Comparing AI and NWP operational forecasts for the energy sector for winter 2025/26

AI forecasts have some impressive high-level statistics; less clear is how useful they are in operational forecasting for end users. 

Many went into winter 2025/26 excited to see if AI models (including the ECMWF AI ensembles) would have better skill at forecasting the significant atmospheric pattern shifts during the winter.  For energy end users the question was especially whether they would forecast any potential cold outbreaks more rapidly, or more reliably, than NWP model output.  And how they handled associated wind, and so renewable output. 

Here we take a few case studies over the last winter and compare the ECMWF AI and NWP ensemble forecasts for 2m temperature and 100m wind over NW Europe.

Winter 2025/26 had some significant stratospheric activity, including an early sudden stratospheric warming in Nov which many operational meteorologists (and some academics) regard as influential on the winter temperatures for Europe throughout winter, tho by no means the only driver!  This raises the question of whether a lack of (upper) stratosphere in the ECMWF AI ensembles impacted forecast skill.  We will look at how well the ECMWF AI ensembles fared in capturing changes propogating from the upper stratopshere.

ECMWF will release new models for NWP and AI in spring 2026 with the new AI model including a 10hPa level for better representation of the stratosphere.  At the time of writing this abstract, it’s unclear when AI v2 test data will be available from: if possible we will look to see whether a better representation of the stratosphere in the next generation of the ECMWF AI model improved  performance during stratospherically-driven events of last winter.