关键词: GＲU 神经网络, 深度学习, Stacking 集成算法, 风功率预测, 风电场

Abstract: In order to improve the accuracy of wind farm short-term power prediction，based on deep learning，
a method for integrating wind farm short-term power prediction using the Stacking algorithm to integrate multiple
GＲU( Gated Ｒecurrent Unit) models is proposed． This method first builds three multi-layer GＲU neural network
models establishing a first-level model，extracts high-dimensional temporal feature relationships in depth，builds a
training set from the prediction results of the first-level model，and then uses the newly generated training set to
train the second GＲU model． The second-level GＲU model uses a single-layer structure to find and correct
prediction errors in the first-level model and improve the overall prediction result． Finally，a two-level model
embedded Stacking fusion model is obtained． Taking the historical data of Ningxia Taiyangshan wind farm as an
example，the accuracy of the model is verified． The experimental results show that the GＲU model fused by the
Stacking algorithm has at least an increase in the average absolute percentage error index of 0． 63 compared to other
algorithms，and the overall prediction effect is ideal． The accuracy of prediction has been improved significantly．

Key words: gated recurrent unit ( GＲU) neural network, deep learning, Stacking integrated algorithm, wind power prediction, wind farm