A Multivariate Ensemble Post-Processing Technique for Physically Consistent Spot Forecasts

As meteorological organisations transition to high-resolution ensemble-based forecasting, they risk leaving behind downstream users who rely on deterministic data: a need that may arise from the inability to process large volumes of data or difficulty integrating probabilistic information into decision-making processes. Proposed solutions for such users typically involve providing the control (unperturbed) member of the ensemble or deriving a forecast through the independent treatment of variables (such as the median). However, relying solely on the control member undermines the benefits of ensemble forecasting, while univariate approaches can result in forecasts that lack physical consistency across variables. To address this, we propose a novel method to select ‘most-likely’ ensemble realisations, combining techniques from pre-existing ensemble post-processing methods. For a given location, we construct a timeseries of ‘most-likely values’ for variables of interest by extracting the mode from multivariate probability density distributions created at each timestep. We then select the ensemble member most similar to this timeseries using clustering techniques. Since the chosen realisation is a complete forecast from an individual model run, this allows us to deliver a spot forecast for that location that maintains physical consistency across all variables, including those not directly analysed. As a demonstration, we apply this method to output from the Met Office convective-scale ensemble MOGREPS-UK at 240 locations across the Met Office synoptic observation network, focusing on near-surface air temperature and windspeed. We find that the chosen member performs comparably to the control member at short lead times, but is able to outperform the control member at longer lead times. This is an important finding as it demonstrates an alternative to the control member for users who require physically consistent spot forecasts, utilising the additional information available in the ensemble. In addition to improving forecast accuracy, this method also offers the ability to tailor solutions for individual users.