Ensemble Weather Forecasting Scenarios Tailored to Users' Needs: Application to Wind Energy

ABSTRACT

Ensemble Prediction Systems (EPS) are commonly used in weather forecasting to account for uncertainties, but their exploitation for operational prediction remains a challenge and often lacks a user‐oriented perspective. To address this issue, this article examines how a scenario‐based processing of an EPS, which summarizes ensemble information in a few meaningful scenarios, can be adapted to an end‐user. The proposed methodology follows a 2‐step procedure where ensemble information is first projected in a reduced space with a convolutional autoencoder, and a clustering algorithm is then applied in this reduced space to define climatological patterns that will guide the final ensemble members classification. Our main contribution is to integrate a user impact variable in the dimension reduction step, in order to structure the latent space and subsequent clusters according to the user's need. The approach is illustrated for wind power prediction, with 100 m wind fields as predictors and the capacity factor as the impact variable. Compared to a pure wind‐based clustering, leveraging an impact variable provides more coherent scenarios that better represent the correlation between wind and energy production. An application to a case of ramp event shows that our clustering provides a relevant decision support tool. Finally, our design is generic enough to be applied easily to other case studies with different impact variables.