Using Conditional Verification to describe Situation-dependant Model Biases for Surface Weather – Applications and Implications

Gridscale forecasts of surface weather delivered by operational global NWP suffer from biases which depend strongly on the weather situation and on geographical factors. Such biases also plague re-analyses, such as ECMWF’s ERA5, as operational models are the engines of those re-analyses. This presentation will itemise a number of different gridscale biases identified through a conditional verification exercise in which millions of station measurements were compared with short range Control run forecasts of the ECMWF operational ensemble. We will postulate what physical reasons might underpin these biases. There is for example a strong dependence of rainfall forecast bias on model near surface relative humidity, which seems to relate to the handling of droplet evaporation and other cloud physics processes. All such errors can in principle be addressed via ECMWF’s “ecPoint” post-processing approach; indeed the conditional verification activity here was managed via ecPoint calibration software. The resulting corrections will be illustrated.

Whilst data-driven AI models are currently delivering better predictions of the synoptic pattern than classical physics-based global NWP, the fact remains that those AI models are generally using unadjusted re-analyses for training, and so the situation-dependant biases will clearly put a cap on skill attainable by them for surface weather parameters, even when the forecast synoptic pattern is ‘perfect’. Some ECMWF views on how to overcome this barrier, to deliver even better predictions, will be very briefly presented.