Probability density prediction (PDP) of wind power is essential for planned scheduling of generation through coordination with other power sources and limited-capacity storage, and for improving the profitability of wind producers by enabling effective market participation. Ensemble weather forecasts (EWFs) provide valuable information on wind uncertainty, but their computational cost and infrequent updates constrain applicability in intraday decision-making where timely forecasts are required. We propose a PDP framework that dynamically aggregates probability distributions generated by individual EWF members using the continuous ranked probability score (CRPS). Each member first produces its own distribution, and adaptive weights are learned so that the aggregated forecasts best reflect recent observations. To avoid sharpness degradation from naive blending, sparsity regularisation selectively emphasises informative members, resulting in sharper and more accurate forecasts based on a subset of reliable members. By jointly capturing EWF uncertainty and the latest observations, the mechanism enables frequent updates of PDPs even when EWFs are updated only a few times per day. Numerical experiments using real-world data from an operational wind power plant (WPP) show that the proposed approach improves both coverage reliability and sharpness compared with typical methods, including best-member selection and CRPS-inverse weighting, underscoring its potential for practical deployment.