With the large-scale integration of renewable energy sources, probabilistic forecasting of wind power output has become essential for supporting the economical and sustainable utilization of wind energy. Ensemble weather forecasts (EWFs) based on numerical weather prediction models have been widely recognized as a promising approach, as they provide multiple weather scenarios that capture the uncertainty in wind conditions. However, these EWFs are computationally intensive and typically updated only a few times per day, which may be insufficient for power producers who need accurate and timely output forecasts at various arbitrary times, such as when generating production schedules or submitting market bids in time for gate closures in intra-day or day-ahead electricity markets. To overcome this limitation, it is crucial to incorporate data-centric approaches that leverage the most recent monitoring data of actual wind power output, in conjunction with EWFs, for frequent and responsive forecast updates. In particular, it is important to appropriately utilize the multiple wind power scenarios provided by EWFs by identifying those that are most consistent with recent observations and thus more reliable.

This study proposes a Continuous Ranked Probability Score (CRPS)-guided approach for the dynamic integration of EWFs in probabilistic wind power prediction. The framework builds on a machine learning model that first generates a sequence of probability density forecasts—specifically, probability density functions of wind power output—individually from each ensemble member. These member-wise forecasts are then dynamically recombined by optimizing ensemble weights to minimize the CRPS, using the latest observations of actual wind power output as ground truth. To address the risk of reduced sharpness due to overly diffuse coverage from blending multiple probabilistic forecasts, the optimization process is regularized to promote sparsity, thereby encouraging the selective use of ensemble members that best match recent observations. By explicitly incorporating CRPS as the objective function, the method adaptively enhances the alignment between predicted distributions and observed outcomes, providing probabilistic forecasts that are well calibrated in terms of sharpness and coverage. This dynamic scheme supports flexible and frequent forecast updates by leveraging up-to-date observations, even when EWFs are updated infrequently. The proposed approach maintains a favorable balance between sharpness and coverage under rapidly changing wind conditions.

The proposed approach is evaluated through numerical experiments using real-world datasets from operational wind farms in Japan. Results suggest the potential for improved forecast sharpness and coverage relative to traditional methods with static weighting.