Generative probabilistic forecasting of wind power: A Denoising-Diffusion-based nonstationary signal modeling approach

The large-scale integration of wind generation results in considerable uncertainties in power systems because of the nonstationary and stochastic nature. However, limited studies have been focused on the nonstationary properties of wind power. Also, accurate modeling the uncertainty of wind power is yet to be achieved. In this study, an integrated model with Diffusion as the backbone and nonstationary enhancement as the kernel was proposed for probabilistic wind power forecasting. First, Denoising Diffusion was established to simulate the uncertainty of the wind power series through diffusion and denoising processes. Subsequently, the transition probabilities of Diffusion in reverse process were learned by a novel nonstationary enhancement, which was designed to prevent over-stationarization and enhance temporal dependencies. As case study reveals, the proposed method can improve stability and robustness, which can fulfill the requirements of wind power probabilistic forecasting from 10 min to 2.5 h.