Enhanced probabilistic forecasting of wind power using a distribution-free PDF evolution model

Probabilistic forecasting is vital for power grid systems, supporting decision-making under uncertainty and enabling long-term planning. The Probability Density Function (PDF) provides a complete representation of wind power's stochastic behavior. However, most existing PDF modeling methods are static and assume a predefined distribution, limiting their ability to reflect the true variability of wind power. This paper proposes a novel approach to model the dynamic evolution of wind power PDFs without prior distribution assumptions. Historical PDF curves are first projected into a functional space using a radial basis function model, from which representative feature vectors are extracted. These vectors are then combined with weather features and enhanced using a multiple dual attention mechanism to capture temporal and contextual dependencies. The augmented features are fed into a fully connected neural network to predict the evolving PDF of wind power. By directly approximating the true probability distribution, the method effectively captures wind power's stochastic volatility. Experiments on publicly available datasets demonstrate the effectiveness of the proposed model, which consistently outperforms existing methods in terms of PDF forecasting accuracy and robustness.