Ensemble weather forecasting scenarios tailored to users' needs

Ensemble forecasting has become the standard approach to represent uncertainties and produce multiple scenarios. Ensembles therefore provide rich information, but it often remains challenging to identify the useful signal for decision-making.

Ensemble information can be synthesized using clustering algorithms, which gather similar members in a limited number of meaningful scenarios. The current work builds on the strategy proposed by Mounier et al. (from EMS Annual Meeting 2023) for clustering convective-scale precipitation forecasts of the French Arome-EPS. We demonstrate that the approach can be applied to other meteorological fields, and we propose a way to adapt the design of the scenarios to users' needs.

In our methodology, scenarios are obtained by assigning ensemble members to climatological classes defined after applying a classification method to a meteorological database. Members assigned to the same class are grouped together to form a scenario. To reduce the dimensionality of the problem, the first step of the methodology performs a dimension reduction of input meteorological fields, using a convolutional autoencoder. The clustering algorithm is then applied to the meteorological training database projected in the autoencoder's latent space in order to issue climatological clusters.

To obtain scenarios that best fit the users' needs, the baseline autoencoder has been modified to account for users' variables.

This is achieved by adding a dense neural network in the decoder part of the architecture, that aims at predicting the user variable. We show how we can control the latent representation learned by the autoencoder by adding this supervised learning term to the reconstruction.

This user-oriented clustering approach will be illustrated for the case study of wind energy production. Thanks to a collaboration with the Compagnie Nationale du Rhône, the method has been trained considering a 5-year dataset of 100m wind analysis from the Arome model as meteorological inputs, and wind energy production values in north-western France as the user variable.