Strategies for generating physically consistent realizations from a multi-model blend

Blending weather forecasts from different sources aims to generate a single forecast which is temporally seamless, spatially consistent and more skilful than the individual inputs. This forecast can capitalise on all available forecast sources throughout the lead time range, such as extrapolation-based nowcasts and convection-permitting ensemble models at shorter lead times, and coarser ensemble models at longer range. Optimising the individual models contributing to the blend using post-processing techniques like neighbourhood processing and calibration helps to optimise overall forecast skill.

A multi-model blend can be created either in physical space or in probability space depending upon the desired output format and the interaction of multi-model blending with other processing steps. The IMPROVER codebase (https://github.com/metoppv/improver) utilises multi-model blending in probability space, which results in spatially and temporally smooth probability and percentile forecasts that are ideal for some use cases. Blending in probability space can avoid artefacts that can be difficult to overcome when blending in physical space. The aim for generating physically consistent realizations from the multi-model blend in IMPROVER is therefore to retain the benefit of the forecast source-specific processing and the relative ease of the probability space blending. Physically consistent realizations are desired by multiple users, including as inputs to hydrological models that require physical consistency including between diagnostics, such as, precipitation and snow melt.

This presentation outlines the requirements for physically consistent realizations, including as input for hydrological models. Prior work focusing on the usage of Ensemble Copula Coupling (ECC) and Schaake Shuffle will be reviewed and challenges with implementing ECC where the choice of dependence template is not obvious will be discussed. Proposals and initial findings will be presented for creating realizations that are self-consistent spatially, across lead times and across diagnostics following multi-model blending in probability space.