Short-term Power Prediction For Wind Farm Clusters Based On Deep Learning And Transfer Learning

Wind power is with the characteristics of randomness and volatility.Therefore,the large-scale integration of wind power into the power grid brings challenges to the safe and stable operation of the grid.The future output power of wind farm is predicted in advance,and the space reserved for consumption according to the predicted results is an important technical means to improve the level of new energy consumption and ensure the safety of the power system.The power prediction accuracy directly affects whether the reserved consumption space is reasonable.One of the effective measures to improve the capacity of new energy consumption is to establish a high-accuracy wind power prediction system,formulate a dispatch plan scientifically based on the prediction results,and reserve a reasonable consumption space.In this paper,the short-term power prediction of wind farm clusters was studied from four aspects: feature construction and selection,deep learning modeling,sample transfer learning and feature transfer learning.1)The feature construction and selection of wind farm clusters were studied,and the optimal feature set was analyzed.Based on the multi-dimensional meteorological data of wind farms,the high-dimensional candidate feature library was constructed;the composite meteorological feature selection method based on data mining was studied,the core feature combination suitable for power prediction of wind farm cluster was selected.The feature selection provides a theoretical basis for the subsequent modeling of power prediction for wind farm clusters.2)A short-term power prediction model for wind farm clusters based on deep learning was established.Based on the results of feature construction and selection,the deep learning prediction method based on Bi-LSTM was studied in this paper,a deep learning prediction model based on high-dimensional feature input was established,and the LSTM based on PSO optimization and the SDAE-SVR based on BA optimization were compared.The prediction results of each model under different wind processes are respectively counted,and the optimal prediction model suitable for different wind processes was obtained.The results show that the predicted RMSE for 24 hours of Bi-LSTM is 1.1% lower than that of traditional BP.3)A short-term wind power prediction model based on multi-level transfer learning was established.For the problem of insufficient operating data for newly built,expanded or renovated wind farms in wind farm clusters,the correlation classification of source data based on cross-correlation was studied,effective transfer samples from the massive source wind farm data was selected,and divided into different levels according to the correlation between the samples and the target wind farm data.Based on the multi-level similar samples of the source wind farm,the network parameters of the source wind farm prediction model were transfered hierarchically to the target wind farm,and a multi-level migration learning prediction model was established.Compared with the model without transfer learning,the predicted RMSE for 24 hours was reduced by 6.7%.4)A wind power short-term power prediction model based on feature migration was established.The screening of similar features between source and target wind farms was studied,and a migration learning prediction method based on feature screening was proposed.The feature mapping technology between source and target wind farms was studied,and a migration learning prediction method based on feature mapping was proposed.Based on feature screening and feature mapping methods,the integrated prediction method was studied,and a short-term wind power integrated prediction model based on feature migration was established.The forecast accuracy of the integrated forecasting model is higher than that of the single forecasting model,which can realize the complementary advantages of each single model,thereby improving the overall forecasting accuracy.Compared with the sample transfer learning model,the predicted RMSE for 24 hours was reduced by 0.27%.