Ultra-short-term prediction for wind power via intelligent reductional reconfiguration of wind conditions and upgraded stepwise modelling with embedded feature engineering

With the increasing penetration of grid-connected wind power, its ultra-short-term prediction has become critical to actively support the efficient operation of power system. Due to high spatio-temporal dispersion, power prediction of mountain wind farms with temporal resolution less than 15-min faces great challenges. This paper proposes a novel ultra-short-term wind power receding interval prediction approach. Firstly, intelligent reductional reconfiguration of inflow wind conditions for the wind property of wind farms is realized, yielding several feature wind conditions at specific wind turbines' sites. Then, stepwise prediction includes step 1 (multi-step receding prediction of wind conditions) and step 2 (modelling of wind farm power generation characteristics). Herein, time finite difference (TFD) regression vector is defined, involved with the autoregression (AR) or piecewise autoregression with extra inputs (PWARX) structure, yielding a kind of embedded feature engineering processing to input-output data. As a result, using different machine learning (ML) algorithms, the AR-TFD-ML and PWARX-TFD-ML frameworks are presented for step 1 and step 2, respectively. Finally, ultra-short-term interval prediction of wind farm output power can be achieved and evaluated. Finally, above systematic approach is validated in a mountain wind farm from North China, showing excellent prediction accuracy and robustness.