Operational and prototype ensemble clustering post-processing approaches at the Met Office

Ensemble clustering is an efficient ensemble post-processing approach that distils an ensemble forecast into its prevalent forecast scenarios by grouping similar ensemble members together - something which is increasingly important in a world where ensemble data volumes are rapidly increasing. The application of a suitable clustering method combined with appropriate forecast visualisation allows a forecaster to effectively focus on the key possible outcomes, simplify the message and characterise and communicate forecast uncertainty more easily. For example, output may be presented as (1) probabilities of each cluster occurring, (2) representative or central members from each cluster showing alternative forecast directions, and (3) probabilities of threshold exceedance under each cluster.

Operationally, the Met Office runs a probabilistic weather pattern forecasting tool called Decider (Neal et al., 2024), where ensemble members are clustered according to their allocation to one of 30 predefined weather patterns (circulation types). This allows for changes in the large-scale circulation to be identified at a range of lead times and is useful for many downstream applications where the same weather patterns have been related to specific impacts. To be used alongside this, a prototype feature-based clustering approach is being trialled, which is the focus of this talk. Here, k-medoids clustering is applied to Fractions Skill Score (FSS) distances between members, using identified features in each member. These features may represent areas of hazardous weather or mesoscale or synoptic features, such as areas of heavy rainfall, damaging wind speeds, or weather fronts. The methods being trialled follow those initially developed by Boykin (2022) in PhD work at the University of Reading in collaboration with the Met Office, and include several post-processing steps. These steps are designed to (1) determine spatial distances between objects and cluster on those distances, (2) identify an optimal number of clusters, (3) identify windows of interest where clusters become more distinct, and (4) identify representative members within each cluster, within the time window, to provide plausible forecast scenarios or evolutions with associated probabilities. Multiple configurations of the prototype feature-based clustering approach have been trialled and early results will be discussed.

References

Boykin, K.A. (2022) Extracting likely scenarios from Ensemble Forecasts in real time. PhD in Atmosphere, Oceans and Climate, Department of Meteorology, University of Reading, DOI: 10.48683/1926.00111270.

Neal, R., Robbins, J., Crocker, R., Cox, D., Fenwick, K., Millard, J., Kelly, J. (2024) A seamless blended multi-model ensemble approach to probabilistic medium-range weather pattern forecasts over the UK. Meteorological Applications, 31(1), e2179.