Abstract: In order to avoid the influence of the fluctuation and uncertainty of wind power output on power grid operation, a combined algorithm of improved variational modal decomposition algorithm and integrated gated cyclic neural network is proposed for the short-term power prediction of wind farm. Firstly, aiming at the problems of low prediction accuracy and unstable prediction of single GRU(Gated Recurrent Unit) model, the stacking model is used for fusion and the SGRU(Stacking-GRU) prediction model is established. Secondly, for the common endpoint effect problem in VMD (Variational Modal Decomposition)decomposition algorithm, GRU extension method is used, and ALO ( Ant Lion Optimizer ) algorithm is introduced to optimize the super parameters in the model to establish GAVMD(GRU-ALO-VMD) decomposition model. Firstly, the wind power data is decomposed by GAVMD model, which reduces the instability of the data and improves the training difficulty of the prediction model. Then, the SGRU model is used to predict and reconstruct the decomposed IMF (Intrinsic Mode Function) component, and finally the predicted value of the original wind power data is obtained. The performance of the model is verified by the actual wind farm sample data. The simulation results show that the combined forecasting model based on GAVMD-SGRU maintains the average absolute percentage error index within 6% in different seasons compared with other models, which meets the requirements of the power system for wind power dispatch.

Key words: gated recurrent unit ( GRU) neural network, stacking ensemble algorithm, variational mode decomposition, endpoint effect, wind power prediction