain. I’m sure others will improve on it!!

( I used Chat GPT which did moan about the amount of data it was processing but it then got on and did it….)

When I request a SOTA weather check, run a full SOTA Weather Scan using EVERY rule below.

Please carry out a full check forinsert date. The start point location is

The summits requested are those in all English Welsh and Scottish regions specified by SOTA

In general summits or summit combinations with higher scores would be preferred

-–

ALWAYS USE ONLY THE LATEST AVAILABLE DATA

Use the newest updates from:

MWIS (all regions)

Met Office mountain area forecasts

Met Office summit forecasts

Mountain-Forecast

Local providers (Glencoe, Cairngorms, Nevis Range etc, Lake District, Snowdonia)

Webcams

ECMWF

GFS

UKV

Never reuse previous data.

If uncertain → include it with an honest rating, never omit.

-–

ALWAYS SCAN EVERY REGION BELOW (NO EXCEPTIONS)

Arrochar Alps

Assynt & Far North

Bridge of Orchy

Cairngorms

Cairnwell Hills

Crianlarich Hills

Cruachan Hills

Dundonnell & Fisherfield

Fannaichs & Ullapool

Glen Affric & Strathfarrar

Glen Etive Hills

Glen Shiel Hills

Glencoe Hills

Knoydart & Loch Quoich

Laggan & Monadh Liath

Loch Ericht to Loch Laggan

Loch Arkaig & Loch Eil

Loch Treig to Loch Ossian

Mamlorn Hills

Nevis Region

Rannoch & Glen Lyon

Tarf & Tilt

Tyndrum Hills

Pennines

Snowdonia

Brecon Beacons

Lake District

ALWAYS EXCLUDE ENTIRE RANGES:

Mamores

Lawers Range

Skye & Mull

Achnashellach & Torridon

-–

EXCLUDE ALL OF THIS YEARS COMPLETED SUMMITS

Never suggest any of the following

-–

EXCLUDE THESE “WINTER-ONLY / BOGGY” HILLS UNLESS FULLY FROZEN

Only include these if ground is genuinely frozen or snow-covered:

Tolmount

Tom Buidhe

Meall Buidhe

Am Faochagach

Beinn Chabhair

Ben Lui

Ben Challum

Carn Sgulain

Ben Oss

Fionn Bheinn

-–

Do Not INCLUDE MULTI summit ROUNDS

-–

OUTPUT MUST FOLLOW THIS EXACT 4-SECTION FORMAT

Clear-summit

Borderline

Cloud-inversion

Near-miss

Never reorder, rename, or merge sections.

Cloud-inversion must NEVER be joined with Near-miss.

-–

FOR EVERY HILL OR ROUND, PROVIDE THESE DETAILS

For each option:

Summit Points

Winter Bonus

Rating /10

Best summit-weather window (use insert start time, or default 09:00)

Distance (km)

Total time (hours)

Total ascent (m)

Vehicle Parking options

Time to first summit

Drive time from start point location shown at the start of the query

Scramble notes where relevant

Must obey completed list + frozen-only list

Must include all viable regions, no filtering

-–

ZERO BIAS

Never prioritise one region over another.

NEAR-MISS RULE

Near-miss must include:

1–2 hour clear windows

Cloud bases just below summits

Short, usable breaks

These ARE valid options — not rejects.

-–

HOME LOCATION FOR TRAVEL TIMES

Whenever providing drive times, use:

Start Point Location specified at the start of the query

START TIME RULE

Assume insert start time.

If none is provided, default to 09:00.

-–

WHEN I SAY “RUN THE FORECAST”

Immediately run this full 4-section forecast.

No confirmations.

No clarifying questions.

-–

WHEN I SAY “RUN AGAIN”

Perform a 100% fresh scan using the latest data.

SNOW COVER & AVALANCHE RISK SECTION (MANDATORY)

After completing the four main output sections (Clear-summit, Borderline, Cloud-inversion, Near-miss), you must ALWAYS include a dedicated Snow Cover & Avalanche Risk section for the specific day’s forecast. This section is mandatory for every run.

For every Hill suggested, you must provide:

Snow Cover Assessment

Expected snowline and snow depth by elevation

Surface type: fresh powder, graupel, windslab, refrozen crust, neve, verglas, or mixed winter conditions

Expected changes during the day (e.g., drifting, loading, thaw–freeze, consolidation, crust formation, solar softening)

Hazard notes specific to the route (e.g., corniced rims, icy slabs, verglas-prone boulders, scoured ridges, loaded lee slopes)

Identify any terrain traps (gullies, bowls, corrie rims, convex rolls)

Avalanche Risk / Instability Indicators

Because SAIS may not be in season, you must infer snow instability and avalanche likelihood using ALL available sources:

MWIS comments on drifting, slab formation, unstable accumulations, deep deposits

Met Office snowpack wording

Mountain-Forecast wind + precip data for predicting lee-side loading

Wind direction + speed vs slope aspect to identify likely windslab zones

Temperature profiles (freezing level, overnight low, daytime warming) to highlight crust weakness or rapid loading

Identification of slopes/aspects with elevated instability

Provide a plain-English avalanche risk rating for the day using:

Low, Moderate, or Considerable-like

(Not SAIS ratings — but equivalent qualitative guidance.)

This section must always be included and must always be specific to each route listed, not generalised.

FOLLOW THIS TEMPLATE EXACTLY

----- END OF PROMPT -----

[M0LEP]

@G4IPB

did moan about the amount of data it was processing

…which leaves me wondering how long it’ll be before the providers of the individual forecasts the LLM is ingesting will decide to proactively block (as best they can) visits by LLMs. Said blocking is becoming less trivial as the LLM scrapers mostly no longer declare what they are in any honest form, so they now generally need to be identified by behaviour, and there’s a bit of a digital arms-race going on at present…

[MW0PDV]

Thanks for this. I have just run the query using Copilot app (ChatGPT5)
No grumbles, it just took a bit of time to run.
Very impressive

[G7ADF]

@G4IPB

Using AI to to get weather predictions

Firstly I think it is sensible to say that this is NOT foolproof but it certainly provides some food for thought

I think it is fair to say , that no AI however advanced will ever be foolproof. For a start there are too many fools

But it made me think, about how we used to decide where to go back in the 80’s and 90’s, before readily available internet and AI. We used to have a saying, the weather tommorow will be the same as today. So combined with a bit of John Kettley and Michael Fish we could get a good idea what areas might be suitable for a day out climbing (so dryness over all else).

We also used to have various phone numbers, Rock N Run in Ambleside was always a favourite with the question being “Can you see the top of Loughrigg?”. If heading to Wales, you could ring the Snowdon Mountain railway on the pretence of booking a ride and asking if the weather was suitable for photography or something. This was usually done from a call box (remember those) from somewhere down the A55 so you could divert off to Tremadog or even back towards the Peak if it was really bad.

I think we only had one real bad washout day out of the probably hundreds, so I wonder how the human method actually compares to the AI method.

An interesting point to note is that one of the metrics used in measuring the effectiveness of some AI predictions is based on the ROC curve (Receiver Operating Characteristic) developed during WWII for measuring the performance of radar operators.

Ian

[G8ADD]

@G7ADF

But it made me think, about how we used to decide where to go back in the 80’s and 90’s, before readily available internet and AI. We used to have a saying, the weather tommorow will be the same as today.

And now you have made me think about how we did it in the 60’s! We tended to go out to an area for the week (no mountains around Brum and the motorways just being built, pre-motorways it was a full days drive to get to the Scottish border) and were proud of the best tech available - to have a pocket transistor radio to get the weather forecasts, particularly the ones covering sea areas, valuable for the synoptic situation - but often were camped in locations where there was little or no reception! We had an aneroid barometer in the kit, we learned to read the sky, we learned to dig a trench to judge snow conditions and how to recognise things like windslab, and we survived and thrived!

[MM0EFI]

We used the premium rate phone line that used to give mountain forecasts for east and west Scottish mountain areas.

We were poor students who couldn’t afford it, and all the phones at work had call blocking for premium rate numbers.

Except for the one in the staff restaurant, which got hammered at least couple of times a week during winter!

[GM5ALX]

None of your pictures have uploaded properly.

@G4IPB

31_20e3.png

-–

[MW0PDV]

I realise the use of AI is going to be a bit like Marmite.
And of course and with all the caveats, that the information cannot be taken as completely accurate.
However, it is interesting to see that the weather forecast results below, are more or less, aligned with the results I get when looking at the individual data sources.
It seems that the GIGO principle, (garbage in, garbage out) in the context of AI, is very much about the quality of the prompt.
Using Paul’s original prompt, in the same chat history that I had previously used, I entered the prompt :
“Run the prompt for Brecon on the 4th December, strictly follow all of the rules in the first prompt in this chat history”
I then get a nice summarised outlook, just for Brecon on the 4th December. Of course it’s only a forecast, and this far out it is only a rough guide.