Value of Adding AI Weather Model Ensembles to a Multi-model Probabilistic Blend
Many weather forecasting organizations use a multi-model blend of ensemble members from different prediction systems to obtain a more accurate forecast. Recent studies have shown that blending deterministic AI-based models alongside physics-based models improves the overall forecast skill. However, the incorporation of an ensemble AI-based model into a multi-model blend is yet to be investigated. This study explores the value of blending the ensemble version of the Artificial Intelligence Forecasting System (AIFS-CRPS) with the Met Office's convection permitting physics-based ensemble, MOGREPS-UK. We show that blending site temperature (and wind speed) forecasts over the United Kingdom from both ensemble physics-based and ensemble AI-based models leads to an increase in overall forecast skill spatially, across lead times and across exceedance thresholds, especially when the models have different error characteristics. By decomposing the continuous ranked probability score of the blend into component means and diversity terms, we show that a more accurate blend can be achieved by assigning more weight to more skillful models and using models with complementary errors. For example, we show that a blend of MOGREPS-UK with either AIFS-CRPS or the physics-based Integrated Forecasting System (IFS) is better than a blend of IFS and AIFS-CRPS which are not very diverse. This has implications for the importance of weather model diversity and may lend weight to different model development approaches being pursued in tandem. We also demonstrate that calibrating AIFS-CRPS forecasts yields an increase in forecast skill comparable to calibrated ensemble physics-based forecasts, and calibrating the source forecasts before blending and the blended forecast after blending improves the overall forecast accuracy.