非参数随机微分方程在风电功率超短期概率预测中的应用

Ultra-short-term probabilistic wind power forecasting provides paramount uncertainty information for power system real-time operation. However, the stochastic dynamics of wind power generation are not well clarified in existing studies. To transcend such a research barrier, a nonparametric stochastic differential equation (NSDE) combined with deep neural networks is developed for ultra-short-term probabilistic wind power forecasting. Without prior assumptions of the functional structures, an improved Gaussian process regression method is proposed to adaptively infer NSDEs that flexibly capture the evolving temporal dynamics and stochastic attributes inherent in wind power. To tackle issues of sparse observations and analytic solution deficiency, a novel stochastic dynamics-informed network is embedded with a recurrent temporal interpolator and an energy-guided forecaster. An innovative two-stage training algorithm is presented to optimize the network efficiently. Consequently, probabilistic wind power forecasts are derived via precise solutions of the well-inferred NSDEs for future states. Comprehensive case studies based on actual wind farm data demonstrate the superior performance of the proposed approach.