How Google's AI Weather Models Are Revolutionizing Solar Power Forecasting

Introduction: AI is quietly transforming weather forecasting

Over the past few years, a quiet revolution has been underway in weather forecasting. At its center are the **AI weather models** developed by Google (DeepMind).

Traditional forecasting relied on "numerical weather prediction" — solving massive physical equations on supercomputers for hours. Now, AI can produce forecasts of equal or better accuracy in minutes. And that accuracy now exceeds ECMWF, the European Centre for Medium-Range Weather Forecasts — the world's reference institution.

The shock of Google's AI weather models

Google has evolved its AI weather models step by step:

- **GraphCast (2023)**: a graph neural network for 10-day forecasting. It outperformed ECMWF's deterministic forecast on **90%** of evaluation metrics, with inference taking about a minute.

- **GenCast (2024)**: a diffusion model producing 15-day **ensemble forecasts** — drawing 50 possible futures to express uncertainty. It beat ECMWF's ensemble system on **97.2%** of metrics.

- **WeatherNext (2025)**: Google's next-generation production model family built on the above.

In short, **"hours of supercomputing" became "minutes of AI inference"** — the cost and speed of weather forecasting have fundamentally changed.

For solar power, accurate weather forecasting is a lifeline

A solar plant's output is almost entirely determined by **irradiance** — how much the sun shines. And irradiance is driven by weather. In other words:

**Knowing "how cloudy tomorrow will be" means knowing "how much power you will generate tomorrow."**

That is a matter of money for plant owners:

- **Revenue planning**: in post-FIT market-linked sales, forecast accuracy directly affects income

- **Maintenance planning**: separating "weather-caused" dips from "equipment failure"

- **Output-curtailment readiness**: advance forecasts enable faster operational decisions

The higher accuracy of AI weather models is a tailwind that lifts solar generation forecasting to a new level.

Our approach: ensemble forecasting for solar generation

AiPowerAdvisor uses **ensemble forecasting** — combining multiple forecasts — to predict solar generation. By bundling several forecast members, we smooth out the jitter of any single forecast and present a **"most likely generation range"** that is far more practical for planning.

On top of that, our **generation diagnosis** cross-references actual historical weather data to separate "is this drop just weather?" from "is this panel degradation or soiling?"

Conclusion

Weather forecasting has seen dramatic improvements in accuracy, speed, and cost thanks to AI. For solar power operators, this is a tailwind that raises the bar on both generation forecasting and equipment diagnosis.

"AI predicts the weather, and from that weather, predicts generation" — we are pushing this forward to support Japan's solar operators.

---

*This article references technical information published by Google DeepMind on GraphCast, GenCast, and WeatherNext.*