Probabilistic prediction of wind farm power generation using non-crossing quantile regression

The probabilistic prediction of energy generation by a wind farm quantifies the volatility of wind power. Thus, accurate probabilistic predictions can provide valuable information for grid dispatching and a basis for reliability assessment for safe operation. However, the inherent stochasticity and instability of wind power generation and the quantile crossover problem of traditional quantile regression neural networks pose challenges for prediction. Therefore, this study proposes a wind power probabilistic prediction model using a non-crossing quantile regression neural network (NCQRNN). A data preprocessing method using time-varying filtering empirical mode decomposition (TVFEMD) is introduced to reduce the volatility and sophistication of the wind power series. The NCQRNN model is designed to incorporate the monotonicity constraints and predict the results of multiple quantiles simultaneously. Furthermore, the predicted conditional quantiles are mathematically proven to not exhibit any crossover phenomena. The wind power data from Elia Grid, Belgium, is used to verify the prediction effectiveness of the proposed method. The obtained results indicate that the proposed probabilistic prediction model addresses the quantile crossover issue while adequately extracting the nonlinear and temporal features of the wind power series. This method accurately quantifies the uncertainty of wind power with high prediction efficiency and accuracy.