AI weather forecasting & weather impact prediction

Weather drives energy prices, retail demand, construction schedules, harvests, road safety and insurance losses. For decades, forecasts came only from physics-based numerical weather prediction (NWP) run on supercomputers. Since 2025, machine-learning weather models have run operationally next to physics models: ECMWF's AIFS Single since February 2025 and the 51-member AIFS ENS since July 2025. They are as skilful as physics models or better on many scores, and they cost a tiny fraction of the compute. The business value comes from the second step, impact prediction. That step turns "12 mm rain, gusts 70 km/h" into "higher umbrella sales in store Y", "crane stop Thursday afternoon" or "elevated grid-fault risk in district X", with a probability attached.

Problem & value

• Who has the problem: energy traders and grid operators, retailers and food producers, logistics and transport operators, construction firms, farmers and agri-cooperatives, winter road services and municipalities, insurers (event response, claims staffing), event and hospitality operators, and private people (garden, PV, travel, hobby flying or sailing).
• Today: staff read a generic weather app or buy a point forecast from a weather service and decide by gut feeling. National-scale NWP is good, but it is coarse for one site, and it says nothing about the impact on your KPI.
• Why AI helps (two layers): 1. AI weather models produce large ensembles cheaply, which gives better probabilities for extremes. ECMWF reports that AIFS ENS beats physics-based models on many measures, with gains of up to 20 % for surface temperature, runs more than 10× faster and uses about 1,000× less energy [1]. AIFS Single improves tropical-cyclone tracks by up to 20 % [2]. GenCast outperformed the ECMWF ensemble on 97.4 % of 1,320 targets [3]. 2. Impact models (gradient boosting, time-series models) learn from your own history how weather moves your KPIs, and they make forecasts actionable.
• Free inputs: ECMWF publishes AIFS output under CC BY 4.0, four times a day [4]. DWD has had a legal mandate since 2017 to offer most of its model and observation data at no or low cost [5].

Variants

• Energy: wind/PV/load forecasts → handled in PRED-007. This file supplies the weather input.
• Retail & consumer goods: weather-driven demand uplift (beverages, ice cream, garden, apparel) as a feature for PRED-001; staffing of stores and call centres (PRED-011).
• Agriculture: frost, spray windows, disease-pressure models, irrigation timing (OPT-014), and yield prediction (PRED-013).
• Construction & logistics: crane and wind limits, concrete pouring windows, delay and ETA impact (PRED-018), road weather and winter maintenance (gritting), traffic demand (PRED-019).
• Aviation & shipping: weather routing is handled in OPT-026. Certified aviation meteorology stays with licensed providers.
• Insurance & public sector: hail and storm event alerts for claims staffing (AGENT-014), and impact-based warnings for civil protection. Floods → PRED-027. Long-horizon climate risk → PRED-036. Air quality → PRED-042. Climate research → SCI-010.
• Private use: hyperlocal apps, smart-home PV/heat-pump control and garden irrigation, built on consumer weather APIs.

How it works

1. data ingest (AIFS/IFS/ICON/GFS GRIB, radar, satellite, own stations)
2. (optional) own AI model run: global ML model from analysis
3. ensemble
4. downscaling / bias correction to site (statistical post-processing, generative downscaling)
5. nowcasting for 0–6 h from radar + satellite
6. impact model: weather features + calendar + business history
7. KPI forecast (quantiles)
8. decision rules & thresholds
9. alert / schedule / order / staffing proposal
10. LLM briefing ("what does this mean for site X tomorrow?")
11. human decision

• AI weather models are graph neural networks, transformers or diffusion/generative models trained on decades of ERA5 reanalysis. They still need initial conditions from data assimilation, which is mostly physics-based today [1].
• Post-processing (classic MOS, quantile regression, gradient boosting) removes local bias. For a single site it often adds more value than switching the global model.
• Impact layer: a gradient-boosted or time-series model per KPI, trained on the past forecasts that were actually available at decision time (not on observed weather), so that it does not leak hindsight.
• Human-in-the-loop: dispatchers, site managers or duty meteorologists confirm high-impact actions. Official public warnings are issued only by the national weather service.

Data preparation

Two data groups are needed. Weather data (model output, radar, own stations) comes from public providers and is usually clean, but arrives as GRIB/NetCDF and must be converted to tables for site locations. The business side is the harder part: sales, load, delays or damage history live in the ERP or shop system, in Excel files and in the building-management or SCADA system, often with gaps and changed definitions.

Typical gaps: no site coordinates or thresholds in master data, own sensors with missing readings, history without timestamps at hourly level, and no record of which forecast was available at decision time.

Preparation steps: - Inventory the sources and owners (PREP-001, PREP-004) and export ERP data (PREP-005). - Clean master data: coordinates, heights, thresholds per site. - Align time series to hourly values, mark outliers and closed days. - Store each forecast run as received, so impact models train on real forecasts. - Collect past weather-related incidents as labels (PREP-020).

Tools: Python with xarray, pandas, or a managed weather-API; for SMBs a spreadsheet export is often enough. Share of total effort: about 30 to 40 percent, mostly for the business history. See DATA-003 for time-series basics.

Implementation paths

Which path for whom

• smb (A): Mature weather-API and alerting products with EU hosting exist at low monthly prices, and site forecasts are a standard process.
• enterprise (B): Impact forecasts only pay off when joined with own sales, load or asset data, so cloud APIs plus own models fit best.
• developer (C): Open weather data (DWD ICON, ECMWF open data) and open-weight AI models run on own hardware and allow full control.
• private (A): Ready-made weather apps and alerts cover the need; self-hosting adds no value.
• public (B): Official data from DWD and EU services via EU-hosted cloud APIs keeps procurement simple and sovereignty intact.

Path A — Buy (recommended for most SMBs)

• Weather + AI forecast APIs: Meteomatics (Switzerland; one API with 1,800+ parameters, which also serves GraphCast, FourCastNet, AIFS and AIFS ENS output) [6]; Jua (Zurich; EPT-2 AI models and wind/solar MW forecasts for energy traders) [7]; Open-Meteo (open-source API with servers in Europe; offers ECMWF AIFS) [8]; Google Maps Platform Weather API (240 h hourly forecasts) [9].
• Domain products: energy-trading forecast tools, farm-management platforms with spray and frost windows, and road-weather services. Prefer EU/CH vendors with a DPA and an EU hosting statement.
• Choose A when: the impact logic is standard (alerts, thresholds) or the vendor already models your domain.

Path B — Build on cloud APIs

• Pull forecasts from Meteomatics, Open-Meteo (commercial plan) or the DWD/ECMWF open data, then build your own impact model in your data platform.
• Google WeatherNext 2 (FGN-based; available through Earth Engine, BigQuery and a Vertex AI early- access programme) produces hundreds of scenarios in under a minute on one TPU [10].
• Use an EU-hosted LLM (e.g. Mistral or Azure OpenAI in an EU region) for briefings, and n8n or Airflow for scheduling.
• Choose B when: the impact model on your own KPIs is the differentiator and you have no on-prem constraint.

Path C — Self-host (open source)

• Data only: self-host Open-Meteo (AGPLv3, Docker) on a CPU server [8]. Ingest DWD ICON and ECMWF open data, and run the impact model on CPU (HW1).
• Own AI weather runs: ECMWF AIFS Single (CC BY 4.0 weights, about 1.2 GB checkpoint, run via anemoi-inference) [11]; the Anemoi framework (Apache 2.0) [12]; NVIDIA Earth2Studio (Apache 2.0; wraps FourCastNet 3, AIFS, GraphCast and FuXi-S2S) [13]; Microsoft Aurora (non-commercial by default, commercial use on request) [14]. Check each checkpoint's licence. GraphCast and GenCast weights are restricted to non-commercial use.
• Hardware: HW3 (one 48–141 GB data-centre GPU) for running global 0.25° models and ensembles with headroom. The impact layer alone runs on HW1.
• Ops effort: fetching initial conditions every 6 h, GRIB tooling (ecCodes, xarray), monitoring and verification. This only makes sense with in-house meteorology or data-science skills.

Path D — Custom training

Worth it for regional high-resolution models, special variables (icing, hail, fog, turbine-height wind), or a new region. Fine-tune a foundation model (Aurora was designed to be fine-tuned at modest cost [15]) or train with Anemoi on ERA5 plus regional reanalysis or radar archives. This needs years of gridded data and multi-GPU nodes (HW4) or more. National weather services, energy majors and reinsurers do this. SMBs should not.

State of the art (2026)

• Operational ML forecasting: ECMWF AIFS Single and AIFS ENS were upgraded to v2 on 12 May 2026 [4]. Google WeatherNext 2 beats its predecessor on 99.9 % of variables and lead times, and it powers Search, Gemini, Pixel Weather and the Maps Weather API [10]. NVIDIA Earth-2 adds open medium-range ("Atlas"), nowcasting, AI data assimilation and CorrDiff downscaling models [16].
• Benchmarks: WeatherBench 2 compares Pangu-Weather, GraphCast, FuXi, Aurora, GenCast, FGN, NeuralGCM and others against IFS HRES and ENS [17].
• Trends 2025–26: ensembles by default (generative models), end-to-end AI pipelines including data assimilation, km-scale generative downscaling, earth-system foundation models (air quality, waves), and LLM agents that translate forecasts into decisions.
• Limits: AIFS ENS runs at 31 km versus 9 km for the physics ensemble [1]. Older deterministic ML models smooth out extremes, local convection and fog remain hard, and models trained on the past may under-represent unprecedented extremes. Impact models break when the business process changes. Always verify against your own observations.

Shared resources & combinations

• Same weather data pipeline (GRIB ingest, ERA5, post-processing): PRED-007 energy forecasting, PRED-027 floods, PRED-042 air quality, PRED-036 climate risk, PRED-013 yields, OPT-026 weather routing. One ingest and feature store can serve all of them.
• Same tabular/time-series stack (impact layer): PRED-001 demand, PRED-011 staffing, PRED-018 ETA, PRED-019 traffic. Add weather as a shared feature set.
• Downstream optimisation: OPT-011 battery dispatch, OPT-022 grid redispatch, OPT-014 irrigation.
• Remote sensing: VIS-021 satellite analysis, VIS-013 wildfire monitoring, VIS-058 post-disaster damage mapping (after the event).
• Hardware: one HW3 GPU server can run global AI weather models alongside other scientific-FM workloads (SCI-010). LLM briefings can share a general LLM server.

Effort, cost & KPIs

• Path A: days to weeks; subscription priced by API calls, locations or parameters.
• Path B: 1–3 months for an impact model with back-testing; the main cost is data engineering, not compute.
• Path C: 2–4 months; one GPU server plus GRIB/ops know-how.
• Path D: quarters; multi-GPU training and meteorological expertise.
• KPIs: forecast skill against your stations (MAE, CRPS, Brier score for threshold events); hit and false-alarm rate of alerts; impact-forecast error (e.g. MAPE on demand); avoided costs (spoilage, idle crews, balancing energy, frost damage); lead time gained for decisions.

People & change

Involve management for budget and scope first, then the department lead who owns the weather-sensitive process (site operations, logistics, grounds). IT connects data sources and accounts; data owners release ERP and sensor data. End users (dispatchers, site managers) join the pilot early because they decide on thresholds and actions.

Works council: usually no, because the system forecasts weather and asset impacts, not staff behaviour. It becomes relevant if forecasts drive shift planning or individual workload; then §87(1) no. 6 and §90(1) no. 3 BetrVG apply, and the council should be informed in good time. In the public sector the Personalrat applies equally.

Data protection: weather data is not personal; the DPO is needed only if staff data enters the impact model, then a DPIA check under Art. 35 GDPR. Conclude an AVV (Art. 28 GDPR) with any API provider that processes personal data.

AI literacy: support staff with a short briefing on probabilities, ensembles and forecast limits (GUIDE-006). Start with one site or team, compare forecasts with outcomes for a season, and keep humans in control: alerts are proposals, dispatchers confirm high-impact actions. Typical concern: "the forecast was wrong". Address it by showing probabilities and tracked hit rates (GUIDE-008, GUIDE-009).

Risks, regulation & compliance

• EU AI Act: a weather or impact forecast for business planning is minimal risk (AI literacy, Art. 4). It can become high-risk under Annex III point 2 when the AI system is a safety component in managing road traffic or the supply of water, gas, heating or electricity. It can also become high-risk under point 5(d) when it is used to dispatch or prioritise emergency services [18]. Classification therefore depends on use. Consumer chatbots that explain the weather need Art. 50 transparency.
• GDPR: weather data is not personal. Impact models on sales or staff data, or on precise locations of private persons (consumer apps), fall under GDPR.
• Sector rules: official warnings stay with the national meteorological services (DWD in Germany). Aviation needs certified providers. DORA and NIS2 matter for dependence on third-party forecast feeds in finance and critical infrastructure.
• Licences: observe CC BY attribution (ECMWF, Open-Meteo) and the non-commercial restrictions on some model weights.
• Mitigations: ensembles and probabilities instead of single values, fallback to physics NWP, verification dashboards, human sign-off for safety-relevant actions.

Getting started

1. List decisions that depend on the weather and the cost of a wrong decision. Pick 1–2 with clear thresholds (frost, wind limit, demand peak).
2. Collect 2–3 years of your KPI history plus the forecasts that were archived at the time (ERA5 or vendor archives as a proxy).
3. Trial an AI-enabled forecast API (Meteomatics, Open-Meteo with AIFS, Jua for energy) and verify it against your own station for a season.
4. Build a simple gradient-boosted impact model with quantile output. Back-test it, then wire alerts into your workflow.
5. Only then consider self-hosting AIFS via Anemoi or Earth2Studio, if volume, latency or sovereignty requires it.

Sources

1. ECMWF — "ECMWF's ensemble AI forecasts become operational" (2025) — https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ensemble-ai-forecasts-become-operational
2. ECMWF — "ECMWF's AI forecasts become operational" (Feb 2025) — https://www.ecmwf.int/en/about/media-centre/news/2025/ecmwfs-ai-forecasts-become-operational
3. Price et al. — GenCast: Diffusion-based ensemble forecasting for medium-range weather — https://arxiv.org/abs/2312.15796
4. ECMWF — AIFS machine learning data (versions, licence) — https://www.ecmwf.int/en/forecasts/datasets/aifs-machine-learning-data
5. Deutscher Wetterdienst — Open Data — https://www.dwd.de/EN/ourservices/opendata/opendata.html
6. Meteomatics — Weather API — https://www.meteomatics.com/en/weather-api/
7. Jua — AI weather models for energy — https://www.jua.ai/
8. Open-Meteo — ECMWF API docs & GitHub — https://open-meteo.com/en/docs/ecmwf-api ; https://github.com/open-meteo/open-meteo
9. Google Maps Platform — Weather API overview — https://developers.google.com/maps/documentation/weather/overview
10. Google — "WeatherNext 2: our most advanced weather forecasting model" (Nov 2025) — https://blog.google/technology/google-deepmind/weathernext-2
11. ECMWF — AIFS Single model card on Hugging Face — https://huggingface.co/ecmwf/aifs-single-1.0
12. ECMWF — Anemoi core — https://github.com/ecmwf/anemoi-core
13. NVIDIA — Earth2Studio — https://github.com/NVIDIA/earth2studio
14. Microsoft — Aurora repository — https://github.com/microsoft/aurora
15. Bodnar et al. — Aurora: A foundation model for the Earth system — https://arxiv.org/abs/2405.13063
16. NVIDIA — Earth-2 — https://www.nvidia.com/en-us/high-performance-computing/earth-2/
17. WeatherBench 2 — https://sites.research.google/weatherbench/
18. EU AI Act Annex III (Regulation (EU) 2024/1689) — https://artificialintelligenceact.eu/annex/3/